The cellular architecture of memory modules in Drosophila supports stochastic input integration

  1. Omar A Hafez
  2. Benjamin Escribano
  3. Rouven L Ziegler
  4. Jan J Hirtz
  5. Ernst Niebur  Is a corresponding author
  6. Jan Pielage  Is a corresponding author
  1. Zanvyl Krieger Mind/Brain Institute, Johns Hopkins University, United States
  2. Division of Neurobiology and Zoology, Department of Biology, University of Kaiserslautern, Germany
  3. Physiology of Neuronal Networks Group, Department of Biology, University of Kaiserslautern, Germany
  4. Solomon Snyder Department of Neuroscience, Johns Hopkins University, United States

Abstract

The ability to associate neutral stimuli with valence information and to store these associations as memories forms the basis for decision making. To determine the underlying computational principles, we build a realistic computational model of a central decision module within the Drosophila mushroom body (MB), the fly’s center for learning and memory. Our model combines the electron microscopy-based architecture of one MB output neuron (MBON-α3), the synaptic connectivity of its 948 presynaptic Kenyon cells (KCs), and its membrane properties obtained from patch-clamp recordings. We show that this neuron is electrotonically compact and that synaptic input corresponding to simulated odor input robustly drives its spiking behavior. Therefore, sparse innervation by KCs can efficiently control and modulate MBON activity in response to learning with minimal requirements on the specificity of synaptic localization. This architecture allows efficient storage of large numbers of memories using the flexible stochastic connectivity of the circuit.

Editor's evaluation

In light of the ongoing emergence of volume electron microscopy connectomics, detailed morphologies at the nanometre scale for many neurons are now available, ready for functional and computational analysis. Building on these foundational resources, this work delivers compelling evidence that synaptic inputs onto the dendritic arborisations of readout neurons (MBONs) of the learning and memory system of the adult Drosophila melanogaster contribute with equal weight to the depolarization of the neuron, independently of their location on the arbor, a phenomenon known as synaptic democracy. These important findings establish the validity of computational models based on passive dendritic propagation for simulating fly brain circuits and highlight the differences between the much larger mammalian neurons that present active propagation strategies as part of their approach to synaptic democracy.

https://doi.org/10.7554/eLife.77578.sa0

Introduction

Operating successfully within a complex environment requires organisms to discriminate between a vast number of rewarding and stressful interactions. Learning to avoid potentially harmful interactions will promote survival of individual animals and requires long-term memory mechanisms. In recent years, significant progress has been made toward our understanding of the cellular and circuit mechanisms underlying learning and memory. Despite these advances, at the computational level it remains largely unknown how the cellular and circuit architectures contribute to the efficient formation and storage of multiple memories.

Here, we use the Drosophila melanogaster mushroom body (MB) as a simple model system to investigate the computational principles underlying learning and memory in the context of decision making. A key advantage of the MB, the learning and memory center of the fly, is its relative simplicity in terms of the number and types of neurons compared to the mammalian brain. It has been demonstrated that the MB is required for the acquisition and recall of associative olfactory memories (de Belle and Heisenberg, 1994), and a wide variety of genetic tools are available to access and manipulate key cell types within the memory circuitry (Owald et al., 2015b; Lai and Lee, 2006).

Odor information is transmitted from the antennal lobe via projection neurons (PNs) to the Kenyon cells (KCs) of the MB. PNs are connected in a largely stochastic manner to the approximately 2000 KCs per brain hemisphere, resulting in a unique KC odor representation in individual flies (Litwin-Kumar et al., 2017; Aso et al., 2014a; Caron et al., 2013; Gruntman and Turner, 2013; Lin et al., 2007; Betkiewicz et al., 2020; Eichler et al., 2017). The exception to this is a partial non-random connectivity of a population of food-responsive PNs (Zheng et al., 2022). The formation of olfactory memories within the MB circuit is assayed through experimental manipulations in which normally neutral odors are associated with either positive (reward learning) or negative (avoidance learning) valences (Livingstone et al., 1984).

Electrophysiological and calcium imaging studies demonstrated that any given odor activates approximately 3–9% of KCs (Campbell et al., 2013; Honegger et al., 2011; Turner et al., 2008; Siegenthaler et al., 2019). KCs then transmit this odor information to mushroom body output neurons (MBONs) that subdivide the MB lobes into distinct structural and functional modules by virtue of their dendritic arborizations (Owald et al., 2015a; Aso et al., 2014b; Hige et al., 2015a; Cognigni et al., 2018). During associative olfactory conditioning, the behavioral response to any given odor can be altered to encode positive (approach) or negative (avoidance) behavior. The valence of the conditioned response is encoded by the activation of dopaminergic neurons (DANs) that selectively innervate the structural MB modules defined by the MBON dendrites (Aso et al., 2014a; Saumweber et al., 2018).

DANs of the PPL1 cluster preferentially provide input to the vertical lobe of the MB and are activated during negative reinforcement, while DANs of the PAM cluster are activated during positive reinforcement and modulate activity in the horizontal lobe of the MB (Handler et al., 2019; Aso and Rubin, 2016; Tomchik and Davis, 2009; Séjourné et al., 2011; Cohn et al., 2015; Burke et al., 2012; Rohwedder et al., 2016; Saumweber et al., 2018; Selcho et al., 2009). The majority of MBONs of the vertical lobe utilize acetylcholine as a neurotransmitter, and artificial activation of these MBONs often mediates approach behavior (Aso et al., 2014b; Eschbach et al., 2021). In contrast, the majority of MBONs of the horizontal lobe release glutamate as a neurotransmitter, and these neurons mediate avoidance behavior (Aso et al., 2014b; Eschbach et al., 2021).

In vivo calcium imaging and electrophysiological experiments demonstrated that the activity of individual MBONs can be altered by conditioning paradigms (Hige et al., 2015a; Aso and Rubin, 2016; Plaçais et al., 2013; Owald et al., 2015a; Jacob and Waddell, 2020; Perisse et al., 2016; Felsenberg et al., 2018). This modulation of MBON activity depends on the pairing of DAN and KC activity and results in either an enhancement or depression of the strength of the KC>MBON connection (Handler et al., 2019; Owald et al., 2015a; Cohn et al., 2015; Burke et al., 2012). These changes in KC>MBON connection strength occur in either the horizontal or the vertical lobe depending on the administered stimulus-reward contingencies. As a consequence, the resulting net activity of all MBONs is shifted to either approach or avoidance behavior, and the behavioral response toward the conditioned stimulus during memory recall is changed (Handler et al., 2019; Aso and Rubin, 2016; Cohn et al., 2015; Burke et al., 2012). Evidence for this MBON balance model (Cognigni et al., 2018) has been provided for short-term memory (Aso et al., 2014a; Hige et al., 2015a; Perisse et al., 2016; Cohn et al., 2015) but to what extent, if any, such a model applies to long-term memory remains unclear. Recently, genetic tools have been created that enable genetic tagging of long-term memory engram cells in the MB (Miyashita et al., 2018; Siegenthaler et al., 2019). These studies indicate that modulation of a relatively small number of KCs is sufficient to significantly alter memory recall.

Here, we generate a realistic computational model of a core KC>MBON module that is essential for the incorporation of memories to determine the cellular basis for decision making. Prior studies (Gouwens and Wilson, 2009) used ‘synthetic’ (randomly generated) models of dendritic trees of Drosophila projection neurons to gain first insights into the computational processes of central neurons. Using linear cable theory, they demonstrated that voltage differences are much smaller within the dendritic trees than between the dendrites and the soma. Using a similar approach, Scheffer and coworkers simulated voltage distributions in the dendritic tuft of another Drosophila neuron (EPG) but used its measured morphology, rather than random geometries (Scheffer et al., 2020). Here, we go beyond these studies by combining precise structural data including the full synaptic connectivity from the electron microscopy-based synaptic connectome (Takemura et al., 2017) of the Drosophila MB with functional (electrophysiological) data of an individual MBON. This approach allows us to determine the impact of populations of KCs on MBON activity and to resolve the potential computational mechanisms underlying memory encoding. We focus on MBON-α3 (MBON-14 in the terminology of Aso et al., 2014a), an MBON at the tip of the α-lobe that is relevant for long-term memory (Aso et al., 2014b; Plaçais et al., 2013). We first determine the physiological parameters of MBON-α3 by patch-clamp electrophysiology. We then use these parameters to generate an anatomically and physiologically realistic in silico model of MBON-α3 and its KC inputs. Using a variety of computational simulations, we provide evidence that MBON-α3 is electrotonically compact and perfectly suited to incorporate a substantial number of memories based on largely stochastic KC inputs.

Results

Electrophysiological characterization of MBON-α3

MBON-α3 is particularly suitable for computational modeling due to the availability of its complete dendritic reconstruction through electron microscopy (Takemura et al., 2017). In addition, we can access MBON-α3 genetically by using a specific Gal4 line (MB082C-Gal4) to label MBON-α3 by the expression of membrane-bound GFP (10xUAS-IVS-mCD8::GFP; Figure 1A). In immunohistochemical co-labelings with the active zone marker Bruchpilot (Brp), we were able to visualize the dendritic input sites, the axonal projections and the cell soma of MBON-α3 (Figure 1A, arrow). We used this labeling to perform fluorescence-guided electrophysiological whole-cell patch-clamp recordings of this neuron ex vivo (Figure 1C and D). We carried out standard patch-clamp protocols in five cells of independent preparations and recorded all essential physiological parameters (Table 1). It is of note that we cannot discriminate between the two different MBON-α3 cells that are marked by the split Gal4 line and that are almost identical with regard to their dendritic tree position and synaptic input number (Takemura et al., 2017). First, we performed a continuous current-clamp protocol over 60s at a sampling rate of 50kHz which allowed us to determine the average resting membrane potential as -56.7±2.0mV. This value for the resting membrane potential is in line with prior measurements of MBONs in vivo and of other central neurons in Drosophila (Hige et al., 2015a; Wilson et al., 2004; Gu and O’Dowd, 2006; Groschner et al., 2022). In these recordings, we observed spontaneous firing activity of MBON-α3 with an average frequency of 12.1Hz. Next, we performed a current-clamp de- and hyper-polarization protocol where we recorded 400ms trials at a sampling rate of 50kHz. Each trial started with 0pA for 10ms, followed by a current injection of -26pA to 32pA, with 2pA increments for 400ms. These protocols allowed us to evoke action potentials and to determine the resulting changes in membrane potential. An example displaying selected traces of a representative recording is shown in Figure 1F. Two additional examples are illustrated in Figure 1—figure supplement 1. From this data, we determined the average membrane time constant as τm=16.06±2.3ms. We then performed a voltage-clamp step protocol without compensating for the series resistance. In this experiment, we recorded at 20kHz while applying a 5mV voltage step for 100ms. This allowed us to calculate the average membrane resistance as Rm=926±55MΩ, a rather high membrane resistance when compared to similarly sized neurons (Gouwens and Wilson, 2009). MBON-α3 is thus highly excitable, with small input currents being sufficient to significantly alter its membrane potential.

Figure 1 with 2 supplements see all
Electrophysiological characterization of the MBON-α3 neuron.

(A) Immunohistochemical analysis of the MBON-α3 (green, marked by GFP expression) reveals the unique morphology of the MBON with its dendritic tree at the tip of the α-lobe and the cell soma (arrowhead) at a ventro-medial position below the antennal lobes. The brain architecture is revealed by a co-staining with the synaptic active zone marker Bruchpilot (Brp, magenta). (B) Artificial superimposition of a partial rendering of MBON-α3 (green) and of αβγ-KCs (magenta) to illustrate KC>MBON connectivity. A simplified schematic αβ-KC projection is illustrated in white. (C) Transmitted light image of an MBON-α3 cell body attached to a patch pipette. (D) Visualisation of the GFP-expression of MBON-α3 that was used to identify the cell via fluorescence microscopy. The patch pipette tip is attached to the soma (right side). The scale bar in C applies to C and D. (E) Recording of spontaneous activity of an MBON-α3 neuron in current clamp without current injection. (F) Recording of evoked neuronal activity of an MBON-α3 neuron with step-wise increasing injection of 400ms current pulses. Pulses start at -26pA (bottom) with increasing 2pA steps and end at +32pA (top). (G) Mean trace of the induced membrane polarization resulting from a 200ms long current injection of -10pA. 105 trials from three different cells (35 each) were averaged. Gray shading indicates the SD. (H) Relative change of membrane potential after current injection. Vm is equal to the maximum depolarization of the membrane. (I) Correlation between the frequency of action potential firing and changes in membrane potential. (J) Relative change of spike amplitude after current injections. Spike amplitude represents the peak of the largest action potential minus baseline. Also see Figure 1—figure supplement 1 and Figure 1—figure supplement 2.

Table 1
Summary of passive membrane properties of MBON-α3, calculated from electrophysiological measurements from five neurons.
SampleVm[mV]τm[ms]Cm[pF]Cspec[μF/cm2]Cpass[S/cm2]
Cell 1–59.213.5412.350.2001.87×10−5
Cell 2–52.314.4620.790.3372.12×10−5
Cell 3–56.224.5821.600.3501.48×10−5
Cell 4–52.815.5015.830.2571.66×10−5
Cell 5–63.112.2013.210.2141.76×10−5
Mean–56.716.0616.760.2721.78×10−5
SD4.54.924.260.0690.24×10−5
SEM2.02.201.900.0310.11×10−5
Table 1—source data 1

Summary of the passive membrane properties of MBON-α3, calculated from electrophysiological measurements from four neurons marked by cytoplasmic EGFP.

https://cdn.elifesciences.org/articles/77578/elife-77578-table1-data1-v4.xlsx

The amplitude of action potentials was rather small (4.3±0.029mV) but similar to previously reported values for different MBONs (Hige et al., 2015a). The small amplitudes are likely a result of the unipolar morphology of the neuron with a long neurite connecting the dendritic input region to the soma (see below). As a result, signal propagation may partially bypass the soma (Figure 1A, B and F; Gouwens and Wilson, 2009). We calculated the neuron’s membrane capacitance (Cm=τm/Rm) as Cm=16.76±1.90pF, which classifies MBON-α3 as a mid-sized neuron.

To enable the calibration of the in silico model, we then recorded 35 trials per cell in a current-clamp injection protocol with injections of 10pA current pulses for 200ms at a sampling rate of 20kHz. We averaged 105 traces from three selected example cells to a standard curve, describing the membrane kinetics of MBON-α3 in response to current injections (Figure 1G). To further characterize MBON-α3, we analyzed the resulting data from the current-clamp de- and hyper-polarization protocols in the three example cells (Figure 1F and Figure 1—figure supplement 1). We recorded the absolute deflection, the action potential firing frequency, and the amplitude of action potentials for all individual cells. In all three cells, we observed that the absolute deflection of the membrane potential was proportional to the increasing current injections (Figure 1H). The action potential firing frequency increased gradually with increasing membrane potential deflections (Figure 1I), with no significant change in action potential amplitudes (Figure 1J). The same effect was observed with an increasing current injection (Figure 1—figure supplement 1). This is consistent with the general idea that action potentials are "all or nothing" events, and that MBON-α3 is a spike-frequency adapting neuron constistent with prior observations of other MBONs (Hige et al., 2015a). We repeated these recordings with a cytosolic EGFP as an independent and non-membrane associated marker and observed consistent results (Figure 1—figure supplement 2 and Table 1—source data 1). We next used these data to generate an in silico model based on realistic passive membrane properties.

Constructing an in silico model of MBON-α3

For our in silico model of MBON-α3, we used the recently published electron-microscopy-based connectome of the Drosophila mushroom body (Takemura et al., 2017). This dataset includes the precise morphological parameters of the entire dendritic tree including the precise location data for all 12,770 synaptic connections from the 948 innervating KCs. It is important to note that this dataset represents a single fly and the EM-based reconstruction was not able to faithfully assign all synaptic connections (potentially missing 10% of synapses; Takemura et al., 2017). The comparison with the hemibrain data set (Scheffer et al., 2020) that provides a second independent reconstruction of the MBON from a different fly showed very few differences in KC to MBON-α3 connectivity. Thus, despite potential shortcomings, this dataset represents the most accurate template to reconstruct the dendritic tree and connectivity of the MBON. To build a complete morphological reconstruction of MBON-α3, we included the axonal reconstructions from the hemibrain dataset (Scheffer et al., 2020) and determined the length of the connection between the soma and the dendritic region via confocal light microscopy. We incorporated these data together with all other values into the NEURON simulation environment (Hines and Carnevale, 1997; Figure 2A and C and Figure 2—figure supplement 1).

Figure 2 with 2 supplements see all
Construction of a computational model of MBON-α3.

(A) Electron microscopy based reconstruction of MBON-α3. Data was obtained from NeuPrint (Takemura et al., 2017) and visualized with NEURON (Hines and Carnevale, 1997). Data set MBON14 (ID 54977) was used for the dendritic architecture (A–C) and MBON14(a3)R (ID 300972942) was used to model the axon (light gray) and synaptic terminal (dark gray, A, C). The proximal neurite (bright green) was defined as the proximal axonal region next to the dendritic tree. This is the presumed site of action potential generation. Connectivity to the soma (blue) is included for illustration of the overall morphology and is not drawn to scale. (B) A magnified side view of the dendritic tree of MBON-α3. Within the dendritic tree a total of 4336 individual membrane sections were defined (light gray area). The 3121 membrane sections that contain one or more of the 12,770 synaptic input sites from the 948 innervating KCs are highlighted in magenta. (C) Simplified in silico model of MBON-α3 highlighting the size of the individual neuronal segments. We simulated recordings at the proximal neurite (specified as the potential site of action potential generation based on morphological parameters; this area is still included in the dendritic tree reconstruction of Takemura et al., 2017) or at the soma. Values for the different sections are provided in Figure 2—figure supplement 1, Figure 2—source data 1. (D) Normalized trace of a simulated membrane polarization after injection of a 200ms long square-pulse current of -10pA at the soma. (E) Normalized and averaged experimental traces (blue) with standard deviation (light blue) of our measured depolarization. (F) Comparison of a normalized induced depolarization from the model (red) and from the experimental approach (black). The model was fitted within the NEURON environment to the measured normalized mean traces with a mean squared error between model and measured data of 0.048mV2. Also see Figure 2—figure supplement 1, Figure 2—figure supplement 2 and Figure 2—source data 1.

Figure 2—source data 1

Numerical values of the individual sections of the neuron utillized for the computational model.

https://cdn.elifesciences.org/articles/77578/elife-77578-fig2-data1-v4.xlsx

Based on the published vector data (Takemura et al., 2017) we subdivided the dendritic tree into 4336 linear sections, of which 3121 are postsynaptic to one or more of the 12,770 synaptic contacts from 948 KCs (Figure 2B). We used linear cable theory (Niebur, 2008) to generate the in silico model of MBON-α3. Linear cable theory is the baseline model for neurites and the most rational approach in the absence of detailed information about potential nonlinear currents.

In previous simulations of Drosophila neurons, the kinetics of membrane polarisation as a response to current injections were adjusted to in vivo measurements through a fitting procedure (Gouwens and Wilson, 2009). We first defined the boundaries of the passive membrane parameters within NEURON’s PRAXIS (principle axis) optimization algorithm and incorporated the experimentally defined passive membrane properties (Table 1). We then determined the biophysical parameters of the in silico neuron by current step injections of -10pA for 200ms at the soma and by recording the neuronal voltage traces. This enabled us to fit the membrane parameters of our model neuron to the experimental data of the three example cells (Figure 2E). The biophysical parameters that resulted from the fitting procedure are provided in Table 2.

Table 2
Electrophysiological values applied for the in silico model of MBON-α3-A.

Parameter values were obtained from literature or were based on the fitting to our electrophysiological data. The first column shows the parameters, with the names used in the NEURON environment in parentheses. The maximal conductance was determined to achieve the target MBON depolarization from monosynaptic KC innervation.

Electrophysiological Properties
VariableDescriptionFittedLiterature
Ra (Ra)Cytoplasm resistivity [Ω×cm]85.4130–400 (Borst and Haag, 1996; Gouwens and Wilson, 2009)
Cm (cm)Specific membrane capacitance [μF/cm2]0.69610.6–2.6 (Borst and Haag, 1996)
gpas (pas.g)Passive membrane conductance [S/cm2]9.399×10-63.8×10-5-1.2×10-4 (Cassenaer and Laurent, 2007)
epas (pas.e)Leak reversal potential [mV]–55.64–60 (Berger and Crook, 2015)
Synaptic Parameters
τs (tau)Time to max. conductance [ms]N.A.0.44 (Su and O’Dowd, 2003)
ESynaptic current reversal potential [mV]N.A.8.9 (Su and O’Dowd, 2003)
Gmax (gmax)Maximal conductance [μmho ]N.A.1.5627×10-5 (Hige et al., 2015b)

We obtained membrane kinetics that closely resemble ex vivo current injections of -10pA, with a mean squared error between model and measured data of 0.0481mV2. For comparison, we normalized both the modelled current injection (Figure 2D) and the average of the traces from the ex vivo recordings (Figure 2E) and overlayed the resulting traces in Figure 2F.

Note that the faster voltage change at the onset and offset of the current injection seen in the simulation is caused by the small size of the soma which results in a much faster ‘local’ somatic time constant than that of the whole cell (Figure 2—figure supplement 2). This is also visible in the average of the recorded traces (Figure 2F), but more obvious in individual (not trial-averaged) recorded traces (Figure 2—figure supplement 2C). Overall, these results indicate that we were able to generate a realistic in silico model of MBON-α3.

Effect of KC>MBON synaptic inputs

As KCs are cholinergic (Barnstedt et al., 2016) we utilized previously determined parameters of cholinergic synapses to simulate synaptic input to MBON-α3. We used a reversal potential of 8.9mV and a synaptic time constant τs=0.44ms as target values (Su and O’Dowd, 2003). Individual synaptic contacts were simulated as alpha functions (gmax(t-ti)×exp(-(t-ti-τs)/τs)) where ti is the time of the incoming spike and gmax=1.56×10-11S, chosen to obtain response levels in agreement with the target values. The average number of synaptic contacts from a KC to this MBON is 13.47 (Takemura et al., 2017), Figure 4B. To determine the computational constraints of MBON-α3, we first stimulated each of the 3,121 dendritic sections that receive KC synaptic innervation using the described alpha-function conductance change. We then "recorded" (from our simulation) the resulting voltage excursions at the dendritic input site (Figure 3A, red), at the proximal neurite (putative action potential initiation site, green) and at the cell soma (blue). The comparative analysis of the responses at these three locations revealed two notable features: (1) voltage excursions directly within the dendrite are faster and substantially larger than in the proximal neurite and soma, and (2), the resulting voltages in the latter two compartments are much less variable than in the dendrite and distributed within a very small voltage range (Figure 3D and E). These results indicate that the morphological architecture and biophysical parameters of MBON-α3 promote a ‘compactification’, or ‘democratization’, of synaptic inputs, with all inputs resulting in similar voltage excursions at the proximal neurite and soma, regardless of the strength of the initial dendritic voltage or the position on the dendritic tree.

Figure 3 with 1 supplement see all
Computational characterization of MBON-α3.

(A) Simulated stimulation of individual synaptic input sites. Voltage deflections of synaptic currents were analyzed locally at the dendritic segment (red, next to electrode tip), the proximal neurite (green) and the soma (blue). The soma is depicted as a square that reflects its size in the NEURON environment. Color code also applies to B-F. (B) Resulting mean voltages (lines) and standard deviations (shading) from synaptic activations in 3121 individual stimulus locations (dendrite), and at the proximal neurite and soma. (C–E) Violin plots of the simulated maximal amplitudes at the dendritic section (C), the proximal neurite (D) and the soma (E) in response to synaptic activations. Dotted lines represents the quartiles and dashed lines the mean. Proportions of the violin body represent the distribution of individual data points. Note the difference in scale and compactification of the amplitudes in the proximal neurite and soma. (F) Distribution of the percentage of voltage attenuation recorded at the soma. (G) Distribution of distance between all individual dendritic sections and the soma. This color code is applied in H and I. (H) The elicited local dendritic depolarization is plotted as a function of the local dendritic section volume. (I) The soma amplitude is plotted as a function of the amplitude at individual dendritic segments. Note different scales between abscissa and ordinate; the blue line represents identity. Also see Figure 3—figure supplement 1.

To a first approximation, compactification can be understood from a simple analytical estimate of the space constant of this neuron. It is obtained by assuming a non-branching cylinder with the mean diameter of the dendritic segments, 0.29μm, as obtained from the morphological data from Takemura et al., 2017. Taking the values for the resistivity of the cytoplasm (Ra=85Ωcm) and the transmembrane conductance (gpas=9.4×10-6S/cm2) from Table 2, the characteristic length over which excitations decay can then be computed by linear cable theory (Niebur, 2008) as λ1,300μm. This is about twice the maximal path length between any two segments of the cell, see Figure 2C. Using the average diameter of all neuronal compartments, both in the dendritic tree and in the other components of the neuron as listed in Figure 2—figure supplement 1, increases λ by about 10% because the non-dendritic segments have larger diameters. In either case, synaptic input even at distal locations on the dendritic tree can influence somatic voltage deflections nearly as strongly as more proximal synapses, see Figure 3E.

To understand additional influences of cell morphology for the input computation, we analyzed the relation between synaptic position and local dendritic volume and the resulting voltages in the dendrite, proximal neurite, and soma. Representing the distance between dendritic inputs and soma (Figure 3G) by a color code, we find that, as in many tree-like structures, there is a systematic tendency that more distal segments (more distant from the soma) have a smaller volume and diameter, Figure 3H. The same figure also shows that the simulated voltage excursions are many times larger in distal, small segments than in proximal, large segments, over a range from 0.03mV to 1mV. This is expected: placing an electrical charge (integrated synaptic current) on a small capacitor (low-volume segment with small surface area) results in a larger voltage excursion than placing the same charge on a large capacitor (large segment). It is highly remarkable, however, that the morphology of the MBON-α3 dendritic tree seems finely tuned, with the goal of maximizing the equality of the functional weight of all synapses. Specifically, the decrease of the voltage excursion in the soma caused by the injection of a current is nearly perfectly (within less than 0.5μV) compensated by the increase of the local dendritic voltage due to the smaller volume of distal segments, see Figure 3I. The approximately 30-fold difference of local dendritic voltages (0.03mV to 1mV) is nearly exactly compensated, such that activation of each synapse has exactly the same effect at the soma, within less than one μV.

We next analyzed the impact of individual KCs on MBON-α3. We simulated the activation of all synapses of each individual KC that innervates MBON-α3, using the synaptic parameters described above, and we recorded the simulated voltage excursions in the soma, Figure 4A. The number of synapses varied between KCs, ranging from 1 to 27 (plus one outlier with 38 synaptic contacts), with a mean of 13.47 synapses per KC (Figure 4B).

MBON-α3 is electrotonically compact.

(A) Simulated activation of all synapses from all 948 individual KCs innervating MBON-α3. A representative example of a KC (KC12) with 13 input synapses to MBON-α3 is shown. (B) Histogram of the distribution of the number of synapses per KC. (C) Membrane potential traces from simulated activation of all 948 KCs. (D) Distribution of the elicited amplitudes evoked by the individual activations of all KCs. (E) Correlation between the somatic amplitude and the number of activated synapses in the trials. (F) Blue: Distribution of the somatic voltage change evoked by the activation of all KCs with exactly 12, 13, or 14 input synapses. Each dot represents one simulated KC. Bars represent mean and SD. Red: Same, but for activation of 13 random synapses per simulation in 1000 independent trials.

In these simulations, we observed a wide range of voltage excursions (note the outlier in Figure 4B, C, D and E). The neuron is firmly in a small-signal operation mode, as we observed a highly linear relation between the voltage excursion and the number of synapses per KC, Figure 4E. Activating a single KC leads to a voltage excursion at the soma with a mean of 0.37mV, Figure 4D. To further analyze this linear relationship between synapse number and somatic voltage excursion, we analyzed all KCs with exactly 12, 13, or 14 synaptic inputs to MBON-α3. This analysis revealed a highly stereotypical depolarization at the soma for all these KCs, despite wide variations in the position of the individual synaptic contacts (Figure 4E and F). Activation of 13 randomly selected synapses in 1000 independent trials resulted in similar voltage excursions as observed for the KCs with 13 synaptic connections. Interestingly, in these simulations the responses showed less variability compared to the activation of individual KCs with synapses at anatomically observed locations (Figure 4F). Together, these results provide further support for the electrotonical compactness of MBON-α3.

Physiological and tuned activation of MBON-α3

A given odor activates only a small fraction of KCs reliably, reported as about 5% in some studies (Honegger et al., 2011; Siegenthaler et al., 2019) and 6% in others (Turner et al., 2008; Campbell et al., 2013). This one-percentage point difference may well be caused by differences in the experimental preparations. However, in absolute terms a change from 5% to 6% in the same system represents a 20% increase. To understand the physiological effect of differences in the number of activated KCs of this size, we first established a baseline condition by simulating voltage excursions in the soma resulting from the simultaneous activation of sets of 50 KCs (corresponding to 5%), randomly selected from the 948 KCs innervating MBON-α3 and addressed increases in KC number afterwards.

In 1000 independent activation trials (Figure 5A), we observed highly stereotypical activation patterns, with a mean somatic depolarization of 15.24mV and only small variations in the time course (Figure 5B and C (blue)). These results are consistent with an activation of 50 KCs being sufficient to induce changes in the firing frequency of MBON-α3 (Figure 1H, I). The fact that all simulations resulted in a significant depolarization consistent with action potential generation or alterations of the firing frequency together with the small range of variation within this dataset, provides support for the current model that odor encoding in KCs is, at least to a large extent, random (stochastic; Caron et al., 2013; Zheng et al., 2022; Li et al., 2020; Eichler et al., 2017).

Effect of KC recruitment and synaptic plasticity on MBON-α3 responses.

(A) Schematic of the simulated recording paradigm. We modeled the activation of ≈5% of random combinations of identified KCs, that is we activated the anatomically correct synapse locations of 50 randomly selected KCs to mimic activation by an odor. The panel shows the distribution of activated synapses for one of these trials. We then varied the number of activated KCs by ±25% around its mean of 50, that is 38, 50, and 63 KCs, and we varied synaptic strength ±25% around its original value, that is 75%, 100%, and 125%. (B) Mean somatic depolarizations after activating 1000 different sets of either 38 (light blue), 50 (blue), or 63 (dark blue) KCs, all at standard synaptic strength (100%). Shades represent the standard deviations. (C) Mean somatic depolarizations after activating 1000 different sets of 50 KCs at either 75% (red), 100% (blue), or 125% (purple) synaptic strength. Shades represent the standard deviations. (D, E) Violin plots of the relative amplitudes evoked by the different KC sets from (B, C). (F–G) Violin plots for the comparison of the amplitudes between the different plasticity paradigms. (F) 63 KCs at 100% vs 50 KCs at 125% synaptic strength. (G) 38 KCs at 100% vs 50 KCs at 75% synaptic strength. Plots in D-G use the color code from (B) and (C). (H) 63 KCs at 75% (dark red) vs. 50 KCs at 100% (blue) and 38 KCs at 125% (light purple) synaptic strength. See Figure 3 for symbols in violin plots. (I) Relation between somatic depolarization and the number and strength of activated synapses. Illustrated are all tested conditions for synaptic plasticity and KC recruitment. Stimulation paradigms are color coded as shown. (J) Relation between the slope of the condition-specific linear regressions from (I) and the synaptic strength of the activated KC sets. (K) Relation between the slope of the condition-specific linear regressions from (I) and the number of recruited KCs. (L) Slopes of the condition-specific linear regressions from (I). Color codes of (J–L) as in (I). Statistical significance was tested in multiple comparisons with a parametric one-way ANOVA test (D, E and H) resulting in p=<0.0001 for all conditions, or in an unpaired parametric t-test (F and G) resulting in p=<0.0001 for all conditions. Numerical results are summarized in Table 3.

Based on in vivo data, two different modes of plasticity may act at the level of the KC-MBON module to alter MBON output to conditioned odors. In vivo imaging of short-term memory formation demonstrated alterations of KC input strength (Owald and Waddell, 2015; Hige et al., 2015a; Perisse et al., 2016; Plaçais et al., 2013; Pai et al., 2013; Séjourné et al., 2011; Jacob and Waddell, 2020) while recent observations of long-term memory demonstrated a specific change (increase) in the number of KCs representing the conditioned odor (Delestro et al., 2020; Shyu et al., 2019; Baltruschat et al., 2021). To directly address the relative impact of these two modulations on MBON depolarization, we repeated the same simulation but now modified input strength by either altering activated KC numbers or by modulating the synaptic gain. To enable a direct comparison, we modulated each factor by 25%. First, we increased the number of activated KCs per set by 25%, from 50 to 63. This lead to a mean somatic depolarization of 18mV, a significant increase by 19% (Figure 5B and D; dark blue). Decreasing the number of activated KCs by 25%, from 50 to 38, lead to a mean somatic depolarization of 12mV, a significant decrease by 20% (Figure 5B and D; light blue). Next, we either increased or decreased synaptic strength within sets by 25%. Increase lead to a mean somatic depolarization of 18mV, a significant increase by 19% (Figure 5C and E; purple) while decrease lead to a mean somatic depolarization of 12mV, a significant decrease by 21% (Figure 5C and E; red). Averaged over all conditions, a change of input of 25% lead to a change of 20±1% in membrane depolarization.

Even though the integrated synaptic conductance was changed by the same amount by our manipulation of either synaptic strength or number of activated KCs, we observed small but significant differences in the resultant mean somatic amplitudes (Figure 5F and G). Finally, we combined both manipulations by compensating an increase in KC number with a corresponding decrease in synaptic strength and vice versa. Both manipulations resulted, on average, in a significant decrease of MBON membrane depolarization (Figure 5). An overview of the performed simulations and numerical results is provided in Table 3.

Table 3
Summary of data from the plasticity simulations of Figure 5.

Simulations vary the number of active KCs and/or the strength of the KC-MBON synapses by scaling the maximal conductance parameter of the synaptic alpha-conductance function. Results include the mean number of activated synaptic sites as well as the mean amplitude of the somatic voltage depolarization across the 1000 trial repetitions of each simulation. Ranges are reported as parentheticals. The slope of the linear regression describing the relation between the number of active synapses and somatic voltage amplitude is reported (Figure 5I–L).

Simulation parametersResults
# KCsSynaptic strengthMean # Activated synapsesMean somatic depolarization amplitude [mV]Slope
63125%847.88 (747 - 939)21.55 (19.66–23.21)0.0184
50125%672.55 (573 - 778)18.14 (15.99–20.29)0.0209
38125%511.77 (425 - 615)14.60 (12.48–16.96)0.0235
63100%847.08 (725 - 977)18.26 (16.15–20.34)0.0166
50100%673.76 (574 - 805)15.24 (13.34–17.58)0.0185
38100%512.48 (429 - 599)12.14 (10.43–13.85)0.0203
6375%849.05 (749 - 965)14.56 (13.11–16.17)0.0141
5075%672.86 (570 - 767)11.98 (10.37–13.41)0.0153
3875%511.97 (420 - 606)9.44 (7.91–10.96)0.0164

Our simulations revealed a nearly linear relationship between the activated number of synapses and the measured amplitude in the soma regardless of synaptic strength (Figure 5I). Interestingly, we observed some differences in the slope of the responses between the different tuning modalities (Figure 5J, K and L), indicating that these modulations are not completely equivalent.

Discussion

For our understanding of the computational principles underlying learning and memory, it is essential to determine the intrinsic contributions of neuronal circuit architecture. Associative olfactory memory formation in Drosophila provides an excellent model system to investigate such circuit motifs, as a large number of different odors can be associated with either approach or rejection behavior through the formation of both short- and long-term memory within the MB circuitry (Aso and Rubin, 2016). In contrast to most axon guidance processes in Drosophila that are essentially identical in all wild-type individuals, processing of odor information in the MB largely, but not exclusively, depends on the stochastic connectivity of projection neurons to KCs that relay odor information from the olfactory glomeruli to the MBONs (Caron et al., 2013; Gruntman and Turner, 2013; Litwin-Kumar et al., 2017; Warth Pérez Arias et al., 2020; Zheng et al., 2022). However, the role of MBONs within the circuit is largely fixed between animals and many MBONs can be classified as either approach or avoidance neurons in specific behavioral paradigms (Aso et al., 2014a; Aso et al., 2014b; Litwin-Kumar et al., 2017; Li et al., 2020; Takemura et al., 2017; Scheffer et al., 2020). Furthermore, individual KCs are not biased in their MBON connectivity but innervate both kinds of output modules, for both approach and avoidance. Learning and memory, the modulation of odor response behavior through pairing of an individual odor with either positive or negative valence, is incorporated via the local activity of valence-encoding DANs that can either depress or potentiate KC>MBON synaptic connections (Aso and Rubin, 2016; Berry et al., 2018; Handler et al., 2019; Yamagata et al., 2015; Cohn et al., 2015; Burke et al., 2012; Hige et al., 2015a; Owald et al., 2015a; Jacob and Waddell, 2020; Pai et al., 2013; Plaçais et al., 2013; Séjourné et al., 2011; Bouzaiane et al., 2015). As KC odor specificity and connectivity differ significantly between flies with respect to the number and position of KC>MBON synapses, this circuit module must be based on architectural features supporting robust formation of multiple memories regardless of specific individual KC connectivity.

Prior computational work addressing properties of central nervous system neurons in Drosophila relied on synthetic (randomly generated) data (Gouwens and Wilson, 2009) or partial neuronal reconstructions (Scheffer et al., 2020; Tobin et al., 2017). We go beyond these previous studies by combining precise structural data from the electron-microscopy based synaptic connectome (Takemura et al., 2017) with functional (electrophysiological) data. The structural data consists of the neuroanatomical structure of MBON-α3, including the 12,770 synaptic inputs of all its 948 innervating KCs. While this EM reconstruction may contain some mistakes in synaptic connectivity as e.g. up to 10% of synaptic sites remained unassigned within the dataset (Takemura et al., 2017), it currently represents the best possible template for an in silico reconstruction. The neuron’s functional properties were determined from ex vivo patch clamp recordings. Near-perfect agreement between experimentally observed and simulated voltage traces recorded in the soma shows that linear cable theory is an excellent model for information integration in this system.

Together, we obtain a realistic in silico model of a central computational module of memory-modulated animal behavior, that is a mushroom body output neuron.

Our data show that the dendritic tree of the MBON is electrotonically compact, despite the complex architecture that includes a high degree of branching. This data is in agreement with a prior electrophysiological characterization of an MBON in locust (Cassenaer and Laurent, 2007), and a similar feature has been reported for neurons in the stomatogastric ganglion of crayfish. Here the electrotonic compactness supports linear integration of synaptic inputs across extensive arborizations and likely serves to functionally compensate for inter-individual variability (Otopalik et al., 2017; Otopalik et al., 2019). The location of an individual synaptic input within the dendritic tree has therefore only a minor effect on the amplitude of the neuron’s output, despite large variations of local dendritic potentials. This effect was particularly striking when we restricted the analysis to a population of KCs with identical numbers of synapses that all elicited highly stereotypical responses. The compactification of the neuron is likely related to the architectural structure of its dendritic tree. Together with the relatively small size of many central neurons in Drosophila, this indicates that in contrast to large vertebrate neurons, local active amplification or other compensatory mechanisms (Rumsey and Abbott, 2006; Sterratt et al., 2012; Froemke et al., 2005; Gidon and Segev, 2009; Schiller et al., 2000; Polsky et al., 2004) may not be necessary to support input normalization in the dendritic tree. In contrast, for axons it has been recently reported that voltage-gated Na+ channels are localized in putative spike initiation zones in a subset of central neurons of Drosophila (Ravenscroft et al., 2020). In case in vivo physiological data quantitatively describing local active currents become available, they can be incorporated into our model to further increase the agreement between model and biological system.

Encoding of odor information and incorporation of memory traces is not performed by individual KCs but by ensembles of KCs and MBONs. Calcium imaging in vivo demonstrated that individual odors evoke activity reliably in approximately 3–9% of KCs (Campbell et al., 2013; Honegger et al., 2011). Our simulations of 1000 independent trials with random sets of 50 KCs, each representing one distinct odor that activates approximately 5% of the KCs innervating the target MBON, demonstrate that such activation patterns robustly elicit MBON activity in agreement with in vivo observations of odor-induced activity that elicited robust increases in action potential frequency in MBONs (Hige et al., 2015a; Warth Pérez Arias et al., 2020). The low variability of depolarizations observed in these simulations indicates that information coding by such activation patterns is highly robust. As a consequence, labeled line representations of odor identity are likely not necessary at the level of KCs since relaying information via any set of ≈50 KCs is of approximately equal efficiency. Such a model is supported by a recent computational study demonstrating that variability in parameters controlling neuronal excitability of individual KCs negatively affects associative memory performance. The authors provide evidence that compensatory variation mechanisms exist that ensure similar activity levels between all odor-encoding KC sets to maintain efficient memory performance (Abdelrahman et al., 2021).

Optical recordings of in vivo activity of MBONs revealed selective reductions in MBON activity in response to aversive odor training (Jacob and Waddell, 2020; Pai et al., 2013; Owald and Waddell, 2015; Perisse et al., 2016; Séjourné et al., 2011) or to optogenetic activation of selective DANs (Hige et al., 2015a). More generally, both depression and potentiation of MBON activity have been previously observed in different MBON modules in vivo (Owald and Waddell, 2015; Hige et al., 2015a; Perisse et al., 2016; Plaçais et al., 2013; Pai et al., 2013; Séjourné et al., 2011; Jacob and Waddell, 2020). In addition, recent studies have observed changes in KC stimulus representations after conditioning (Delestro et al., 2020; Shyu et al., 2019) that may be due to learning-dependent modulations of synaptic PN input to KCs (Baltruschat et al., 2021). Our computational model allowed us to implement and compare these two mechanisms changing MBON output: One is a change in the strength of the KC>MBON synapses (over a range of ±25%); the other is a change in the number of activated KCs (over the same range). Interestingly, we found almost linear relationships between the number of active KCs and the resulting depolarizations, and the same between the strength of synapses and MBON depolarization. Decreasing or increasing either of these variables by 25% significantly altered the level of MBON activation with only minor differences in the extent to which these modifications contributed to MBON depolarization. The two different mechanisms of altering MBON output are potentially utilized for the establishment of different types of memory with fast and local alterations of synaptic transmission likely essential for short-term memory while structural changes may ensure maintenance of memory over long periods of time. In addition, differential modes of plasticity may be required at potentially more static parts of the MB circuitry like the food-related part that is not entirely based on stochastic connectivity (Zheng et al., 2022).

Our simulation data thus shows that the KC>MBON architecture represents a biophysical module that is well-suited to simultaneously process changes based on either synaptic and/or network modulation. Together with the electrotonic nature of the MBONs, the interplay between KCs and MBONs thus ensures reliable information processing and memory storage despite the stochastic connectivity of the memory circuitry. While our study focuses on the detailed activity patterns within a single neuron, the availability of large parts of the fly connectome at the synaptic level, in combination with realistic models for synaptic dynamics, should make it possible to extend this work to circuit models to gain a network understanding of the computational basis of decision making.

Materials and methods

Fly stocks

Request a detailed protocol

Flies were reared at 25°C and 65% humidity on standard fly food. The following stocks were used in this study: 10XUAS-IVS-mCD8::GFP (Pfeiffer et al., 2010) (BDSC 32186), MB082C-Gal4 (Aso et al., 2014b) (BDSC 68286), 2xUAS-EGFP (BDSC 23867) (all from the Bloomington Drosophila Stock Center, Indiana).

Immunohistochemistry

Request a detailed protocol

Male flies expressing GFP in MBON-α3 were fixed for 3.5 hr at 4°C in 4% PFA containing PBST (0.2% Triton-X100). Flies were then washed for at least 3x30min at RT in PBST before dissection. Dissected brains were then labelled with primary antibodies rabbit anti-GFP (A6455, Life technologies, Carlsbad, USA 1:2000) and mouse anti-Brp (nc82, Developmental Studies Hybridoma Bank, Iowa, USA, 1:200) for two nights at 4°C. After incubation, brains were washed at least 6x30 min in PBST. Secondary antibodies (Alexa488 (goatαrabbit) and Alexa568 (goatαmouse) coupled antibodies, Life technologies, Carlsbad, USA, 1:1000) were applied for two nights at 4°C. After a repeated washing period, brains were submerged in Vectashield (Vector Laboratories, Burlingame, USA), mounted onto microscope slides, and stored at 4°C.

Images were acquired using a LSM 710 confocal scanning microscope with a 25 x Plan-NEOFLUAR, NA 0.8 Korr DIC, oil objective (Carl Zeiss GmbH, Jena, Germany) and a Leica STELLARIS 8 confocal microscope with a 20 x HC PL APO NA 0.75 multi-immersion objective (Leica Microsystems, Wetzlar, Germany). Raw images were projected with Fiji (Schindelin et al., 2012) and cropped in Photoshop (Adobe, San José, USA). Uniform adjustments of brightness and contrast were performed.

Electrophysiology

All patch clamp recordings were performed at room temperature similar to prior descriptions (Hige et al., 2015a; Hige et al., 2015b). Flies were flipped 1d after hatching to obtain 2- to 3-day-old flies. Due to the ventral location of the MBON-α3 somata within the brain, we performed the recordings ex vivo. For preparation, flies expressing GFP in MBON-α3 were briefly anesthetized on ice before removing the entire brain out of the head capsule. The preparation was performed in oxygenated (95% and 5%) high glucose (260mM) extracellular saline. Analogous to a prior study, the brains were incubated for 30s in 0.5mg/ml protease (from Streptomyces griseus, CAS# 90036-06-0, Sigma-Aldrich, St. Louis, USA) containing extracellular saline (Wilson et al., 2004). The brain was then transferred to standard extracellular saline (Wilson et al., 2004; in mM: 103 NaCl, 3 KCl, 5 TES, 10 trehalose, 10 glucose, 7 sucrose, 26 NaHCO3, 1 NaH2PO4, 1.5 CaCl2, 4 MgCl2, pH 7.3, 280-290mOsmol/kg adjusted with sucrose). For physiological recordings, the brain was transferred to a glass bottom chamber with continuously perfused (2ml/min) oxygenated standard extracellular saline and held in place by a custom made platinum frame.

We used glass capillaries (GB150 (F-)–10 P, Science Products, Hofheim, Germany) and a horizontal puller (Flaming Brown Micropipette Puller P-1000, Sutter instruments, Novato, USA) to obtain pipettes with a resistance of 7-10MΩ. The puller was equipped with a 2.5 mm box filament and patch pipettes were pulled with a program consisting of 2 lines (Line 1/Line 2: Heat 472/463 with a ramp value of 468, Pull 0/41, Velocity 52/70, Delay 1/10, Pressure 350). For recordings, patch pipettes were filled with internal solution (Mauss et al., 2014) (in mM: 140 K-aspartate, 10 HEPES, 4 Mg-ATP, 0.5 Na-GTP, 1 EGTA, 1 KCl, pH 7.29 adjusted with KOH, 265mOsmol/kg). Whole-cell recordings were made using the EPC10 amplifier (HEKA, Reutlingen, Germany) and the PatchMaster software (HEKA, Reutlingen, Germany). Signals were low-pass filtered at 3kHz, and digitized at 10kHz via a digital-to-analog converter. The liquid junction potential of 13mV was corrected online. Patch pipettes were guided under visual control with an upright microscope (BX51WI; Olympus, Tokio, Japan) equipped with a 60x water immersion objective (LUMPlanFl/IR; Olympus, Tokio, Japan) and a CCD camera (Orca 05 G, Hamamatsu, Japan). GFP signal of target cells was visualized through fluorescence excitation (100 W fluorescence lamp, Carl Zeiss GmbH, Jena, Germany), and a dual-band emission filter set (512/630 HC Dualband Filter GFP/DsRed, AHF Analysentechnik, Germany). We used GFP expression in soma and neurite to identify MBON-α3 neurons. During recordings, the fluorescence excitation was shut off to minimize phototoxic effects. The complete setup was mounted on an air-damped table while being shielded by a Faraday cage. Series resistance was maintained below 90MΩ and compensated for up to 35% through the amplifier’s compensation circuitry. In current-clamp mode, cells were held at a baseline of -60mV to measure the relevant parameters for the model. Signals were recorded with a sample rate of 20kHz or 50kHz and low-pass filtered at 5kHz.

To determine passive electrical membrane properties and qualitative patch characteristics, a series of protocols were performed:

Protocol 1

Request a detailed protocol

In voltage-clamp mode, a voltage-step (100ms, 5mV) is applied without compensating the series resistance. The corresponding current trace was then used to determine Rseries and Rinput using Ohms law and Rm was determined as (Rseries-Rinput) (Numberger and Draguhn, 1996) in Igor Pro (WaveMetrics, Portland, USA).

Protocol 2

Request a detailed protocol

In current-clamp mode, a step-protocol was performed and recorded at 50kHz. Each sweep starts with 0pA for 10ms, which is followed by a current injection (-26pA to 32pA, with 2pA increments) and a step duration of 400ms. The resulting changes in membrane potential and induced action potentials were then analyzed and plotted with Matlab (MathWorks, Natick, USA). The maximum depolarization was calculated as a mean of the last 70 sample points of the stimulation sweep in each trace. The resting membrane potential was calculated as the mean of the first 500 sample points before the stimulation in each trace. The absolute action potential was measured as the maximum of each trace. The relative action potential amplitude resulted from subtracting the baseline from the absolute value. The baseline resulted from the mean of a 400 sample points window, starting 700 sample points upstream of the maxima in each trace. The resulting trace of the -10pA injection was used to calculate the membrane time constant τm of each cell.

Protocol 3

Request a detailed protocol

In current-clamp mode, no current was injected throughout a 60s recording at 50kHz. The resting membrane potential was determined as the mean of the baseline in the recording with Igor Pro and spontaneous activity was plotted with Matlab.

Protocol 4

Request a detailed protocol

In current-clamp mode, a short current pulse (-10pA, 200ms) was injected and recorded at 20kHz. We performed 35 iterations per cell. The resulting data of three example cells was used for fitting the membrane kinetics of the model as described previously (Gouwens and Wilson, 2009). Averaging, normalisation ((Vm-Vm0)/Vm0) and plotting was performed with Matlab. This data was used to perform the model fitting. The baseline (Vm0) resulted from the average of the sample points before stimulation.

Cells displaying a resting membrane potential higher than -45mV and/or where the series resistance was too high (> 90MΩ) were excluded from the analysis. Only cells that were firing action potentials were included in the analysis. A list of the five included cells using mCD-GFP as a marker and of the four cells using cytoplasmic GFP as a marker and the corresponding calculated values are provided in Table 1 and Table 1—source data 1.

Computational model

Request a detailed protocol

Morphology data for the Drosophila MBON-α3-A was originally characterized using scanning electron microscopy (Takemura et al., 2017) and for the present study obtained from the ‘mushroombody’ dataset hosted in the database neuPrint (https://neuprint-examples.janelia.org/; cell ID 54977). These data describe the structure of its dendritic arborization (Figure 2B, and Figure 2—figure supplement 1), with coordinates specified at 8 nm pixel resolution and converted to μm for model implementation. The reconstruction is limited to the portion of the neuron in the mushroom body, not including its soma and axon. The geometry of the axon was determined from the corresponding MBON14(α3)R (Figure 2A), obtained from neuPrint’s ‘hemibrain:v1.1’ dataset (https://neuprint.janelia.org/; cell ID 300972942). The axon and synaptic terminal of this neuron were approximated as five separate sections, distinguished by the major branch points identified in the electron microscopic reconstruction (Figure 2—figure supplement 1). Each section was characterized by its total length and average diameter and divided into segments to maintain the average section length of that region in the original reconstruction. They were then appended to the MBON-α3-A proximal neurite as individual linear segments (Figure 2—figure supplement 1). The dimensions of the soma were characterized by confocal microscopy of MBON-α3-A (Figure 1A). Synaptic locations of the KC innervation of MBON-α3-A were also obtained from neuPrint’s ‘mushroombody’ dataset. MBON-α3-A is innervated by 948 distinct KCs, forming a total of 12,770 synapses with an average of 13.47 synapses per KC. This data set potentially misses up to 10% of synaptic connections and includes errors associated with electron microscopy reconstructions (Takemura et al., 2017).

Electrophysiological properties of the model neuron (Table 2) were determined by fitting the model response to the recorded data, an approach consistent with prior work modeling Drosophila central neurons (Gouwens and Wilson, 2009). Somatic voltage excursions in response to 200ms current injections of -10pA were recorded ex vivo (Figure 1G) and compared with the simulated membrane potential (Figure 2D). This current injection was replicated in silico and the membrane potential change at the soma was fit to our recorded data by varying the cell’s cytoplasm resistivity, specific membrane capacitance, and passive membrane conductance. The fitting was performed using NEURON’s principle axis optimization algorithm, where parameters were tuned until the computed voltage excursion at the soma matched our electrophysiological recordings. The model optimization searched parameter space to minimize an error function defined as the squared differences between the averaged experimental responses to current pulses (Figure 1G) and the computed model responses (Figure 2D–F). The fitting was then verified by normalizing both the model and experimental voltage excursions. The model was first allowed to equilibrate for 30ms simulated time, after which the 200ms current pulse was initiated. The region of data used for the fitting ranged from 32.025ms to 228.03ms, excluding approximately the first and last 2ms of current injection. The domain of the parameter space was restricted to physiologically plausible values: 0.5μF/cm2 to 1.5μF/cm2 for specific membrane capacitance, 1E-7S/cm2 to 1E-4S/cm2 for passive membrane conductance, and 30Ωcm to 400Ωcm for cytoplasm resistivity (Table 2). The final values for each parameter are well within these margins. Resting membrane potential and the leak current reversal potential were set at -55.64mV based on the initial conditions of the recorded data to which the model was fit.

Cholinergic KC to MBON synapses were modeled as localized conductance changes described by alpha functions (Table 2). KC innervation of MBON-α3-A was determined to be cholinergic (Barnstedt et al., 2016). Currents recorded in Drosophila KCs were demonstrated to have a reversal potential of 8.9mV and a rise time to a maximal conductance of 0.44ms (Su and O’Dowd, 2003). The maximal conductance for the synapses was set at 1.5627*10-5μS, the value determined to achieve the target MBON depolarization from monosynaptic KC innervation (Hige et al., 2015b).

All simulations were performed using the NEURON 7.7.2 simulation environment (Hines and Carnevale, 1997) running through a Python 3.8.5 interface in Ubuntu 20.04.1. The morphology data divided the MBON-α3-A dendritic tree into 4336 sections. We added 6 additional sections describing the axon, synaptic terminals, and soma to a total of 4342 sections with an average of 1.24μm in length and 0.29μm in diameter (Table 4). Passive leak channels were inserted along the membrane of the dendrite sections as well as throughout the axon and soma. Each dendritic section was automatically assigned an index number by NEURON and modeled as a piece of cylindrical cable, with connectivity specified by the neuron’s morphology. The coordinate location of each KC synapse was mapped to the nearest section of the MBON dendrite, and the conductance-based synapse model was implemented in the center of that section. Simulations were performed in which each synapse-containing section was individually activated, followed by additional simulations activating groups of synapses corresponding to single or multiple active KCs. A list of all performed simulations related to Figure 5 is provided in Table 3. Membrane potential data was analyzed and plotted with MATLAB R2020b (MathWorks, Natick, USA) and Prism 9 (GraphPad, San Diego, USA).

Table 4
Morphological parameters of the model.
ParameterValue
Number of sections4342
Average section length1.24μm
Total length5377.39μm
Average diameter0.29μm
Average segment surface area1.26μm2
Total surface area6168.32μm2
Soma diameter6.45μm

All code and data files necessary to replicate the simulations are available at: https://doi.org/10.7281/T1/HRK27V.

Statistical analysis

Request a detailed protocol

Statistical analysis and visualization of data was performed in Prism 9 (GraphPad, San Diego, USA) and Matlab (MathWorks, Natick, USA). Before a statistical comparison was performed, individual groups were tested in Prism for normality and lognormality with an Anderson-Darling test. Statistical significance was tested with an unpaired parametric t-test or in case of multiple comparisons, with an ordinary corrected parametric one-way ANOVA test. Further information is provided in figure legends.

Data availability

All data generated or analysed in this study are included in the manuscript. All simulation files and the code and data files needed to replicate the simulations are available as a permanent and freely accessible data collection at the Johns Hopkins University Data Archive: https://doi.org/10.7281/T1/HRK27V. This includes the simulation code itself (python), the structural EM reconstruction of MBON-alpha3 (swc), the EM reconstruction of the related MBON used to model the axon and synaptic terminal structures (swc), the synapse locations as coordinate data (json), and the synapse locations by MBON section (json). Parameter values for model definition and individual simulations are specified within the code files and outlined in each figure legend where appropriate.

References

  1. Book
    1. Numberger M
    2. Draguhn A
    (1996)
    Patch-Clamp-Technik
    Heidelberg: Spektrum Akademischer Verlag.
    1. Owald D
    2. Lin S
    3. Waddell S
    (2015b) Light, heat, action: neural control of fruit fly behaviour
    Philosophical Transactions of the Royal Society of London. Series B, Biological Sciences 370:20140211.
    https://doi.org/10.1098/rstb.2014.0211

Decision letter

  1. Albert Cardona
    Reviewing Editor; University of Cambridge, United Kingdom
  2. K VijayRaghavan
    Senior Editor; National Centre for Biological Sciences, Tata Institute of Fundamental Research, India
  3. Albert Cardona
    Reviewer; University of Cambridge, United Kingdom

Our editorial process produces two outputs: (i) public reviews designed to be posted alongside the preprint for the benefit of readers; (ii) feedback on the manuscript for the authors, including requests for revisions, shown below. We also include an acceptance summary that explains what the editors found interesting or important about the work.

Decision letter after peer review:

Thank you for submitting your article "The cellular architecture of memory modules in Drosophila supports stochastic input integration" for consideration by eLife. Your article has been reviewed by 3 peer reviewers, and the evaluation has been overseen by a Reviewing Editor and K VijayRaghavan as the Senior Editor. The following individual involved in review of your submission has agreed to reveal their identity: Albert Cardona (Reviewer #1).

The reviewers have discussed their reviews with one another, and the Reviewing Editor has drafted this to help you prepare a revised submission.

Essential revisions:

Please follow carefully the many detailed comments by the reviewers. To emphasize comments with regard to the caveats of using partially EM-reconstructed neuronal morphologies, and in the statements regarding electrotonic compactness.

Reviewer #1 (Recommendations for the authors):

Hafez and collaborators describe the construction and analysis of a computational model of a mushroom body neuron. The anatomy derives from a combination of electron microscopy reconstructions of MBON-α3 and also from light microscopy. The physiological parameters derive from publications that measured them, in addition to the author's own electrophysiological recordings with patch-clamp.

There are two main findings. First, the dendritic arbor of MBON-α3 is electrotonically compact, meaning, individual connections from Kenyon cells will similarly elicit action potentials independently as to where, spatially, the synapses lay on the arbor. Second, in simulation, exploration of changes in the strength of Kenyon cell inputs illustrate two possible ways to alter the strength of the KC-MBON physiological connection, showing that either could account for the observed synaptic depression in the establishment of associative memories. The properties of each approach differ.

Overall, the manuscript clearly describes the journey from connectomics and electrophysiology to computational modeling and exploration of the physiological properties of a circuit in simulation.

The discussion ought to be expanded to include the implications of two possible approaches to physiologically altering the KC-MBON synapse and the consequence of their combination in expanding the space of alterations induced by associative memory paradigms.

In general, the results are clear, but some details remain underdetermined and I have listed them below in the detailed comments. The introduction and discussion present some inaccuracies that can be swiftly addressed by the authors

Detailed comments:

Line 45: Language: "potential rewarding": potentially.

Line 49: Language: "cellular and circuit architecture contributes": architectures contribute.

Line 72: instead of 5%, the number of KCs active at any one time seems to be 6% as per Turner et al. 2008 and Campbell et al. 2013. What is the robustness of the analysis to this small change? Did you explore a range of possible single-digit percent KC activations?

Line 99: Aso & Rubin 2016 belongs with citations in line 95.

Line 122: "Drosophila" needs italics, throughout the manuscript.

The authors devise a computational model of an MBON using a neuronal arbor reconstructed from volume electron microscopy by Takemura et al. 2017. That paper details that only 93% of all synapses were connected to an arbor, and only 86% of the synapses had known pre- and postsynaptic arbors. For the MBON that was used for modeling, what was the fraction of terminal ends labeled as uncertain, and where these clustered or scattered across the arbor? Furthermore, the volume imaged with FIBSEM did not fully enclose the vertical lobe of the MB. Any estimate of what fraction of the chosen MBON's arbor is contained within the imaged volume? In other words, what analysis has been done here to ensure that the modeled arbor is representative of an MBON arbor in vivo, and what mitigating measures were taken to account for the potentially missing 14% or more of the arbor synapses and terminal dendrites?

The authors report using the 10XUAS-IVS-mCD8::GFP to label the MBON, so that they can then record electrophysiologically with patch-clamp. What is the effect of inserting so many mCD8 proteins (a large transmembrane protein) into the neuron's membrane on the voltage potential and action potential formation and transport? The 10XUAS is particularly strong. How does the morphology of the imaged neuron differ from that of the EM-reconstructed neuron, regarding calibers and amount of cable? For this purpose, a cytosolic GFP targeting the soma or nucleus and poorly diffusing into the arbor would have been far preferable, as the effect of inserting transmembrane proteins in neurons' membranes on resting potentials is well reported.

Line 155: average resting potential for the MBON is reported at -56.7 mV +/- 2.0 mV. In Hige et al. 2015a, cells were held at -70 or -60 mV. Nowhere does Hige et al. 2015a report on the resting potential.

Line 174: amplitude of action potentials was rather small, but in Hige et al. 2015a action potentials typically exceeded 200 pA. Is this what the authors mean by small? Just how small were the recorded action potentials?

Line 176: by small amplitudes and the explanation on the long neurite connecting the dendritic arbor with the soma, you mean that the signal is attenuated over long distances?

Line 179: how was the membrane capacitance calculated?

Table 1: 5 neurons were used. How do you know they are all MBON-α3? Has it been confirmed that the GAL4 line doesn't have stochastic expression among similar yet different sibling neurons of the same lineage? How many other MBONs innervate the tip of the α lobe and do any of them share neuroblast lineage with MBON-α3? The large differences in measured values listed in Table 1 could be explained by having measured similar yet different neurons. Did you run a battery of tests before and after the measurements to ensure the recorded neurons remained in good health throughout the measurement session? Such tests often consist in a ramp of current injections and the recording of the neuron's responses, which are then compared between before and after the experimental measurements of membrane properties (like the current step protocol of Figure 1F).

Line 184: why only 3 cells? In Drosophila, recording from e.g. 10 cells, all homologous cells across 10 individuals, gives e.g., 8 responding with excitation to a sensory stimuli with some variation and 2 responding with inhibition. There is a lot of variability in the responses. Recording from only 3 cells seems risky, statistically speaking. What justifies this low number?

Line 186: "To to further".

Line 199: when you say that the measurements are in "good agreement with prior recordings" of other neurons in Drosophila, what do you mean exactly? How similar, how far off, by what parameters?

Line 203: might as well mention that there were 948 reconstructed KCs synapting onto MBON-α3, so 5% is 50. Spare the reader remembering where the 50 was picked from. (If you correct this to 6%, would be 57 KCs). And you seem to not keep in mind that the KCs responding to a specific odor may be correlated in their synaptic connectivity strength onto MBON-α3. Data to this end may be included in Li et al. 2020 eLife where the whole mushroom body is reconstructed, including the olfactory projection neurons, so such correlations if any may be evident in that data set.

Line 210: a complete reconstruction of MBON-α3 now exists, from either the FAFB volume or the Hemibrain volume. In the methods you mention you used the Hemibrain data set for the axon.

Line 209: the 12,770 synaptic connections aren't "all", these are the ones reported from the anatomical reconstruction from volume electron microscopy. According to the source papers (Takemura et al. 2017) about 10% of all synapses are missing. An analysis of how these missing synapses impacts the structure of the arbor is absent from the paper.

In addition, sample preparation for electron microscopy with chemical fixation alters the fine anatomical details, including the length of terminal dendrites and the calibers of neurites throughout. See e.g., Korogod et al. 2015 eLife "Ultrastructural analysis of adult mouse neocortex comparing aldehyde perfusion with cryo fixation" and the follow up paper Tamada et al. 2020 eLife "Ultrastructural comparison of dendritic spine morphology preserved with cryo and chemical fixation". What measures were taken to correct or mitigate these artifactual differences with in vivo neurons?

Later, Figure 2F strongly supports the appropriateness of the model, yet, the above points merit discussion and even exploration: how much of the dendritic arbor can you miss and still get the same result? What does the response to current injection depend on, cable, number of synapses, synapse spatial location, cable calibers, tapering of cable? What cable truncations are tolerable? This is very important information towards future computational studies based on neuronal morphologies reconstructed from volume electron microscopy.

Figure 2 legend: what is the evidence that the "proximal neurite" in green in Figure 2B is the site of axon potential generation? Gouwens and Wilson 2009 pointed at a region anywhere between the root of the dendritic tree and half-way through the axon of the uniglomerular olfactory projection neuron they modeled.

Does the site of axon potential generation emerge from your model, or did you specify it in the model?

Why is the two-tailed non-parametric Spearman correlation the correct statistic to compare the modeled and the experimentally measured membrane potential in Figure 2F?

Figure 2 legend reads "see appendix" but there isn't any appendix to the manuscript?

Line 249: "the average number of synaptic contacts from a KC to this MBON is 13.47". This statement ought to be qualified: for the single MBON-α3 measured in Takemura et al. 2017, and with the caveat of ~10% of synapses potentially missing. You could just as easily apply a correction factor and say the average number is about 14.7 + 1.47 = 16.2. Would this change the outcome of your model?

Please don't use "PN" as an acronym for proximal neurite. First, eLife doesn't restrict the length of your test. Second, PN is an established acronym, universally across all neuroscience literature, for projection neuron. Plus, the "proximal neurite" (as per figure 2B) might as well be called the putative AIS (axon initial segment; pAIS for "putative") where the integration of inputs across the entire dendritic tree take place and the axon potential is initiated.

Figure 3H: in the measurement of "local dendritic section volume", did you correct for volume artifacts induced by using (in purpose!) an incorrect osmolarity of the buffers when fixing the tissue in the sample preparation protocol for electron microscopy? See Korogod et al. 2015 eLife and Tamada et al. 2020 eLife.

Line 291: "this value is in good agreement with in vivo data for MBONs". Please could you specify what this agreement is, how close, some details.

Line 293: in line with analyzing all KCs with exactly 13 synaptic inputs onto MBON-α3, what's the result of analyzing the voltage excursion from drawing random subsets of 13 synapses? (or 16 as per the correction, see above). Are the natural groups of 13 synapses different in their effect on the neuron's voltage than artificial groupings?

Line 299: inaccurate statement: "Given that ≈ 5% of the 984 KCs innervating MBON-α3 are typically activated by an odor". Instead, what is known from the literature is that, given the presence of the GABAergic neuron APL in the mushroom body which acts as the inhibitory unit of a winner-take-all configuration, only 6% (not 5%) of KCs simultaneously respond to any one odor. Plus when the APL neuron is inhibited, a huge double-digit percent of KCs are active in response to an odor.

Line 311: there is now far better evidence of stochastic odor encoding by KCs than Caron et al. 2013. See Zheng e t al. 2020 bioRxiv "Structured sampling of olfactory input by the fly mushroom body", and Li et al. 2020 eLife, and also, for larvae, Eichler et al. 2017 Nature.

Line 319: see also Baltruschat L, Prisco L, Ranft P, Lauritzen JS, Fiala A, Bock DD, Tavosanis G. Circuit reorganization in the Drosophila mushroom body calyx accompanies memory consolidation. Cell reports. 2021 Mar 16;34(11):108871.

Line 337: the differences in the somatic amplitudes may be significant statistically, but are they meaningful? In other words, the effect size looks like near zero. The real, and important difference, is in Line 345 where it is stated that "we observed some differences in the slope of the responses between the different tuning modalities (Figure 5J,K,L)."

Line 350: Scheffer et al. 2020 is not an appropriate citation for the statement "Te ability of an animal to adapt its behavior to a large spectrum of sensory information requires specialized neuronal circuit motifs". Rather, a textbook such as Kandel et al. Principles of Neural Science, or no reference at all, would be appropriate. You could also delete the sentence without loss.

Line 353: Drosophila must go in italics, it's a species name. Multiple occurrences throughout.

Line 356: through both short-term and long-term memory. A good example is Aso & Rubin 2016 eLife.

Line 359: note the olfactory system of the fly has a sort of "fovea", Zheng et al. 2020 bioRxiv.

Line 362: this sentence needs work, I am not sure what it means: "Individual flies display idiosyncratic, apparently random connectivity patterns that transmit information of specific odors to the output circuit of the MB."

Line 366: MBONs can only each be classified as approach or avoidance within specific behavioral paradigms. In different contexts, including different physiological states (e.g., hunger, satiation, others), the classification changes.

Line 381: prior work includes Tobin et al. 2017 eLife, where EM-reconstructed dendrites of olfactory projection neurons were modeled to understand the impact of dendritic arbor size on neuronal function.

Line 386: again, these numbers aren't precise. There's about 10% of missing synapses to consider, and potentially additional Kenyon cells. And some of the KCs, particularly those with low number of synaptic connections to MBON-α3, may have been connected in error.

Line 397: Strongest finding of this work: "The location of an individual synaptic input within the dendritic tree has therefore only a minor effect on the amplitude of the neuron's output, despite large variations of local dendritic potential." Would be best to surface it more.

Line 402: a comparison would be appropriate with neurons from the crayfish stomatogastric ganglion (STG) as described by Eve Marder lab's, with published findings such as neurons being electrotonically compact despite their large size in mature adult animals. For example, Otopalik et al. 2017 eLife "When complex neuronal structures may not matter", where the authors "quantify animal-to-animal variability in cable lengths (CV = 0.4) and branching patterns in the Gastric Mill (GM) neuron". And also Otopalik et al. 2019 eLife "Neuronal morphologies built for reliable physiology in a rhythmic motor circuit".

Line 406: you forgot to cite Jackie Schiller's work on cortical pyramidal cells and their tuft dendrites, some of which predates all the cited work in this statement.

Line 412: your model, as per Figure 2F, rather very closely matches the observed electrophysiological responses of the neuron under study. In what further ways could you model more closely match experimental observations? This would be very instructive to the reader.

Line 415: by ensembles of both KCS and MBONs, not just KCs, if you are including the memory part and not only the input representation part.

Line 416: for most of the paper, you quoted these papers to justify the 5% of KCs being simultaneously active in response to an odor. Now the range is shown as 3 to 9%. This discrepancy ought to be reconciled.

Line 422: again, nowhere in the manuscript so far did you detail in what way your simulation and experimental findings match those of prior experimental reports regarding neuron response and physiological properties.

Line 425: this statement is inaccurate. An individual KC does not encode for a single odor. That would almost never be the case even if a KC was single-claw, as in, exclusively integrated inputs from a single projection neuron: individual projection neurons rarely encode an odor; it's the population of projection neurons that does. Similarly, ensembles of co-active KCs together represent an odor, and do so more narrowly and accurately than the population of projection neurons that excited them.

Line 452: left unresolved remains the question of why would both mechanisms exist, as in, is the combination of altering the KC-MBON synapse and altering the PN-KC synapse better in some dimension than altering either alone? Does this relate perhaps to a possible dynamic nature of the olfactory "fovea" proposed by Zheng et al. 2020 bioRxiv as presumably static?

Line 466: what is standard fly food? Please define it. Choice of food affects very much behavioral assays, for example.

Line 476: no antigen saturation steps in the immunohistochemistry protocol? Please revise.

Line 505: what were the settings of the puller? These would be necessary to reproduce your glass capillaries. Did you grind the tips, and if so, how?

Line 534: how were the R_series, R_input and R_m determined with Igor Pro?

Line 561: why was a threshold of 600 MΩ used to exclude cells? How many cells were recorded in total, and how many were excluded?

Line 584: once again the data is the best human effort in proofreading a semiautomatic segmentation. It is not the absolute truth. Also, each individual fly is somewhat different, so this is merely one fairly complete yet partial reconstruction, and not assured to be error free, particularly of errors of omission, of a single MBON from a single individual. Would be appropriate to remark this clearly.

Did you provide the NeuroML or similar files necessary to run the model in the NEURON simulation software? These should be appended to the manuscript as supplemental data.

Line 596: if "parameters were tuned until the computed voltage excursion at the soma matched our electrophysiological recordings", what is the rationale for comparing the voltage potential of the simulation with those of the experimental observations?

Line 607: where do these ranges of physiologically plausible values come from? Citation, or measurements done in the lab and therefore a figure is needed to show them? Are these ranges from the literature listed in Table 2? Likely yes, would be appropriate to cite them here too or at least explicitly point to Table 2.

Line 616: innervation of MBONs is cholinergic because KCs are cholinergic, but that's not from Takemura et al. 2017, instead from Barnsted et al. 2016 Neuron.

Line 622: would be appropriate to include the python scripts used to configure and run the NEURON simulation as supplemental material. Likewise for the matlab scripts used to load the analyzed data and plot it. The raw data ought to be included as CSV files or similar.

Reviewer #2 (Recommendations for the authors):

"The cellular architecture of memory modules in Drosophila supports stochastic input integration" is a classical biophysical compartmental modelling study. It takes advantage of some simple current injection protocols in a massively complex mushroom body neuron called MBON–a3 and compartmental models that simulate the electrophysiological behaviour given a detailed description of the anatomical extent of its neurites.

This work is interesting in a number of ways:

– The input structure information comes from EM data (Kenyon cells) although this is not discussed much in the paper.

– The paper predicts a potentially novel normalization of the throughput of KC inputs at the level of the proximal dendrite and soma.

– It claims a new computational principle in dendrites, this didn't become very clear to me.

Problems I see:

– The current injections did not last long enough to reach steady state (e.g. Figure 1FG), and the model current injection traces have two time constants but the data only one (Figure 2DF). This does not make me very confident in the results and conclusions.

– The time constant in Table 1 is much shorter than in Figure 1FG?

– Related to this, the capacitance values are very low maybe this can be explained by the model's wrong assumption of tau?

– That latter in turn could be because of either space clamp issues in this hugely complex cell or bad model predictions due to incomplete reconstructions, bad match between morphology and electrophysiology (both are from different datasets?), or unknown ion channels that produce non–linear behaviour during the current injections.

– The PRAXIS method in NEURON seems too ad hoc. Passive properties of a neuron should probably rather be explored in parameter scans.

Questions I have:

– Computational aspects were previously addressed by e.g. Larry Abbott and Gilles Laurent (sparse coding), how do the findings here distinguish themselves from this work.

– What is valence information?

– It seems that Martin Nawrot's work would be relevant to this work.

– Compactification and democratization could be related to other work like Otopalik et al. 2017 eLife but also passive normalization. The equal efficiency in line 427 reminds me of dendritic/synaptic democracy and dendritic constancy.

– The morphology does not obviously seem compact, how unusual would it be that such a complex dendrite is so compact?

– What were the advantages of using the EM circuit?

– Isn't Figure 4E rather trivial if the cell is compact?

Overall, I am worried that the passive modelling study of the MBON–a3 does not provide enough evidence to explain the electrophysiological behaviour of the cell and to make accurate predictions of the cell's responses to a variety of stochastic KC inputs.

Reviewer #3 (Recommendations for the authors):

This manuscript presents an analysis of the cellular integration properties of a specific mushroom body output neuron, MBON-α3, using a combination of patch clamp recordings and data from electron microscopy. The study demonstrates that the neuron is electrotonically compact permitting linear integration of synaptic input from Kenyon cells that represent odor identity.

Strengths of the manuscript:

1) The study integrates morphological data about MBON-α3 along with parameters derived from electrophysiological measurements to build a detailed model.

2) The modeling provides support for existing models of how olfactory memory is related to integration at the MBON.

Weaknesses of the manuscript:

1) The study does not provide experimental validation of the results of the computational model.

2) The conclusion of the modeling analysis is that the neuron integrates synaptic inputs almost completely linearly. All the subsequent analyses are straightforward consequences of this result.

3) The manuscript does not provide much explanation or intuition as to why this linear conclusion holds.

In general, there is a clear takeaway here, which is that the dendritic tree of MBON-α3 in the lobes is highly electrotonically compact. The authors did not provide much explanation as to why this is, and the paper would benefit from a clearer conclusion. Furthermore, I found the results of Figures 4 and 5 rather straightforward given this previous observation. I am sceptical about whether the tiny variations in, e.g. Figures 3I and 5F-H, are meaningful biologically.

1) My biggest question is about the claim of extreme electrotonic compactness of this neuron. Figure 3D,E suggests that the voltage change at the proximal neurite and at the soma varies by only about 1% depending on stimulation location. Since this is supported only by simulation, it is worth asking how robust this conclusion is.

a) Given that the variability in 3I is so small in magnitude, any dependence would be swamped by other sources of heterogeneity, so the statement that this correlates with distance (line 278) is likely irrelevant.

b) Can the authors provide a confidence interval for the fit of biophysical parameters to their recordings?

c) On lines 271-274, the authors state, "This architecture with the smallest dendritic sections at the most distant sites may contribute to the compactness of the dendritic tree, ensuring that even the most distant synaptic inputs result in somatic voltage deflections comparable to the most proximal ones." Is a dependence of dendritic size on distance required for the results? It seems like the result is simply that there is no attenuation within the dendritic tree at all. In general, the authors don't actually provide an explanation for the compactness. In the discussion, it is stated that, "The compactification of the neuron is likely related to the architectural structure of its dendritic tree," which again is rather vague. The authors should strive to provide a clear explanation for this, since it is their key result.

d) Can the authors report what the electrotonic length for such a dendrite would be? How long before we expect to see a significant spread in 3E?

2) My other concern is that, once we assume this perfect integration, the subsequent analyses are all a straightforward consequence. In particular, Figures 4 and 5 just repeatedly convey that it doesn't matter which synapse is being activated, the effect on the somatic voltage is the same. I was particularly confused about the conclusions of 5F,G,H. The authors claim "small but significant differences" here, but practically speaking, I can't imagine any of the differences in these plots being meaningful.

3) Reference to the data used to constrain the model is confusing.

a) At various places in the manuscript references are made to in vivo recordings, but it appears that all of the recordings were done ex vivo.

b) On lines 389-392, the authors state: "Near-perfect agreement between experimentally observed and simulated voltage distributions in the dendritic tree shows that linear cable theory is an excellent model for information integration in this system." What recordings of the voltage distribution in the dendritic tree were performed?

4) It seems like the conclusions are different than those of Gouwens and Wilson (2009), who described their reconstructed PNs as electrotonically extensive. The authors should comment on what about MBONs and PNs is different.

https://doi.org/10.7554/eLife.77578.sa1

Author response

Reviewer #1 (Recommendations for the authors):

Hafez and collaborators describe the construction and analysis of a computational model of a mushroom body neuron. The anatomy derives from a combination of electron microscopy reconstructions of MBON-α3 and also from light microscopy. The physiological parameters derive from publications that measured them, in addition to the author's own electrophysiological recordings with patch-clamp.

There are two main findings. First, the dendritic arbor of MBON-α3 is electrotonically compact, meaning, individual connections from Kenyon cells will similarly elicit action potentials independently as to where, spatially, the synapses lay on the arbor. Second, in simulation, exploration of changes in the strength of Kenyon cell inputs illustrate two possible ways to alter the strength of the KC-MBON physiological connection, showing that either could account for the observed synaptic depression in the establishment of associative memories. The properties of each approach differ.

Overall, the manuscript clearly describes the journey from connectomics and electrophysiology to computational modeling and exploration of the physiological properties of a circuit in simulation.

The discussion ought to be expanded to include the implications of two possible approaches to physiologically altering the KC-MBON synapse and the consequence of their combination in expanding the space of alterations induced by associative memory paradigms.

We now extended this discussion. However, as we do not provide any experimental evidence supporting either one of these hypothesis, we think that this point would be better addressed in a review format as we can only speculate at this point of time.

In general, the results are clear, but some details remain underdetermined and I have listed them below in the detailed comments. The introduction and discussion present some inaccuracies that can be swiftly addressed by the authors

Detailed comments:

Line 45: Language: "potential rewarding": potentially.

We have changed the text accordingly.

Line 49: Language: "cellular and circuit architecture contributes": architectures contribute.

We have changed the text accordingly.

Line 72: instead of 5%, the number of KCs active at any one time seems to be 6% as per Turner et al. 2008 and Campbell et al. 2013. What is the robustness of the analysis to this small change? Did you explore a range of possible single-digit percent KC activations?

We now state the number as 3-9% of KCs and highlight that different studies observed reliable calcium signals in 5 or 6% of KC upon odor stimulation. Our own results (Siegenthaler et al. 2019) indicate that approximately 5% of KCs are activated by each odor and we used this number as an approximation for our simulations. It is correct, however, that the single-cell study by Turner et al. (2008) reported an average 6% of activated KCs for each odor. A later optogenetical study by the same group (Honegger et al., 2011) reported 5%; importantly for the present discussion, they however considered this result "extremely similar" to their earlier results (their page 11,777, left column). We take from their formulation that they consider the difference between 5% and 6% as very small It would then make sense to assume that small differences obtained in different experimental preparations, on the order of a percentage points, as a kind of noise, and to evaluate the robustness of results under such perturbations. However, in absolute terms this difference by one percentage point, between 5% and 6%, represents a fifth difference (i.e., 20%) in the synaptic input to the MBON. And, of course, in our results this difference can not be due to different experimental conditions since they are obtained in the same "experimental preparations," namely the same simulations. To answer directly the Reviewer’s question, we have considered a range of single-digit percent KC activations, however our "range" only had 3 entries. We have not simulated a 20% change of the number of activated synapse but we did simulate a slightly larger increase, by 25% (our Figure 5B). Our results show that the 25% increase leads to a substantial change in the somatic MBON voltage (Figure 5B,D). Given the nearly linear dependence of the voltage increase with the number of synapses (Figure 5I), we are confident that a 20% increase can be obtained with high precision from a linear interpolation based on the 25% increase (as can be any other increases within the range we cover with our three values for the number of synapses, i.e. 38 to 63 KCs, Figure 5). An increase from 5% to 6% of activated synapses therefore increases the voltage by about 80% of the value shown in Figure 5B, D which is a substantial difference. We would consider such a difference not a question of robustness but as a change that is likely physiologically relevant.

We have added a short discussion of this topic in the revised manuscript (line 340ff, at the very beginning of Section 2.4)

Line 99: Aso & Rubin 2016 belongs with citations in line 95.

Many thanks for pointing this out. We moved the citation to the correct place.

Line 122: "Drosophila" needs italics, throughout the manuscript.

All instances of "Drosophila" are now set in italics.

The authors devise a computational model of an MBON using a neuronal arbor reconstructed from volume electron microscopy by Takemura et al. 2017. That paper details that only 93% of all synapses were connected to an arbor, and only 86% of the synapses had known pre- and postsynaptic arbors. For the MBON that was used for modeling, what was the fraction of terminal ends labeled as uncertain, and where these clustered or scattered across the arbor?

The reviewer rightfully highlights the limitations of the EM-dataset that we used for our reconstruction. Indeed, Takemura and colleagues report that not all postsynaptic sites within the α-lobe could be assigned to identified cells. Importantly, however, the authors state that these synaptic sites very likely belong to identified cells within the dataset. Thus, for our dataset this would mean that the MBON itself may have up to 7 percent more input synapses but these synapses would belong to the identified 948 KCs. An individual KC may therefore at most have one or two additional input synapses, e.g. KCs currently annotated with 13 input synapses may form 14 input synapses. These differences would not significantly change any of the conclusions of our manuscript. We now addressed this point experimentally and analyzed the activation of KCs with 12, 13 and 14 input synapses and compared this to the activation of 13 random synapses (Figure 4F). We observed significant differences in MBON responses when changing the number of input synapses but not between KCs with 13 synapses and the activation of 13 random synapses. Thus, the number of input clearly changes postsynaptic responses, however, as we assume that these non-annotated synapses are likely uniformly distributed, these changes would not affect our physiological simulations of different sets of 50 KCs. Unfortunately, Takemura and colleagues did not specify the location of these non-annotated sites and we simply cannot re-trace the entire dendritic tree of the MBON. As they state however that "the small size and discontinuous nature of these neurites indicates that they are fragments of identified cells rather than collectively constituting an additional cell type", we assume that these sites were not clustered.

A similar rate of "tracing shortcomings" has been observed in the hemibrain dataset and in the current FlyWire approach. We decided that the most conservative approach would be to use the data as published as any modification, e.g. addition of 10% of synapses, would make later reproductions of our data more difficult. We now state the potential problems clearly in our paper (line 206ff, section 2.2).

Furthermore, the volume imaged with FIBSEM did not fully enclose the vertical lobe of the MB. Any estimate of what fraction of the chosen MBON's arbor is contained within the imaged volume? In other words, what analysis has been done here to ensure that the modeled arbor is representative of an MBON arbor in vivo, and what mitigating measures were taken to account for the potentially missing 14% or more of the arbor synapses and terminal dendrites?

Indeed the dataset of Takemura et al. did not include the entire vertical lobe of the MB. However, it included the entire α-lobe of the MB – only the α prime lobe is missing. As beautifully shown by Takemura et al., the entire dendritic tree of MBON-3alpha that only innervates the α-lobe is included and presented in Figure 2H and videos 4, 6 and 7 with the position of synaptic input sites. The striking visualization of Takemura and colleagues of this dendritic tree was one of the main reasons why we chose this MBON as an example for our combined in vivo and in silico analysis. While Takemura states that they could only precisely identify both pre- and postsynaptic partners for 86% of all synapses, this does not necessarily mean that 14% of KC to MBON synapses are missing in the dataset as e.g. only 7% of the postsynaptic sites remained unassigned (and these are the relevant ones for our MBON). We again carefully compared the annotations of KC to MBON-alpha3 synapses in Takemura et al. and Scheffer et al. (Hemibrain dataset) to evaluate the effects of a potential under-representation of synapses within the dendritic compartment of MBON-alpha3. Interestingly, the numbers for both MBON-alpha3 cells in the hemibrain dataset are below the numbers of Takemura et al. Specifically the numbers are for Takemura: MBON-alpha3A 12,770 synaptic connections from the 948 KCs (average 13.47 synapses), MBON-alpha3B 13,129 from 948 KCs (average 13.85 synapses) vs Scheffler: MBON-alpha3A 11,950 (+1610 which are not classified = 13,560) synapses from 894 KCs (average 12.44 synapses), MBON-alpha3B 11,121 (+ 1,572 not classified = 12,693) synapses from from 897 KCs (average 13.32 synapses).

These numbers are highly similar between the two data sets and also between the two different MBON-alpha3 neurons. Importantly however, the numbers are lower in the Hemibrain data set. Thus, the variation between individual flies (or data sets) seems to be in the range of 10%. As such, the Takemura dataset likely represents a very good approximation despite some shortcomings due to either the EM fixation (see below) or due to annotation problems. As we did not want to introduce any arbitrary mistakes in our analysis we chose to adhere to the numbers provided in the Takemura data set. We now clearly state the potential shortcomings of using a single EM data set as the basis for a simulation (line 206ff). We still think that this is currently the best possible approach.

The authors report using the 10XUAS-IVS-mCD8::GFP to label the MBON, so that they can then record electrophysiologically with patch-clamp. What is the effect of inserting so many mCD8 proteins (a large transmembrane protein) into the neuron's membrane on the voltage potential and action potential formation and transport? The 10XUAS is particularly strong. How does the morphology of the imaged neuron differ from that of the EM-reconstructed neuron, regarding calibers and amount of cable? For this purpose, a cytosolic GFP targeting the soma or nucleus and poorly diffusing into the arbor would have been far preferable, as the effect of inserting transmembrane proteins in neurons' membranes on resting potentials is well reported.

We would like to thank the reviewer for this suggestion. We now performed additional recordings with a cytosolic GFP construct. The data from 4 additional recordings is now presented in Figure 1 Supplement 2. In this new Figure we now provide a direct comparison between the mCD8-GFP based recordings and the cytosolic GFP based recordings and present all raw data traces. This analysis demonstrated that the averaged traces of the new recordings fall within the data range of our initial dataset (Figure 1 Supplement 2 I). Thus, the source of GFP used to label these cells did not have any significant consequences for the recordings. In addition, with this new data set we could clearly demonstrate that our electrophysiological data are reproducible and that the original data set provides accurate values for our computer simulations.

Line 155: average resting potential for the MBON is reported at -56.7 mV +/- 2.0 mV. In Hige et al. 2015a, cells were held at -70 or -60 mV. Nowhere does Hige et al. 2015a report on the resting potential.

The average resting membrane potential of the recorded MBONs was at -56.7 mV +/- 2.0 mV. To our knowledge, we are the first to report precise electrophysiological values for this particular cell type. It is correct that Hige et al., 2015a do not report values however, it is possible to estimate these values from the raw traces provided in their publication. Importantly, our value of approximately -60 mV is in line with data from a number of central neurons in Drosophila (e.g. Gu and O´Dowd, 2006; Wilson et al., 2004; Groschner et al., 2022). We now state more appropriately (line 154ff): This value for the restring membrane potential is in line with prior measurements of MBONs in vivo and of other central neurons in Drosophila (Hige et al., 2015a, Gu and O´Dowd, 2006; Wilson et al., 2004; Groschner et al., 2022)

Line 174: amplitude of action potentials was rather small, but in Hige et al. 2015a action potentials typically exceeded 200 pA. Is this what the authors mean by small? Just how small were the recorded action potentials?

The amplitude of the action potentials (4.3 ± 0.029 mV) represent relatively low values compared to previously published data from different cell types in the central brain of the adult fruit fly (see references above). Even between different MBONs a high variability in action potential amplitudes has been observed (Hige et al., 2015a). While precise values were not reported in this publication, based on the traces the values range between very small (gamma2), 10 mV (alpha2sc) and 25 mV (gamma1). The small amplitudes observed for MBON alpha3 are likely a consequence of the unipolar morphology of the neurons with a very long neurite connecting the dendritic input region to the cell soma. This morphology is especially pronounced for MBON alpha3 (see Figure 1A).

Line 176: by small amplitudes and the explanation on the long neurite connecting the dendritic arbor with the soma, you mean that the signal is attenuated over long distances?

Yes, due to the very long neurite the signal is attenuated significantly.

Line 179: how was the membrane capacitance calculated?

Membrane capacitance was determined by dividing tau by the membrane resistance (Cm = tau/Rm). We now state this in the text (line 179).

Table 1: 5 neurons were used. How do you know they are all MBON-α3? Has it been confirmed that the GAL4 line doesn't have stochastic expression among similar yet different sibling neurons of the same lineage? How many other MBONs innervate the tip of the α lobe and do any of them share neuroblast lineage with MBON-α3? The large differences in measured values listed in Table 1 could be explained by having measured similar yet different neurons. Did you run a battery of tests before and after the measurements to ensure the recorded neurons remained in good health throughout the measurement session? Such tests often consist in a ramp of current injections and the recording of the neuron's responses, which are then compared between before and after the experimental measurements of membrane properties (like the current step protocol of Figure 1F).

The MB082C split-Gal4 line we used for our recordings is a highly specific driver line which is only expressed in the two MBON-alpha3 neurons per hemisphere across all observed preparations. These neurons do not share any overlap (position of dendrites plus position of soma) with any other MBON. Thus we are certain that all our recordings were performed on MBON-alpha3 neurons. However, the two MBON-alpha3 neurons innervating the same position in the α lobe are indistinguishable from each other. As there is no possibility to differentiate between the two and as their morphological features are almost identical (see numbers in our answer to the EM dataset above) we treated them simply as MBON-alpha3. We now state this clearly in our manuscript. As we are certain about the identity of the MBON, any differences likely result from inter-individual biological differences.

Line 184: why only 3 cells? In Drosophila, recording from e.g. 10 cells, all homologous cells across 10 individuals, gives e.g., 8 responding with excitation to a sensory stimuli with some variation and 2 responding with inhibition. There is a lot of variability in the responses. Recording from only 3 cells seems risky, statistically speaking. What justifies this low number?

We agree. We now added 4 additional recordings using a cytoplasmic GFP as a labelling source to the manuscript (Figure 1 Supplement 2, Table 1 Source data 1). These values of these recordings fall in between the values of our prior recordings and thus validate our prior results. We now have 7 detailed recordings for these cells that likely show the range of biological variation. It is technically extremely challenging to perform these recordings.

Line 186: "To to further".

We corrected this mistake.

Line 199: when you say that the measurements are in "good agreement with prior recordings" of other neurons in Drosophila, what do you mean exactly? How similar, how far off, by what parameters?

We revised this section. We now simply report the results of our recordings and state that MBONalpha3 is a spike-frequency adapting neuron like other MBONs. We removed the prior sentence with comparisons to other neurons.

Line 203: might as well mention that there were 948 reconstructed KCs synapting onto MBON-α3, so 5% is 50. Spare the reader remembering where the 50 was picked from. (If you correct this to 6%, would be 57 KCs). And you seem to not keep in mind that the KCs responding to a specific odor may be correlated in their synaptic connectivity strength onto MBON-α3. Data to this end may be included in Li et al. 2020 eLife where the whole mushroom body is reconstructed, including the olfactory projection neurons, so such correlations if any may be evident in that data set.

We now state directly that the number 50 is deduced from the 5% of KCs at the appropriate position in the text.

The reviewer is certainly correct that there is now clear evidence that the PN>KC input is not entirely random. Importantly, Li et al. demonstrated that inputs from different sensory modalities are especially segregated – however, MBON-alpha3/14 receives almost exclusively olfactory input (>97% as deduced from Figure 15, Supplement 2). In their models Li et al., still find some evidence for non-random sampling at the PN>KC synapse but this was modest at best and they observed only minor persistance to the KC>MBON level. In addition Zheng et al., 2022 demonstrate that some specific odors may be overrepresented compared to others to enable preferential encoding of naturally meaningful stimuli. In our first analysis presented here we decided to focus on the robustness of encoding principles by assuming random connectivity. We certainly aim to pursue a targeted analysis in a future study focussing on multiple MBONs.

Line 210: a complete reconstruction of MBON-α3 now exists, from either the FAFB volume or the Hemibrain volume. In the methods you mention you used the Hemibrain data set for the axon.

In the FAFB volume our MBON is not fully reconstructed. The reconstructions from Takemura and the Hemibrain are very similar and consistent regarding the number of KC>MBON synapses (see above). As we started our first analysis before the publication of the Hemibrain focusing only on the dendritic tree (Hafez et al., 2019, bioRxiv) we build our current study on this analysis and complemented the morphology reconstruction with data from the Hemibrain.

Line 209: the 12,770 synaptic connections aren't "all", these are the ones reported from the anatomical reconstruction from volume electron microscopy. According to the source papers (Takemura et al. 2017) about 10% of all synapses are missing. An analysis of how these missing synapses impacts the structure of the arbor is absent from the paper.

Please see our responses above to this topic. We now highlight these issues in the results when first describing our approach (line 206ff).

In addition, sample preparation for electron microscopy with chemical fixation alters the fine anatomical details, including the length of terminal dendrites and the calibers of neurites throughout. See e.g., Korogod et al. 2015 eLife "Ultrastructural analysis of adult mouse neocortex comparing aldehyde perfusion with cryo fixation" and the follow up paper Tamada et al. 2020 eLife "Ultrastructural comparison of dendritic spine morphology preserved with cryo and chemical fixation". What measures were taken to correct or mitigate these artifactual differences with in vivo neurons?

Korogod et al. showed that the main effect of fixating somatosensory mouse cortex using the standard aldehyde method was a substantial decrease of the extracellular space and an increase of astrocytic volume, compared to a ’fresh’ (not fixated) preparation. This was not the case, however, for neurites. They found that "the volume fraction occupied by axons and dendrites was similar between the fixation conditions" (their p 3).

There was also no indication of systematic loss or erroneous addition of synapses (the shape of synaptic vesicles was modified but we do not make use of this feature in our study). Synaptic density in the chemically fixed neuropil was 38% higher than cryo-fixated neuropil, i.e. they measured 38% more synapses per volume. This is slightly lower than the total tissue shrinkage they observed (1/0.7 43%), a difference they explained by the fact that there are no synapses in structures like cell bodies and blood vessels. It seems therefore likely that the increase of the synapse density is primarily due to a decrease in the denominator (the volume) due to the loss of recorded extracellular space, rather than in the numerator (number of synapses).

Given that the main effect of chemical fixation is a sharp reduction of extra-cellular space, with no significant changes of neurite dimensions, we feel that the EM data we use are at least a good first approximation of the physiological state. A previous study of Drosophila neurons from Rachel Wilson’s lab (Tobin et al. 2017, https://doi.org/10.7554/eLife.24838) came to a similar conclusion.

Tamada et al. focused on fixation-induced morphology changes of dendritic spines, again in mouse cerebral cortex. They found that the length of the spines as well as the spine head volume was not significantly changed by traditional (chemical) fixation preceding EM. They did find that chemical fixation resulted in significantly (by 42%) thicker spine necks, however. Assuming identical cytoplasmatic specific resistivity, using the measured (per chemical fixation) spine neck diameter would result in significantly lower resistance between spine head and dendrite than the actual physiological value.

Spine-like structures have been observed in the Drosophila nervous system (e.g. Leiss et al. "Characterization of dendritic spines in the Drosophila central nervous system." Developmental neurobiology 69.4 (2009): 221-234). However, the relevance of these results for our model is unknown. In the published data no spine-like structures or locations on the dendritic tree of MBON-α3 have been characterized; so we are not able to evaluate details of the impact that different diameters of the spine necks would have. In the absence of such information, we can not address (potential) differences in the resistance between individual synaptic contacts and the dendrites. Instead, we fit a common value for the cytoplasmatic resistivity for the whole neuron (our Table 2). We agree that it would, indeed, be very interesting to study the effect of synapses on spines (for those neurons where they exist) compared to that of synapses located directly on the dendritic surface, but this requires the availability of data characterizing location and, ideally, geometry of dendritic spines on neurons of interest. Due to these reasons we did not include any correction factor in our analysis.

Later, Figure 2F strongly supports the appropriateness of the model, yet, the above points merit discussion and even exploration: how much of the dendritic arbor can you miss and still get the same result? What does the response to current injection depend on, cable, number of synapses, synapse spatial location, cable calibers, tapering of cable? What cable truncations are tolerable? This is very important information towards future computational studies based on neuronal morphologies reconstructed from volume electron microscopy.

We agree that these are important points. It is important to note that we provide results for our simulations at two different points. First, at the proximate neurite, the "exit" point of the dendritic tree and potential site of action potential initiation, and second at the soma region. Importantly, the region of the proximate neurite is still within the EM-dataset of Takemura et al. – thus, no relevant dendritc arbor sections are missing and no assumptions regarding the diameter of any neuronal structures have been made. We provide a direct comparison to the soma part that is based on a different dataset and approximations as it is currently not included in any EM dataset (the upcoming FlyWire dataset will hopefully resolve this issue). Our comparison demonstrates very few differences between these two points of analysis indicating that our dataset is very robust – but of course the main computation occurs within the dendritic compartment that was not modified in our analysis (see also responses above for our reasoning not to perform any corrections). We therefore cannot draw any significant conclusions regarding physical alterations of the dendritic compartment and would like to stick to our "conservative" approach for this initial study. We will certainly test more challenging models in future studies.

Figure 2 legend: what is the evidence that the "proximal neurite" in green in Figure 2B is the site of axon potential generation? Gouwens and Wilson 2009 pointed at a region anywhere between the root of the dendritic tree and half-way through the axon of the uniglomerular olfactory projection neuron they modeled.

We currently do not have any evidence that this is the position of action potential initiation. Based on immunohistochemical analyses of the Drosophila para channel in other neurons the position is likely in close proximity to the dendritic arbor. Here, we used this position for our comparative analysis with the soma as this region was still confined within the Takemura EM reconstruction – we state this now more clearly in the revised manuscript in the legend to Figure 2. As we observed consistent results between the two regions it would not make a significant difference if this region would be moved. In future work we hope to precisely determine the site of action potential generation using genetic methods.

Does the site of axon potential generation emerge from your model, or did you specify it in the model?

Please see above, we specified the site in the model.

Why is the two-tailed non-parametric Spearman correlation the correct statistic to compare the modeled and the experimentally measured membrane potential in Figure 2F?

Many thanks for highlighting this. This comparison is not necessary as the fitting was done in NEURON using the mean squared error. We now state this appropriately in the Methods section.

Figure 2 legend reads "see appendix" but there isn't any appendix to the manuscript?

We changed this and now only refer to the figure supplement.

Line 249: "the average number of synaptic contacts from a KC to this MBON is 13.47". This statement ought to be qualified: for the single MBON-α3 measured in Takemura et al. 2017, and with the caveat of ~10% of synapses potentially missing. You could just as easily apply a correction factor and say the average number is about 14.7 + 1.47 = 16.2. Would this change the outcome of your model?

Please see our discussion of this issue above. We now clearly state that 10% of synapses potentially missing at the beginning of the section and provide our reasoning. We also now tested the responses to 12, 13 or 14 synapses in Figure 4F.

Please don't use "PN" as an acronym for proximal neurite. First, eLife doesn't restrict the length of your test. Second, PN is an established acronym, universally across all neuroscience literature, for projection neuron. Plus, the "proximal neurite" (as per figure 2B) might as well be called the putative AIS (axon initial segment; pAIS for "putative") where the integration of inputs across the entire dendritic tree take place and the axon potential is initiated.

Many thanks for pointing this out. We now use PN as an acronym for projection neuron and do not abbreviate proximal neurite. We now also highlight that this is the likely site of action potential initiation. However, as this is not based on experimental evidence we would like to remain conservative and refer to it only as the proximal neurite.

Figure 3H: in the measurement of "local dendritic section volume", did you correct for volume artifacts induced by using (in purpose!) an incorrect osmolarity of the buffers when fixing the tissue in the sample preparation protocol for electron microscopy? See Korogod et al. 2015 eLife and Tamada et al. 2020 eLife.

Please see our response to potential fixation artifacts above.

Line 291: "this value is in good agreement with in vivo data for MBONs". Please could you specify what this agreement is, how close, some details.

Many thanks for pointing this out. We now removed this sentence as we only deduced this value from odor evoked activations in vivo that activate approximately 5% of KCs and result in action potential generation. Thus a direct comparison is currently not possible.

Line 293: in line with analyzing all KCs with exactly 13 synaptic inputs onto MBON-α3, what's the result of analyzing the voltage excursion from drawing random subsets of 13 synapses? (or 16 as per the correction, see above). Are the natural groups of 13 synapses different in their effect on the neuron's voltage than artificial groupings?

Many thanks for suggesting this very interesting experiment. We now performed the experiment and provide a comparison between KCs with exactly 12, 13 and 14 synapses to a random set of 13 synapses (Figure 4F). This experiment demonstrated no significant difference between the activation of KCs with 13 synapses and 13 random synapses. Interestingly though, slightly more variation could be observed in the dataset of KCs with 13 synapses. This may indicate some biological relevance however, the differences compared to activating one additional synapse are negligible in line with our model.

Line 299: inaccurate statement: "Given that ≈ 5% of the 984 KCs innervating MBON-α3 are typically activated by an odor". Instead, what is known from the literature is that, given the presence of the GABAergic neuron APL in the mushroom body which acts as the inhibitory unit of a winner-take-all configuration, only 6% (not 5%) of KCs simultaneously respond to any one odor. Plus when the APL neuron is inhibited, a huge double-digit percent of KCs are active in response to an odor.

We now corrected this statement and highlight that only 5-6% (range 3-9%) of neurons reliably respond to individual odors. This is the value that is likely relevant for memory association. As we do not address the inhibitory role of the APL neuron in our current study we would like keep our study focused and potentially explore these issues in a future study.

Line 311: there is now far better evidence of stochastic odor encoding by KCs than Caron et al. 2013. See Zheng e t al. 2020 bioRxiv "Structured sampling of olfactory input by the fly mushroom body", and Li et al. 2020 eLife, and also, for larvae, Eichler et al. 2017 Nature.

We are sorry for our oversight and now include these citations.

Line 319: see also Baltruschat L, Prisco L, Ranft P, Lauritzen JS, Fiala A, Bock DD, Tavosanis G. Circuit reorganization in the Drosophila mushroom body calyx accompanies memory consolidation. Cell reports. 2021 Mar 16;34(11):108871.

We added this citation.

Line 337: the differences in the somatic amplitudes may be significant statistically, but are they meaningful? In other words, the effect size looks like near zero. The real, and important difference, is in Line 345 where it is stated that "we observed some differences in the slope of the responses between the different tuning modalities (Figure 5J,K,L)."

We agree that the observed differences may not be biologically relevant but we can not formally exclude this possibility at the moment. We now highlight the differences in the slopes of the responses further in the discussion.

Line 350: Scheffer et al. 2020 is not an appropriate citation for the statement "Te ability of an animal to adapt its behavior to a large spectrum of sensory information requires specialized neuronal circuit motifs". Rather, a textbook such as Kandel et al. Principles of Neural Science, or no reference at all, would be appropriate. You could also delete the sentence without loss.

We agree and deleted the sentence.

Line 353: Drosophila must go in italics, it's a species name. Multiple occurrences throughout.

As mentioned above, "Drosophila" is now set in italics throughout.

Line 356: through both short-term and long-term memory. A good example is Aso & Rubin 2016 eLife.

Many thanks for this useful suggestion – we adjusted the text accordingly.

Line 359: note the olfactory system of the fly has a sort of "fovea", Zheng et al. 2020 bioRxiv.

As the study is now peer-reviewed and published we included it in the reference list and altered the statement to reflect the fact that processing of odor information does not exclusively depend on stochastic connectivity (line 65ff) – in addition we refer to this paper in our discussion of the two possible plasticity modes (line 504ff).

Line 362: this sentence needs work, I am not sure what it means: "Individual flies display idiosyncratic, apparently random connectivity patterns that transmit information of specific odors to the output circuit of the MB."

We agree and deleted the sentence.

Line 366: MBONs can only each be classified as approach or avoidance within specific behavioral paradigms. In different contexts, including different physiological states (e.g., hunger, satiation, others), the classification changes.

Yes, we now state this more appropriately in the revised version of the manuscript (line 409ff).

Line 381: prior work includes Tobin et al. 2017 eLife, where EM-reconstructed dendrites of olfactory projection neurons were modeled to understand the impact of dendritic arbor size on neuronal function.

Many thanks for pointing this out. We now cite this work.

Line 386: again, these numbers aren't precise. There's about 10% of missing synapses to consider, and potentially additional Kenyon cells. And some of the KCs, particularly those with low number of synaptic connections to MBON-α3, may have been connected in error.

We now highlight this point in the discussion and state (line 432ff): "While this EM reconstruction may contain some mistakes in synaptic connectivity as e.g. up to 10% of synaptic sites remained unassigned within the dataset (Takemura et al., 2017), it currently represents the best possible template for an in silico reconstruction."

Line 397: Strongest finding of this work: "The location of an individual synaptic input within the dendritic tree has therefore only a minor effect on the amplitude of the neuron's output, despite large variations of local dendritic potential." Would be best to surface it more.

We agree. We think that the importance of this finding is appropriately highlighted in the current form.

Line 402: a comparison would be appropriate with neurons from the crayfish stomatogastric ganglion (STG) as described by Eve Marder lab's, with published findings such as neurons being electrotonically compact despite their large size in mature adult animals. For example, Otopalik et al. 2017 eLife "When complex neuronal structures may not matter", where the authors "quantify animal-to-animal variability in cable lengths (CV = 0.4) and branching patterns in the Gastric Mill (GM) neuron". And also Otopalik et al. 2019 eLife "Neuronal morphologies built for reliable physiology in a rhythmic motor circuit".

Many thanks for pointing this out. We now include this work in our discussion and cite these papers.

Line 406: you forgot to cite Jackie Schiller's work on cortical pyramidal cells and their tuft dendrites, some of which predates all the cited work in this statement.

We regret this oversight and now appropriately cite the work.

Line 412: your model, as per Figure 2F, rather very closely matches the observed electrophysiological responses of the neuron under study. In what further ways could you model more closely match experimental observations? This would be very instructive to the reader.

This may be a misunderstanding. We simply highlight the possibility to incorporate active currents within the dendritic compartment if these should be discovered in Drosophila in the future. We now make this more clear by starting the sentence with "In case…"

Line 415: by ensembles of both KCS and MBONs, not just KCs, if you are including the memory part and not only the input representation part.

Yes, we changed the sentence accordingly.

Line 416: for most of the paper, you quoted these papers to justify the 5% of KCs being simultaneously active in response to an odor. Now the range is shown as 3 to 9%. This discrepancy ought to be reconciled.

As stated above we now make this point clear throughout the paper and discuss it more carefully (line 340ff).

Line 422: again, nowhere in the manuscript so far did you detail in what way your simulation and experimental findings match those of prior experimental reports regarding neuron response and physiological properties.

We agree that the sentence was misleading. We now compare this activation to the observed odor evoked changes in action potential frequency in MBONs in the studies by Hige and colleagues (line 469ff).

Line 425: this statement is inaccurate. An individual KC does not encode for a single odor. That would almost never be the case even if a KC was single-claw, as in, exclusively integrated inputs from a single projection neuron: individual projection neurons rarely encode an odor; it's the population of projection neurons that does. Similarly, ensembles of co-active KCs together represent an odor, and do so more narrowly and accurately than the population of projection neurons that excited them.

This seems to be a misunderstanding. In line 425 of the original version we did not talk about individual KCs but about sets of 50 KCs. This is in line with the statement of the reviewer.

Line 452: left unresolved remains the question of why would both mechanisms exist, as in, is the combination of altering the KC-MBON synapse and altering the PN-KC synapse better in some dimension than altering either alone? Does this relate perhaps to a possible dynamic nature of the olfactory "fovea" proposed by Zheng et al. 2020 bioRxiv as presumably static?

We now extended this part of the discussion and speculate about the relevance of these different plasticity modes (line 504ff).

Line 466: what is standard fly food? Please define it. Choice of food affects very much behavioral assays, for example.

As we did not perform any behavior assay, we did not provide a detailed recipe. We basically follow BDSC with small modifications.

Line 476: no antigen saturation steps in the immunohistochemistry protocol? Please revise.

Yes, we did not perform any antigen saturation steps as this is not necessary for our anti-GFP stainings.

Line 505: what were the settings of the puller? These would be necessary to reproduce your glass capillaries. Did you grind the tips, and if so, how?

We now provide the detailed settings of the puller.

Line 534: how were the R_series, R_input and R_m determined with Igor Pro?

We now specify this in the Methods section.

Line 561: why was a threshold of 600 MΩ used to exclude cells? How many cells were recorded in total, and how many were excluded?

We now specified the inclusion criteria more carefully. Cells not matching those criteria were not included in any analysis.

Line 584: once again the data is the best human effort in proofreading a semiautomatic segmentation. It is not the absolute truth. Also, each individual fly is somewhat different, so this is merely one fairly complete yet partial reconstruction, and not assured to be error free, particularly of errors of omission, of a single MBON from a single individual. Would be appropriate to remark this clearly.

We now state this clearly throughout the manuscript and again explicitly in the methods section.

Did you provide the NeuroML or similar files necessary to run the model in the NEURON simulation software? These should be appended to the manuscript as supplemental data.

Yes, all code necessary to run the model in the NEURON simulation software is included and documented in the Dataverse deposit, doi.org/10.7281/T1/HRK27V, as listed under ("Availability of data, materials, and code").

Line 596: if "parameters were tuned until the computed voltage excursion at the soma matched our electrophysiological recordings", what is the rationale for comparing the voltage potential of the simulation with those of the experimental observations?

The described part in the methods refers to the establishment of the basic electrophysiological properties of the simulated neuron. Our analysis then address the activation of this neuron.

Line 607: where do these ranges of physiologically plausible values come from? Citation, or measurements done in the lab and therefore a figure is needed to show them? Are these ranges from the literature listed in Table 2? Likely yes, would be appropriate to cite them here too or at least explicitly point to Table 2.

We now point to Table 2 in this section.

Line 616: innervation of MBONs is cholinergic because KCs are cholinergic, but that's not from Takemura et al. 2017, instead from Barnsted et al. 2016 Neuron.

We adjusted it accordingly.

Line 622: would be appropriate to include the python scripts used to configure and run the NEURON simulation as supplemental material. Likewise for the matlab scripts used to load the analyzed data and plot it. The raw data ought to be included as CSV files or similar.

All code and raw data necessary to run the model in the NEURON simulation software and to recapitulate our work is included and documented in the Dataverse deposit, doi.org/10.7281/T1/HRK27V, as listed under ("Availability of data, materials, and code").

Reviewer #2 (Recommendations for the authors):

"The cellular architecture of memory modules in Drosophila supports stochastic input integration" is a classical biophysical compartmental modelling study. It takes advantage of some simple current injection protocols in a massively complex mushroom body neuron called MBON-a3 and compartmental models that simulate the electrophysiological behaviour given a detailed description of the anatomical extent of its neurites.

This work is interesting in a number of ways:

– The input structure information comes from EM data (Kenyon cells) although this is not discussed much in the paper

– The paper predicts a potentially novel normalization of the throughput of KC inputs at the level of the proximal dendrite and soma

– It claims a new computational principle in dendrites, this didn't become very clear to me

Problems I see:

– The current injections did not last long enough to reach steady state (e.g. Figure 1FG), and the model current injection traces have two time constants but the data only one (Figure 2DF). This does not make me very confident in the results and conclusions.

These are two important but separate questions that we would like to address in turn.

As for the first, in our new recordings using cytoplasmic GFP to identify MBON-alpha3, we performed both a 200 ms current injection and performed prolonged recordings of 400 ms to reach steady state (for all 4 new cells 1’-4’). For comparison with the original dataset we mainly present the raw traces for 200 ms recordings in Figure 1 Supplement 2. In addition, we now provide a direct comparison of these recordings (200 ms versus 400 ms) and did not observe significant differences in tau between these data (Figure 1 Supplement 2 K). This comparison illustrates that the 200 ms current injection reaches a maximum voltage deflection that is close to the steady state level of the prolonged protocol. Importantly, the critical parameter (tau) did not change between these datasets.

Regarding the second question, the two different time constants, we thank the reviewer for pointing this out. Indeed, while the simulated voltage follows an approximately exponential decay which is, by design, essentially identical to the measured value (τ≈ 16ms, from Table 1; see Figure 1 Supplement 2 for details), the voltage decays and rises much faster immediately following the onset and offset of the current injections. We believe that this is due to the morphology of this neuron. Current injection, and voltage recordings, are at the soma which is connected to the remainder of the neuron by a long and thin neurite. This ’remainder’ is, of course, in linear size, volume and surface (membrane) area much larger than the soma, see Figure 2A. As a result, a current injection will first quickly charge up the membrane of the soma, resulting in the initial fast voltage changes seen in Figure 2D,F, before the membrane in the remainder of the cell is charged, with the cell’s time constant τ.

Author response image 1
Simplified circuit of a small soma (left parallel RC circuit) and the much larger remainder of a cell (right parallel RC circuit) connected by a neurite (right 100M resistor).

A current source (far left) injects constant current into the soma through the left 100M resistor.

We confirmed this intuition by running various simplified simulations in Neuron which indeed show a much more rapid change at step changes in injected current than over the long-term. Indeed, we found that the pattern even appears in the simplest possible two-compartment version of the neuron’s equivalent circuit which we solved in an all-purpose numerical simulator of electrical circuitry.

Author response image 2
Somatic voltage in the circuit in Figure 1 with current injection for about 4.

5ms, followed by zero current injection for another 3.5ms..

Author response image 3
: Somatic voltage in the circuit, as in Figure 2 but with current injected for approx.

15ms.

(https://www.falstad.com/circuit). The circuit is shown in Author response image 1. We chose rather generic values for the circuit components, with the constraints that the cell capacitance, chosen as 15pF, and membrane resistance, chosen as 1G, are in the range of the observed data (as is, consequently, its time constant which is 15ms with these choices); see Table 1 of the manuscript. We chose the capacitance of the soma as 1.5pF, making the time constant of the soma (1.5ms) an order of magnitude shorter than that of the cell.

Author response image 2 shows the somatic voltage in this circuit (i.e., at the upper terminal of the 1.5pF capacitor) while a -10pA current is injected for about 4.5ms, after which the current is set back to zero. The combination of initial rapid change, followed by a gradual change with a time constant of 15ms is visible at both onset and offset of the current injection. Author response image 3 show the voltage traces plotted for a duration of approximately one time constant, and Author response image 4 shows the detailed shape right after current onset.

While we did not try to quantitatively assess the deviation from a single-exponential shape of the voltage in Figure 2E, a more rapid increase at the onset and offset of the current injection is clearly visible in Author response image 4. This deviation from a single exponential is smaller than what we see in the simulation (both in Figure 2D of the manuscript, and in the results of the simplified circuit in Author response image 4). We believe that the effect is smaller in Figure E because it shows the average over many traces. It is much more visible in the ’raw’ (not averaged) traces. Two randomly selected traces from the first of the recorded neurons are shown in Figure 2 Supplement 2 C. While the non-averaged traces are plagued by artifacts and noise, the rapid voltage changes are visible essentially at all onsets and offsets of the current injection.

Author response image 4
Somatic voltage in the circuit, as in Figure 2 but showing only for the time right after current onset, about 2.

3ms.

We have added a short discussion of this at the end of Section 2.3 to briefly point out this observation and its explanation. We there also refer to the simplified circuit simulation and comparison with raw voltage traces which is now shown in the new Figure 2 Supplement 2.

– The time constant in Table 1 is much shorter than in Figure 1FG?

No, these values are in agreement. To facilitate the comparison we now include a graphical measurement of tau from our traces in Figure 1 Supplement 2 J.

– Related to this, the capacitance values are very low maybe this can be explained by the model's wrong assumption of tau?

Indeed, the measured time constants are somewhat lower than what might be expected. We believe that this is because after a step change of the injected current, an initial rapid voltage change occurs in the soma, where the recordings are taken. The measured time constant is a combination of the ’actual’ time constant of the cell and the ’somatic’ (very short) time constant of the soma. Please see our explanations above.

Importantly, the value for tau from Table 1 is not used explicitly in the model as the parameters used in our simulation are determined by optimal fits of the simulated voltage curves to experimentally obtained data.

– That latter in turn could be because of either space clamp issues in this hugely complex cell or bad model predictions due to incomplete reconstructions, bad match between morphology and electrophysiology (both are from different datasets?), or unknown ion channels that produce non-linear behaviour during the current injections.

Please see our detailed discussion above. Furthermore, we now provide additional recordings using cytoplasmic GFP as a marker for the identification of MBON-alpha3 and confirm our findings. We agree that space-clamp issues could interfere with our recordings in such a complex cell. However, our approach using electrophysiological data should still be superior to any other approach (picking text book values). As we injected negative currents for our analysis at least voltage-gated ion channels should not influence our recordings.

- The PRAXIS method in NEURON seems too ad hoc. Passive properties of a neuron should probably rather be explored in parameter scans.

We are a bit at a loss of what is meant by the PRAXIS method being "too ad hoc." The PRAXIS method is essentially a conjugate gradient optimization algorithm (since no explicit derivatives are available, it makes the assumption that the objective function is quadratic). This seems to us a systematic way of doing a parameter scan, and the procedure has been used in other related models, e.g. the cited Gouwens & Wilson (2009) study.

Questions I have:

– Computational aspects were previously addressed by e.g. Larry Abbott and Gilles Laurent (sparse coding), how do the findings here distinguish themselves from this work.

In contrast to the work by Abbott and Laurent that addressed the principal relevance and suitability of sparse and random coding for the encoding of sensory information in decision making, here we address the cellular and computational mechanisms that an individual node (KC>MBON) play within the circuitry. As we use functional and morphological relevant data this study builds upon the prior work but significantly extends the general models to a specific case. We think this is essential for the further exploration of the topic.

– What is valence information?

Valence information is the information whether a stimulus is good (positive valence, e.g. sugar in appetitive memory paradigms, or negative valence in aversive olfactory conditioning – the electric shock). Valence information is provided by the dopaminergic system. Dopaminergic neurons are in direct contact with the KC>MBON circuitry and modify its synaptic connectivity when olfactory information is paired with a positive or negative stimulus.

– It seems that Martin Nawrot’s work would be relevant to this work.

We are aware of the work by the Nawrot group that provided important insights into the processing of information within the olfactory mushroom body circuitry. We now highlight some of his work. His recent work will certainly be relevant for our future studies when we try to extend our work from an individual cell to networks.

– Compactification and democratization could be related to other work like Otopalik et al. 2017 eLife but also passive normalization. The equal efficiency in line 427 reminds me of dendritic/synaptic democracy and dendritic constancy.

Many thanks for pointing this out. This is in line with the comments from reviewer 1 and we now highlight these papers in the relevant paragraph in the discussion (line 442ff).

– The morphology does not obviously seem compact, how unusual would it be that such a complex dendrite is so compact?

We should have been more careful in our terminology, making clear that when we write ’compact’ we always mean ’electrotonically compact," in the sense that the physical dimensions of the neuron are small compared to its characteristic electrotonic length (usually called λ). The degree of a dendritic structure being electrotonically compact is determined by the interaction of morphology, size and conductances (across the membrane and along the neurites). We don’t believe that one of these factors alone (e.g. morphology) is sufficient to characterize the electrical properties of a dendritic tree. We have now clarified this in the relevant section.

– What were the advantages of using the EM circuit?

The purpose of our study is to provide a "realistic" model of a KC>MBON node within the memory circuitry. We started our simulations with random synaptic locations but wondered whether such a stochastic model is correct, or whether taking into account the detailed locations and numbers of synaptic connections of individual KCs would make a difference to the computation. Therefore we repeated the simulations using the EM data. We now address the point between random vs realistic synaptic connectivity in Figure 4F. We do not observe a significant difference but this may become more relevant in future studies if we compute the interplay between MBONs activated by overlapping sets of KCs. We simply think that utilizing the EM data gets us one step closer to realistic models.

– Isn't Figure 4E rather trivial if the cell is compact?

We believe this figure is a visually striking illustration that shows how electrotonically compact the cell is. Such a finding may be trivial in retrospect, once the data is visualized, but we believe it provides a very intuitive description of the cell behavior.

Overall, I am worried that the passive modelling study of the MBON-a3 does not provide enough evidence to explain the electrophysiological behaviour of the cell and to make accurate predictions of the cell's responses to a variety of stochastic KC inputs.

In our view our model adequately describes the behavior of the MBON with the most minimal (passive) model. Our approach tries to make the least assumptions about the electrophysiological properties of the cell. We think that based on the current knowledge our approach is the best possible approach as thus far no active components within the dendritic or axonal compartments of Drosophila MBONs have been described. As such, our model describes the current status which explains the behavior of the cell very well. We aim to refine this model in the future if experimental evidence requires such adaptations.

Reviewer #3 (Recommendations for the authors):

This manuscript presents an analysis of the cellular integration properties of a specific mushroom body output neuron, MBON-α3, using a combination of patch clamp recordings and data from electron microscopy. The study demonstrates that the neuron is electrotonically compact permitting linear integration of synaptic input from Kenyon cells that represent odor identity.

Strengths of the manuscript:

1) The study integrates morphological data about MBON-α3 along with parameters derived from electrophysiological measurements to build a detailed model.

2) The modeling provides support for existing models of how olfactory memory is related to integration at the MBON.

Weaknesses of the manuscript:

1) The study does not provide experimental validation of the results of the computational model.

The goal of our study is to use computational approaches to provide insights into the computation of the MBON as part of the olfactory memory circuitry. Our data is in agreement with the current model of the circuitry. Our study therefore forms the basis for future experimental studies; those would however go beyond the scope of the current work.

2) The conclusion of the modeling analysis is that the neuron integrates synaptic inputs almost completely linearly. All the subsequent analyses are straightforward consequences of this result.

We do, indeed, find that synaptic integration in this neuron is almost completely linear. We demonstrate that this result holds in a variety of different ways. All analyses in the study serve this purpose. These results are in line with the findings by Hige and Turner (2013) who demonstrated that also synaptic integration at PN>KC synapses is highly linear. As such our data points to a feature conservation to the next node of this circuit.

3) The manuscript does not provide much explanation or intuition as to why this linear conclusion holds.

We respectfully disagree. We demonstrate that this linear integration is a combination of the size of the cell and the combination of its biophysical parameters, mainly the conductances across and along the neurites. As to why it holds, our main argument is that results based on the linear model agree with all known (to us) empirical results, and this is the simplest model.

In general, there is a clear takeaway here, which is that the dendritic tree of MBON-α3 in the lobes is highly electrotonically compact. The authors did not provide much explanation as to why this is, and the paper would benefit from a clearer conclusion. Furthermore, I found the results of Figures 4 and 5 rather straightforward given this previous observation. I am sceptical about whether the tiny variations in, e.g. Figures 3I and 5F-H, are meaningful biologically.

Please see the comment above as to the ’why’ we believe the neuron is electrotonically compact: a model with this assumption agrees well with empirically found results.

We agree that the small variations in Figure 5F-H are likely not biologically meaningful. We state this now more clearly in the figure legends and in the text. This result is important to show, however. It is precisely because these variations are small, compared to the differences between voltage differences between different numbers of activated KCs (Figure 5D) or different levels of activated synapses (Figure 5E) that we can conclude that a 25% change in either synaptic strength or number can represent clearly distinguishable internal states, and that both changes have the same effect. It is important to show these data, to allow the reader to compare the differences that DO matter (Figure 5D,E) and those that DON’T (Figure 5F-H).

The same applies to Figure 3I. The reviewer is entirely correct: the differences in the somatic voltage shown in Figure 3I are minuscule, less than a micro-Volt, and it is very unlikely that these difference have any biological meaning. The point of this figure is exactly to show this!. It is to demonstrate quantitatively the transformation of the large differences between voltages in the dendritic tree and the nearly complete uniform voltage at the soma. We feel that this shows very clearly the extreme "democratization" of the synaptic input! Please see also the reply to your second comment below.

1) My biggest question is about the claim of extreme electrotonic compactness of this neuron. Figure 3D,E suggests that the voltage change at the proximal neurite and at the soma varies by only about 1% depending on stimulation location. Since this is supported only by simulation, it is worth asking how robust this conclusion is.

The results of the simulation are very robust. The linearity of the system makes the solution of the underlying equations very stable; small perturbations in the input are guaranteed (by the nature of linear systems) to only result in small changes in the output.

As for how well the results reflect biological reality, this is a question that eventually needs to be answered by empirical investigations. The results of any theory, like the one in this study, are subject to verification/falsification. We consider our theoretical results as predictions that are easily falsifiable given suitable experimental techniques. Future empirical studies will determine the validity of these predictions.

a) Given that the variability in 3I is so small in magnitude, any dependence would be swamped by other sources of heterogeneity, so the statement that this correlates with distance (line 278) is likely irrelevant.

The reviewer is precisely right, the differences in the somatic voltage shown in Figure 3I are minuscule, less than a micro-Volt. But that is exactly the point of this Figure!

What we show in Figure 3H is that there are substantial differences (by at least an order of magnitude) between "local" voltages, i.e. at the synaptic locations in the dendritic tree. This is expected because of the much smaller diameter of distal parts of the tree than more proximal ones. What is unexpected (at least it was to us) was that these differences are EXACTLY (within a fraction of a micro Volt) compensated by the electrotonic decay from the synapses to the soma. This is the case for ALL 3121 synaptic segments in the dendritic tree. It thus appears as if the morphology of the neuron is fine-tuned to result in extreme ’democratization’ of all synapses.

Note that this goes beyond the equalizing effect resulting from a large space constant (λ) which is also present in this neuron (at least in our simulation; we are not aware of published results reporting empirically determined voltage distributions in the dendritic tree of this neuron). While large λ compresses the range of voltage distributions at distant locations, it is a linear phenomenon and as such it preserves the relative voltage values. The interplay of segment size and, more generally, morphology of the dendritic tree allows for nonlinear phenomena, such as the compression of the voltage range in the dendritic tree which vary over a factor of 30 (Figure 3H) towards a much smaller range (variation by about 1%) in Figure 3I.

Reviewer 2 had a similar concern, so we clearly did a very poor job in explaining this. We therefore rewrote the early parts of section 2. to make this more clear (line 268ff).

b) Can the authors provide a confidence interval for the fit of biophysical parameters to their recordings?

The fitting was performed as an iterative approach in the NEURON environment until we got the best fit of simulated results to neurophysiological data was achieved. Please see details in comments to reviewer 2 above. As such we cannot provide a confidence interval for the fit.

c) On lines 271-274, the authors state, "This architecture with the smallest dendritic sections at the most distant sites may contribute to the compactness of the dendritic tree, ensuring that even the most distant synaptic inputs result in somatic voltage deflections comparable to the most proximal ones." Is a dependence of dendritic size on distance required for the results? It seems like the result is simply that there is no attenuation within the dendritic tree at all. In general, the authors don't actually provide an explanation for the compactness. In the discussion, it is stated that, "The compactification of the neuron is likely related to the architectural structure of its dendritic tree," which again is rather vague. The authors should strive to provide a clear explanation for this, since it is their key result.

We agree that we should have been more explicit in describing the origin of the electrotonic compactness of the neuron. We also agree that the main reason is the low attenuation of electrical signals in the dendritic tree (large space constant λ). We now add a paragraph (second paragraph of Section 2.3) where we compute an estimate for the characteristic length of the cell and show that it substantially exceeds the cell size, resulting in low decay of distal input.

However, we believe that there is a second effect that contributes to the compactness of the cell. Analysis of the dendritic morphology reveals that the volume of dendritic segments is inversely correlated with the distance from the soma (Figure 3G). For a given current input (or synaptic conductance change), this results in much higher voltage excursions in the distal parts of the dendritic tree, with small neurite diameters, than in more proximal parts, with larger diameters, Figure 3H. As a consequence, the additional voltage decreases due to longer path length from distal synapses vs. proximal synapses is nearly exactly compensated by the increase of local dendritic voltage in the distal locations. This is shown in the nearly identical voltages experienced in the soma by distal and proximal synapses, Figure 2. Therefore, it is not only the large space constant which is important but also the interplay between small distal dendritic volumes and long distance between them and the soma.

d) Can the authors report what the electrotonic length for such a dendrite would be? How long before we expect to see a significant spread in 3E?

We now report the electronic length in the rewritten Section 2.3, see also the previous point. Our simple analytical calculation gives λ= 1,340µm for the dendritic tree only, and λ= 1,510µm for the whole neuron, when we base it on the mean of the diameters of all neuronal segments.

2) My other concern is that, once we assume this perfect integration, the subsequent analyses are all a straightforward consequence. In particular, Figures 4 and 5 just repeatedly convey that it doesn't matter which synapse is being activated, the effect on the somatic voltage is the same. I was particularly confused about the conclusions of 5F,G,H. The authors claim "small but significant differences" here, but practically speaking, I can't imagine any of the differences in these plots being meaningful.

It is correct that the neuron’s compactness determines its behavior. It is indeed one of our main conclusions that the influence of all synapses on the somatic voltage is very similar which, as we discuss, is important for the odor representation by KCs and its interpretation by the MBON. We document that with different analyses, e.g. those reported in the figure panels referred to by the reviewer. They could be considered redundant but we feel that they elucidate different aspects of the main message of our study. If the reviewer feels strongly that these panels are not needed, we are willing to do without them but we feel that they convey the message very clearly.

3) Reference to the data used to constrain the model is confusing.

a) At various places in the manuscript references are made to in vivo recordings, but it appears that all of the recordings were done ex vivo.

We apologize for this mistake. Yes, all our recordings were performed ex vivo. We now state this explicitly throughout the manuscript. However, we also refer to some data by other groups that were collected in vivo.

b) On lines 389-392, the authors state: "Near-perfect agreement between experimentally observed and simulated voltage distributions in the dendritic tree shows that linear cable theory is an excellent model for information integration in this system." What recordings of the voltage distribution in the dendritic tree were performed?

Of course the reviewer is correct: Neither we nor, as far as we know, any other groups have recorded detailed voltage distributions in these neurons. We meant to refer to the voltage traces (not distributions) shown in Figure 2D-F that show excellent agreement between recorded and simulated data. The text has been corrected.

4) It seems like the conclusions are different than those of Gouwens and Wilson (2009), who described their reconstructed PNs as electrotonically extensive. The authors should comment on what about MBONs and PNs is different.

This is an excellent question, and it can be answered at several levels.

First, the most obvious one: why do the simulations result in neurons with different (compact vs extended) model neurons? In both cases, the model parameters used in the simulation are determined by finding the optimal fit to electrophysiological (patch clamp) data. Even the numerical methods are identical: both studies use the PRAXIS method in NEURON. In both cases, the fits obtained are in excellent agreement with the measured data. One difference is that we construct the dendritic tree from the morphological (EM) data while Gouwen and Wilson constructed synthetic dendrites since the exact morphology was not available at the time when they conducted their study. We do, however, not believe that this difference is of importance for the question at hand.

The electrotonic length λ of a neurite depends on three variables, its diameter a, membrane resistance rm and inverse cytoplasmatic resistance ra: λ2 = ar2rma. For the diameter, we used the EM data in the dendritic tree and, as available, for other neurites, and the light-microscopic data where EM data was not available. Gouwen and Wilson used synthetic segments in the dendritic tree and light-microscopical data elsewhere. We could not find the exact values they used in their paper, therefore we can not say whether there are substantial differences between their values and ours. We consider it unlikely that this is the case.

There is a factor of nearly three between our best fit for cytoplasm resistance in the denominator; our value (85m) is about three times smaller than Gouwen and Wilson’s ( 250m, depending on the chosen model neuron). There is also a factor of about five between the fits for the membrane resistance in the numerator: our value is about five times larger ( 100Kcm2) than theirs ( 20Kcm2). Since these numbers multiply, we find a factor of about fifteen. Assuming similar diameters, after taking the square root we find that their λ is about 4 times smaller than ours. Would we take their fitted values for rm and ra, we would likewise obtain a neuron whose characteristic length is substantially smaller than its physical length, i.e.an electrotonically extended neuron. The computations are therefore consistent.

On an anatomical level, Scheffer et al., 2020 showed that neurons are commonly electrotonically compact within one anatomical compartment (as defined in their Table 1) but functionally segregated between compartments (their page 46ff). This is consistent with the electrotonically extended projection neuron studied by Gouwen and Wilson, which receives its input in one of the glomeruli and project its output in the mushroom body, clearly two different compartments. In contrast, MBONα-3 is entirely in the α lobe of the mushroom body, i.e. in one of the compartments from Table 1 in Scheffer et al., 2020 and therefore compact.

Finally, on a third level, does the difference between these neurons make sense functionally? We believe that it does. MBONα-3 receives input from KCs that represent odor input in a highly distributed sparse code. We have hypothesized in this study that this code is read by the MBONs by having all synapses count equally. This is not compatible with an extended structure since the impact of a synapse would depend on its location on the cell. In contrast, the PN studied by Gouwen and Wilson receives input in its glomerulus from the olfactory receptor neurons, from between 10 and 50 synapses (release sites). Indeed, their simulations showed that a volley of this size (they assumed 25 simultaneously active synapses) is required to generate a physiologically meaningful effect at the soma, with input from small numbers of synapses having little effect. An electrotonically compact neuron would faithfully transmit the effect of each individual synapse which due to stochastic variability will likely result in a high level of noise. Therefore, the electrotonically extended neuron enforces that only robust input from a large number of near-simultaneous inputs is registered at the soma.

We therefore conclude, that these two neurons have very different decoding roles which require exactly the difference in biophysical properties that are found in these two complementary studies by Wilson and Gouwen and our own. It will of course be interesting to directly evaluate this topic again in a comparative study in which both simulations depend on the EM-based reconstructions and are performed side-by-side. Only then, we can provide definitive answers – or alternatively we aim to highlight it in a review format. Therefore we kept this discussion for now to this place.

https://doi.org/10.7554/eLife.77578.sa2

Article and author information

Author details

  1. Omar A Hafez

    Zanvyl Krieger Mind/Brain Institute, Johns Hopkins University, Baltimore, United States
    Present address
    Yale MD-PhD Program, Yale School of Medicine, New Haven, United States
    Contribution
    Conceptualization, Data curation, Software, Formal analysis, Validation, Visualization, Methodology, Writing – original draft, Writing – review and editing
    Contributed equally with
    Benjamin Escribano and Rouven L Ziegler
    Competing interests
    No competing interests declared
    ORCID icon "This ORCID iD identifies the author of this article:" 0000-0002-7846-9226
  2. Benjamin Escribano

    Division of Neurobiology and Zoology, Department of Biology, University of Kaiserslautern, Kaiserslautern, Germany
    Present address
    German Center for Neurodegenerative Diseases (DZNE), Bonn, Germany
    Contribution
    Conceptualization, Data curation, Formal analysis, Validation, Visualization, Methodology, Writing – original draft, Writing – review and editing
    Contributed equally with
    Omar A Hafez and Rouven L Ziegler
    Competing interests
    No competing interests declared
    ORCID icon "This ORCID iD identifies the author of this article:" 0000-0002-1432-5952
  3. Rouven L Ziegler

    Division of Neurobiology and Zoology, Department of Biology, University of Kaiserslautern, Kaiserslautern, Germany
    Contribution
    Conceptualization, Data curation, Formal analysis, Validation, Visualization, Methodology, Writing – original draft, Writing – review and editing
    Contributed equally with
    Omar A Hafez and Benjamin Escribano
    Competing interests
    No competing interests declared
    ORCID icon "This ORCID iD identifies the author of this article:" 0000-0002-3050-7692
  4. Jan J Hirtz

    Physiology of Neuronal Networks Group, Department of Biology, University of Kaiserslautern, Kaiserslautern, Germany
    Contribution
    Formal analysis, Methodology
    Competing interests
    No competing interests declared
    ORCID icon "This ORCID iD identifies the author of this article:" 0000-0002-4486-3057
  5. Ernst Niebur

    1. Zanvyl Krieger Mind/Brain Institute, Johns Hopkins University, Baltimore, United States
    2. Solomon Snyder Department of Neuroscience, Johns Hopkins University, Baltimore, United States
    Contribution
    Conceptualization, Data curation, Software, Formal analysis, Supervision, Funding acquisition, Validation, Visualization, Writing – original draft, Project administration, Writing – review and editing
    For correspondence
    niebur@jhu.edu
    Competing interests
    No competing interests declared
    ORCID icon "This ORCID iD identifies the author of this article:" 0000-0002-2815-9262
  6. Jan Pielage

    Division of Neurobiology and Zoology, Department of Biology, University of Kaiserslautern, Kaiserslautern, Germany
    Contribution
    Conceptualization, Data curation, Formal analysis, Supervision, Funding acquisition, Validation, Visualization, Methodology, Writing – original draft, Project administration, Writing – review and editing
    For correspondence
    pielage@bio.uni-kl.de
    Competing interests
    No competing interests declared
    ORCID icon "This ORCID iD identifies the author of this article:" 0000-0002-5115-5884

Funding

National Institutes of Health (R01DC020123)

  • Ernst Niebur

National Institutes of Health (R01DA040990)

  • Ernst Niebur

National Institutes of Health (R01EY027544)

  • Ernst Niebur

National Institutes of Health (Medical Scientist Training Program Training Grant T32GM136651)

  • Omar A Hafez

National Science Foundation (1835202)

  • Ernst Niebur

Bundesministerium für Bildung und Forschung (FKZ 01GQ2105)

  • Jan Pielage

Deutsche Forschungsgemeinschaft (INST 248/293-1)

  • Jan Pielage

The funders had no role in study design, data collection and interpretation, or the decision to submit the work for publication.

Acknowledgements

We thank Glenn Turner and Toshihide Hige for insightful comments on the patch clamp analysis, Hans-Peter Schneider for technical help and Oliver Barnstedt for comments on the manuscript. Funding: This work was supported by NIH R01DC020123, NSF 1835202, NIH R01DA040990, NIH R01EY027544 (all EN), NIH Medical Scientist Training Program Training Grant T32GM136651 (OH), and a BMBF (FKZ 01GQ2105) and a DFG grant (INST 248/293–1) (all JP).

Senior Editor

  1. K VijayRaghavan, National Centre for Biological Sciences, Tata Institute of Fundamental Research, India

Reviewing Editor

  1. Albert Cardona, University of Cambridge, United Kingdom

Reviewer

  1. Albert Cardona, University of Cambridge, United Kingdom

Version history

  1. Preprint posted: December 7, 2020 (view preprint)
  2. Received: February 3, 2022
  3. Accepted: March 9, 2023
  4. Accepted Manuscript published: March 14, 2023 (version 1)
  5. Accepted Manuscript updated: March 14, 2023 (version 2)
  6. Accepted Manuscript updated: March 15, 2023 (version 3)
  7. Version of Record published: April 3, 2023 (version 4)

Copyright

© 2023, Hafez, Escribano, Ziegler et al.

This article is distributed under the terms of the Creative Commons Attribution License, which permits unrestricted use and redistribution provided that the original author and source are credited.

Metrics

  • 1,115
    Page views
  • 174
    Downloads
  • 0
    Citations

Article citation count generated by polling the highest count across the following sources: Crossref, PubMed Central, Scopus.

Download links

A two-part list of links to download the article, or parts of the article, in various formats.

Downloads (link to download the article as PDF)

Open citations (links to open the citations from this article in various online reference manager services)

Cite this article (links to download the citations from this article in formats compatible with various reference manager tools)

  1. Omar A Hafez
  2. Benjamin Escribano
  3. Rouven L Ziegler
  4. Jan J Hirtz
  5. Ernst Niebur
  6. Jan Pielage
(2023)
The cellular architecture of memory modules in Drosophila supports stochastic input integration
eLife 12:e77578.
https://doi.org/10.7554/eLife.77578

Further reading

    1. Cell Biology
    2. Neuroscience
    Elisabeth Jongsma, Anita Goyala ... Collin Yvès Ewald
    Research Article Updated

    The amyloid beta (Aβ) plaques found in Alzheimer’s disease (AD) patients’ brains contain collagens and are embedded extracellularly. Several collagens have been proposed to influence Aβ aggregate formation, yet their role in clearance is unknown. To investigate the potential role of collagens in forming and clearance of extracellular aggregates in vivo, we created a transgenic Caenorhabditis elegans strain that expresses and secretes human Aβ1-42. This secreted Aβ forms aggregates in two distinct places within the extracellular matrix. In a screen for extracellular human Aβ aggregation regulators, we identified different collagens to ameliorate or potentiate Aβ aggregation. We show that a disintegrin and metalloprotease a disintegrin and metalloprotease 2 (ADM-2), an ortholog of ADAM9, reduces the load of extracellular Aβ aggregates. ADM-2 is required and sufficient to remove the extracellular Aβ aggregates. Thus, we provide in vivo evidence of collagens essential for aggregate formation and metalloprotease participating in extracellular Aβ aggregate removal.

    1. Computational and Systems Biology
    2. Neuroscience
    Marjorie Xie, Samuel P Muscinelli ... Ashok Litwin-Kumar
    Research Article Updated

    The cerebellar granule cell layer has inspired numerous theoretical models of neural representations that support learned behaviors, beginning with the work of Marr and Albus. In these models, granule cells form a sparse, combinatorial encoding of diverse sensorimotor inputs. Such sparse representations are optimal for learning to discriminate random stimuli. However, recent observations of dense, low-dimensional activity across granule cells have called into question the role of sparse coding in these neurons. Here, we generalize theories of cerebellar learning to determine the optimal granule cell representation for tasks beyond random stimulus discrimination, including continuous input-output transformations as required for smooth motor control. We show that for such tasks, the optimal granule cell representation is substantially denser than predicted by classical theories. Our results provide a general theory of learning in cerebellum-like systems and suggest that optimal cerebellar representations are task-dependent.