Experiment design and basic response properties.

A. (Top) Network schematic of hippocampal CA3-IN-CA1 network. (Bottom) A transverse hippocampal section showing channelrhodopsin expression in orange (tdTomato), with the stimulation grid drawn to scale overlaid on the CA3 network. An extracellular (field) electrode was used to record the total optogenetic excitation of the CA3 layer (black arrowhead), and CA1 recordings were made from individual pyramidal cells with a whole-cell patch clamp electrode (white arrowhead). (scale bar = 500µm) B. (Top) A DIC image of the CA3 cell layer with a few spots of a stimulation grid overlaid, drawn to scale. (Black dotted lines mark the outlines of a few CA3 cells). (Bottom) A schematic of the CA3-IN-CA1 layer showing the recruitment of CA3 pyramidal cells with three different patterns. The downstream CA1 pyramidal cell receives direct monosynaptic excitation from the activated population and disynaptic inhibition via a heterogeneous, shared population of CA1 interneurons. C. Burst recording protocol. Top (red): Trigger to start sequence. Cyan: Optical stimulus. Black: Postsynaptic potential of a patched CA1 cell (black). Orange: Simultaneously recorded field potential (orange). D. Extracellular recording of CA3 layer activation for 1,5,15 square patterns. E. CA3 field response to first pulse increases with number of squares per pattern (Kruskal-Wallis test). F. Correlation of optically stimulated current recorded from patched CA3 cells with the corresponding field response from CA3. G. Post synaptic potentials for first pulse (hence, no STP) recorded in a CA1 pyramidal cell for different pattern sizes. H. Distribution of PSPs across all recorded CA1 cells differs for number of squares (n=16, p < 0.001, Kruskal-Wallis test). I. Channelrhodopsin desensitization during burst is under 10% (linear regression fit with slope 2%, r2=0.05, bars indicate 95% CI). J. Postsynaptic currents (PSC) recorded for first pulse from CA1 cells in voltage clamp show the proportional relationship between excitatory and inhibitory PSCs for each pattern size. K. Distribution of PSC amplitude differs between 1 and 15 square stimuli. L. A sample recording of excitatory (pink) and inhibitory (teal) currents recorded from a CA1 cell in the burst protocol described in panel C, for a 15 square pattern.

PSP, spiking, EPSC and IPSC dependence on stimulus frequency over a train of pulses.

(A) Trial-normalised excitatory postsynaptic potentials across stimulation frequency and pulse index in the train for 5 square (Ai) and 15 square (Aii) patterns across all recorded cells (n=16). B. Spike probability as a function of the stimulus frequency, pulse index and number of squares in the current clamp cells showing higher likelihood in the first half of the train for five squares (top) and 15 square patterns (bottom). C. Schematic for derivation of Post-synaptic responses. Kernel fits were obtained to valley to peak height for each response for both postsynaptic currents (voltage clamp, EPSC and IPSCs) and postsynaptic voltage (current clamp, not shown). D. Normalised EPSC and IPSC responses obtained for an example cell for a sample pattern showing consistent depression in inhibition and biphasic response in excitation. E, F. Both excitatory (Ei-Eii) and inhibitory (Fi-Fii) PSCs show short-term depression along the pulse train. Traces were normalised to a reference pulse 0, delivered 300ms before the burst.

Excitatory-inhibitory balance evolves over the pulse sequence.

A. Postsynaptic potentials have a biphasic profile: they first rise, then fall in a pulse train. Pulse 2 > 1.0 at p=0.0003, Pulse 8 < 1.0 at p = 0.0003, Wilcoxon signed rank test. Stimulus was 20 Hz and response is normalized to pulse 0, for all patterns of size 5 and 15. B. Same stimulus as A, for voltage clamp. Postsynaptic excitatory (red) currents first rise slightly, then fall. Inhibitory (teal) currents fall throughout the train. All points except pulse 2 differ at p < 0.05, Holm-Bonferroni corrected Wilcoxon test. C. E-I ratio rises along the pulse train. Dashed line: saturating exponential fit, τ=2.115±1.0 pulses. D. The onset delay for both excitation and inhibition rises steadily over the pulse train. Dashed lines are linear regression fit with slope = 0.253ms/pulse for E and 0.232ms/pulse for I, slope difference not significant at p=0.123). E. Scaling of observed responses in CA1 for patterns of different sizes (5, 7, 15 spots) as compared to the expected responses obtained by the sum of constituent single-spot responses, shown for cell 3402, to illustrate the response normalisation. Solid purple curve is a fit to the equation with gamma = 3.94 = expected*observed/(expected – observed). Solid green line is best fit linear slope, m = 0.30. F. Gamma shows a leftward shift along the pulse train, here shown by comparing cumulative distribution of gamma for probe pulse vs pulse 8 of the train. (p<1e-3, Mann-Whitney test) G. Gamma depends both on pulse index and frequency. Gamma during the first four pulses is larger than in the later four (p<0.001, ANOVA).

Multiscale model.

A: Model of reaction system in each bouton. Synaptic input triggers entry of extracellular Ca2+ into the bouton, and a pump removes the Ca. CaM buffers the Ca2+. Ca binds/unbinds from successive stages of the readily releasable (RR) pool of vesicles till they are docked, at which point the final Ca-binding step causes synaptic release. The released neurotransmitter (Glu for excitatory synapses, GABA for inhibitory) opens a ligand-gated receptor channel on the postsynaptic CA1 neuron, which is held voltage-clamped.(green arrow). The reaction scheme was identical for Glutamatergic and GABAergic synapses, but the rates were different to fit the respective voltage-clamp recordings. B: Close-up of dendrite of CA1 pyramidal neuron model. The cell has a compartment for soma, a 200-micron compartment for the dendrite, and 100 spines. Each spine is modelled as a head compartment, a neck compartment, and a presynaptic glutamatergic bouton. There are 200 GABAergic presynaptic boutons positioned directly on the dendrite. C: Network model. The CA3 has 16×16 neurons, projecting randomly to 16×16 interneurons. CA3 also projects directly to the Glu synapses on the spines. The interneurons project to GABA synapses.

Temporal summation and EI balance in burst stimuli.

A-D: EPSP measured at soma for representative cell 3402 (maroon) and simulation (yellow) for each frequency. E: Distribution and mean of EPSP over all recorded cells for each frequency for the experiment (maroon) and for a simulated neuron (yellow). There is no significant dependence of EPSP on frequency for the recordings (Linear regression, slope=0.005, p=0.168) or for simulations (Linear regression, slope = 0.01, p=0.134). Simulated slope lies within the 95% range of the confidence interval [-0.002, 0.0126] of the slope for experimental data (Bootstrap, 10000 resamples, Table 1). The simulated peak EPSPs lie within the range of experimental cells at each of the four frequencies (Table 1) F-I: Simulated excitatory and inhibitory currents for each frequency. Note that the currents are measured close to resting potential, so the driving force (EGABA – Vm) is small compared to the voltage-clamp experiments in Figure 1K, where Vm was held at 0 mV. J: EI balance (ratio, black line) is maintained for simulated excitatory and inhibitory current peaks for all frequencies. K-N: Simulated excitatory and inhibitory conductances for each frequency. O: Simulated excitatory and inhibitory conductance peaks. Note that simulated synaptic conductances are reduced by ∼50x from typical cellular values to match the much smaller cell geometry and higher input impedance.

Summary of statistical comparisons between simulation and experiment for figures 5 and 6.

For Figure 5 panel E, the simulation EPSP peaks do not differ significantly from experiment either for the means for each individual frequency, nor does the slope across frequencies differ from experiment. For Figure 6 we obtained the specified statistics for each panel, and asked if the simulation was within the range of values seen for individual cells. In all cases it was within the range, except for panel 6F for the 5 Square stimulus.

Experimental and simulated responses to Poisson spike train input, all patterns.

A: Sample experimental EPSP train aligned with light trigger pulses (green). Blue trace is recorded data; red trace is the fit used to compute peaks and subsequent statistics. B: Same for simulated data. C-G: Comparisons for readouts of experiment (left column) and simulated (right column) data. Blue traces/markers: 5-square stimulation, orange traces/markers: 15-square stimulation. Outcomes are detailed in Table 1. In all cases except panel F-5Sq, the simulated response lies within the range of individual cell responses. C: Power spectral density of EPSP response over the entire dataset. Note 50Hz peak due to line noise in experimental dataset, despite this the cosine distance between experimental mean and simulated response was within the range of individual cell cosine distances to mean. D: probability of trigger to generate a peak in the EPSP trace. E: Scatter plot of EPSP peaks as a function of time over the Poisson train. Note the initial decline due to STP and ChR2 desensitisation. Comparisons were made both for this initial decline (up to 1.8 s) and for the subsequent steady-state EPSP peak distribution. F: Scatter plot of EPSP peaks vs. Inter-spike-interval. The experimental data show an elevation of EPSP at short ISI. We fit this using a simple exponential decay function y=y0.exp(-t/tau)+y1. G: Distribution of EPSP peak amplitudes.

Network parameters and their definitions.

Mismatch detection in simulations and experiment.

In all cases, the stimulus was 8 repeats each of 4 patterns, at the specified frequency. In panels B,C,D the red triangles indicate time of pattern change. A: Deterministic simulations. Excitatory and inhibitory currents contribute different terms to the mismatch signal. Excitatory currents detect the transition, whereas inhibitory currents undergo synaptic depression to balance depression in excitatory currents. B: Same simulations. At each mismatch there is an elevation of EPSP, shown normalized around the mean signal. C: Same as B except using the model configured for 37° C. D: Experimental data. Colored scatter points indicate individual trials, the different colors represent different patterns. The solid blue line is the mean. A significant (p=0.013, Wilcoxon signed-rank left-tailed test) mismatch signal is seen for the third transition on Diii. E: Summary heatmaps showing percent of significant transitions (p < 0.05, Wilcoxon signed-rank left-tailed test) for experiments (Ei) and simulations (Eii and Eiii). In the simulations we assume 2mV of 500Hz lowpass filtered noise. Eii: no jitter in stimulus timing. Eiii: 6ms half-Gaussian distributed jitter. The introduction of jitter substantially changes the probability of selective transitions, however the simulated probabilities remain somewhat different from experiment.

Spiking responses are selective for mismatch in patterned sequences.

All runs were performed using a spiking neuron model with stochastic synaptic chemical kinetics and each run consisted of 50 trials. Red triangles indicate transition times between patterns. Asterisks over the red triangles indicate significance of transitions. For panels A to E and H to K, the stimulus consists of eight repeats of each four patterns as in AAAAAAAABBBBBBBBCCCCCCCCDDDDDDDD. Traces C-K and N-O are spike rates with a moving window of 10 ms and 5 ms respectively. In C-K, Mismatch statistics use the binomial test to compare the counts of the three pulses immediately before vs. the three immediately after a transition. A: Spiking responses on three illustrative trials. B. Raster plot of spiking for reference stimulus. C: Reference model exhibits mismatch responses for the first two pattern transitions (p=0.002, 0.032, 0.434). D: Dense stimuli (12.5% spots nulled) do not give any mismatch signals. E: Sparse patterns (94% of spots nulled) elicit a mismatch response in two transitions (p=1.5e-5, 6.1e-5). F: Oddball response. There was a strong and pattern selective mismatch response (p=9.5e-7) to the third oddball (deviant) stimulus. Oddballs were presented every 8th pulse, against a uniform background pattern: AAAAAAAABAAAAAAACAAAAAAADAAAAAAA. G: Gap stimulus. Similar to oddballs, gaps were presented every 8th pulse, but instead of a deviant stimulus, no stimulus pattern was delivered. There was no significant response, and activity declines over the course of the trial. H. Lack of STP in Glu removes mismatch detection, and spike rates are high. I. Lack of STP in GABA results in sustained strong inhibition and sparse firing, but some mismatch selectivity remains for the second transition (p=0.004). J: Control with uniform pattern (A repeated 32 times). The spiking response rapidly drops to close to zero. K. Control with randomized pattern (one of pattern A to pattern E) on each pulse. There is a slow decay over the 32 pulses. L: Mismatch response is tuned to frequency of pattern repeats. The strongest responses are in the gamma range between 50 and 100 Hz. Response compares three post-transition vs three pre-transition pulses. Y axis is ratio, p-values are Binomial test for spike counts comparing post vs pre.). O: Schematic of gamma burst stimulus (green) repeated at theta frequency of 7.69 Hz (blue). Each burst consisted of a pattern repeated 5 times at 100 Hz. Panels N and O use the Binomial test to compare spike counts in the first theta cycle with those in the next three cycles. N: Theta modulated gamma burst responses showed strong spiking responses when each burst had a different pattern. O: Spiking in second, third and fourth theta cycles was lower than the first when each burst had the same pattern (p=1.9e-07, 3.1e-07, 6.68e-21).

Parameter sensitivity of network model.

A: Schematic of model, indicating the six network parameters. Overlap specifies the fraction of postsynaptic axons which are activated by more than one optical spot as pattern density increases. WGlu and WGABA specify synaptic weights for gluR and GABAR synapses, measured as 1/ohm.m2. PCA3_CA1 is the probability that a given CA3 neuron connects to the CA1 neuron. Similarly for PCA3_inter and PInter_CA1. B: Pattern selectivity as a function of the six parameters. Red triangles indicate parameter values used for the reference model. Note the distinct selectivity profile for 8, 20 and 50 Hz, where 50Hz is almost always more selective over the entire range of parameters. C: Frequency of greatest mismatch detection as a function of the network parameters. The highest mismatch detection frequency can be tuned over a wide range by network parameters. D: Effect of global network changes on mismatch detection, compared to baseline deterministic model. Stochasticity significantly lowers mismatch detection (T=-2.33 p=0.022), but lack of STP completely eliminates it (T=-6.87, p=1.6e-9). Raising the temperature of the network to 37 degrees improves mismatch detection (T=2.88, p=0.0048). Introduction of spiking greatly increases trial-to-trial variability but there is no difference of the mean compared to the reference (T=0.007, p>0.99). For each of the runs and each frequency, the ratio (post-pre)/(post+pre) was taken for the 3 pulses before (pre) and after (post) the transition. % change = 100*(run_mean – baseline_mean)/baseline_mean. The three frequency terms were merged and Welch’s T-Test was used to compare each run against the baseline run.

Proposed mechanisms for mismatch detection.

A: Fresh-afferent model. Left: Circuit involving only excitatory synapses from CA3 to CA1. The two input patterns p and q (in the CA3) have only a little overlap. Right: on successive stimulus p repeats, the STP in the synapses (indicated by spine size) leads first to an elevated response, and then to depression. When the stimulus changes to pattern q, most of the afferents are fresh so we get an elevated response. B: EI-balance shift model. The excitatory component of the circuit is as in A, but in addition there is an intermediate interneuron layer receiving input from the CA3. During a train of pulses of pattern p, E transiently exceeds I, then both undergo STD. When a different pattern q arrives, there is again a transient advantage for the E input because the two-stage connectivity leads to substantially overlapping input on the I inputs, so most of them are still depressed. Overall, the transient response at p2 and q1 are closer in amplitude, or equivalently, the pattern selectivity is less dependent on the averaged activity history.

A) Heatmap of optogenetic depolarisation of a single CA3 pyramidal cell obtained by single spots of light in a 24×24 grid while keeping the patched PC in the centre of the frame (asterisk). B) Distribution of optical depolarization in CA3 PC against patterns of size 1 square (single spots), 5 squares, or 15 squares. Filled circles denote the instances when the patched cell fired an action potential. For patterns of size 5 and 15, all the trials resulted in spikes, hence there is no subthreshold distribution of EPSPs. For 1-square patterns, four out of 576 spots induced a spike.

Optically-triggered ‘ringing’ response in CA3 field recordings.

A: Sample traces. B. Methodology for computing peak heights. For each peak we computed the fEPSP elevation with respect to the immediately preceding valley. C. Peaks decline rapidly. The median of the second peak was ∼40% of the first peak, but the third peak was < 5%. D. Peak times following optical pulse (of 2 ms). The third peak occurred within <10ms. E: fEPSP Peak Width distribution centred around 1.2 ms, but no peak was wider than 1.6 ms, suggesting tight synchrony in case multiple CA3 neurons were spiking.

Fitting the presynaptic kinetic model to EPSC and IPSC data, for cell 7492.

A: Model of reaction system in each bouton. Numbers next to reaction arrows indicate stoichiometry. Synaptic input triggers a reaction causing extracellular Ca2+ to enter the bouton, and a pump removes the Ca. CaM buffers the Ca2+. Ca binds/unbinds from successive stages of the readily releasable (RR) pool of vesicles till they are docked, at which point the final Ca-binding step causes synaptic release. The released neurotransmitter (Glu for excitatory synapses, GABA for inhibitory) opens a ligand-gated receptor channel on the postsynaptic CA1 neuron, which is held voltage-clamped.(green arrow). The reaction scheme was identical for Glutamatergic and GABAergic synapses, but the rates were different to fit to the voltage-clamp recordings. B-E: EPSC fits for 20, 30, 40 and 50 Hz, respectively. Blue traces are experimental peaks and valleys averaged over all repeats and all 5-square patterns. Red traces are simulated EPSCs. F: Cumulative release curve for glutamate vesicles. Blue: expected transmitter release completion within 10 ms based on our observations of rapid postsynaptic responses from Figure 1 H,K. Red: simulated release. G: Distribution of model fitting scores for all cells and for glu and GABA receptors. Scores are normalised RMS differences between experimental and simulated points, as in panels B-F. Any score below 0.3 is a good fit.

Example model fits for cell 6201.

Here, the 40 Hz data was not recorded. A, B, C: EPSC fits for 20, 30, and 50 Hz, respectively. Blue traces are experimental peaks and valleys averaged over all repeats and all 15-square patterns for neuron 6201. Red traces are simulated rectified EPSCs. Note that the experimental traces are noisy hence the fits miss some of the peaks. D: Cumulative release curve for glutamate vesicles. Blue: target release complete within 10 ms. Red: simulated release. E: Distribution of model fitting scores for all cells and for glu and GABA receptors. Scores are normalized RMS differences between experimental and simulated points, as in panels A-D. Any score below 0.3 is a good fit.

Example model fits for Inhibitory synapses on cell 7492.

A, B, C, D: IPSC fits for 20, 30, 40, and 50 Hz, respectively. Blue traces are experimental peaks and valleys averaged over all repeats and all 15-square patterns for neuron 7492. Red traces are simulated IPSCs. Note that the experimental traces are noisy hence the fits miss some of the peaks. E: Cumulative release curve for GABA vesicles. Blue: target release complete within 10 ms. Red: simulated release. F: Distribution of model fitting scores for all cells and for glu and GABA receptors. Scores are normalized RMS differences between experimental and simulated points, as in panels A-D. Any score below 0.3 is a good fit.

STP profiles for each of the 12 voltage-clamp recordings with the pulse-train stimulus.

Some cells were not held long enough to complete all four frequencies.

Presynaptic signalling model dynamics.

All quantities are units of numbers of molecules per bouton unless otherwise stated. Signalling calculations are stochastic. Note that these are only two representative active synapses from among the 100 excitatory and 200 inhibitory synapses on the cell, and there are other synapses with distinct activity. Currents are reported as total currents for all synapses of the specified type. A: Optical stimulus train and corresponding somatic potential. B: Glutamaterigic bouton signalling. C. GABAergic bouton signalling.

Stimulus patterns used in simulations to activate the CA3 layer.

Green spots are on, blue are off. A: Dense patterns used as equivalents to 5-square experimental stimuli, with 12.5% of spots set to zero, resulting in 78±2 spots per pattern. Overlap between successive patterns is 31, 23, and 37 spots. B. Reference sparse patterns to map to 5 square stimuli, with 75% of spots set to zero, resulting in 21±3 spots per pattern. Overlap between successive patterns is 1,1, and 5 spots. C. Simulation implementation of 15-square patterns, with 144, 152, 152, and 160 spots per pattern. Overlap is 68, 80 and 88 spots..

Time course of charging and desensitisation of optically stimulated CA3 neuron model.

At 0.2 seconds we deliver a burst with inter-pulse intervals of 5, 10, 15, 20, 25, 30 ms. At t=0.7 seconds we deliver a single pulse. At t=1 seconds we deliver a train of 32 pulses at the frequencies of 8, 20 and 50 Hz as indicated above each panel.

EPSP response probability rises with larger presynaptic volume.

Reference volume used in simulations is scaled by 0.2, which corresponds to a synaptic volume of 0.0205±0.0008 femtolitres.

Response distributions for Poisson pulse train: field and EPSP data.

A: Sample data. Blue: Raw EPSP data, maroon: field potentials, green: stimulus pulse trigger. B-M: 5 square pattern responses in blue, 15-square pattern responses in orange. B, C: Scatter plots of EPSP vs fEPSP for 5 and 15 square input, respectively. Linear regression fit is plotted in black. P-value < 0.001 in both cases. D: Field EPSP as a function of time. E: EPSP as a function of time. F. Probability that a pulse will elicit an fEPSP peak. This is almost 100% reliable. G. Probability that a pulse will elicit an EPSP. Note that the initial ∼2 seconds of the train have fewer events of low probability. H: Field EPSP as a function of inter-stimulus interval (ISI) for the pair of stimuli immediately preceding the fEPSP. I: EPSP as a function of ISI. Note that, unlike panel H, short ISIs elicit larger EPSPs. J: Distribution of field EPSP responses. K: Distribution of EPSP responses. This differs qualitatively from panel J, except that in both cases, the 15-square distributions are shifted over to the right. L: distribution of standard deviation in field EPSP for successive trials of each given pattern. M: Distribution of standard deviation in EPSP for successive trials for each pattern.

Contributions to EPSP decline with time and ISI.

Yellow dots: 15 square stimuli, blue dots: 5 square stimuli. A: Contribution of STP to the initial dip in EPSPs. Ai: Experimental data. Note the decline in EPSP from t = 0 to t ∼2 seconds. Aii. Reference simulated model has a similar dip in EPSP. Aiii. Simulated model with STP disabled. In this case the 5 square case has a dip in EPSP, but not the 15 square. Aiv: Incorporation of NMDAR activity has negligible effect because the potential is well below that for release of magnesium block. B: Contributions to the dependence of EPSP on ISI. Yellow dots: 15 square stimuli, blue dots: 5 square stimuli. Bi: Experiment. There is a sharp decline, tau_5Sq = 0.3 to 18ms, tau_15Sq=8ms to 1s. Bii: Reference model. Tau_5sq=34.3±6.94ms, Tau_15sq=49.7±12.5ms. Biii: Model without STP in either GluR or GABAR synapses. Tau_5sq=58.7±9.8ms, Tau_15sq=117±35ms. The ISI dependence remains, though the time-course is slower in the absence of STP. Note that the response amplitudes are larger in the absence of STP, suggesting that presynaptic depression reduces EPSP amplitudes in the reference but not non-STP model. Biv: Model without NMDA receptors. The response is very similar to the reference response, suggesting that most of the current is carried by the Glu receptors, and the NMDA receptor opening does not lead to amplification of EPSP at short ISIs.

Overlap among synapses stimulated by each of 8 stimulus patterns.

Patterns 0 through 4 are 5-square patterns, and patterns 5,6,7 are 15-square patterns. Diagonals indicate self-overlap, and are therefore the count of the number of stimulated synapses. A: Glutamatergic synapses. B: GABAergic synapses.

Overlap between CA1 synapses following 15 square pattern declines with the model construction parameter ZeroIndices.

ZeroIndices is the number of CA3 inputs set to zero, out of a maximum of 256.