Adaptation invariant concentration discrimination in an insect olfactory system

Reviewer #1 (Public Review):

The authors use electrophysiological and behavioral measurements to examine how animals could reliably determine odor intensity/concentration across repeated experiences. Because stimulus repetition leads to short-term adaptation evidenced by reduced overall firing rates in the antennal lobe and firing rates are otherwise concentration-dependent, there could be an ambiguity in sensory coding between reduced concentration or more recent experience. This would have a negative impact on the animal's ability to generate adaptive behavioral responses that depend on odor intensities. The authors conclude that changes in concentration alter the constituent neurons contributing to the neural population response, whereas adaptation maintains the 'activated ensemble' but with scaled firing rates. This provides a neural coding account of the ability to distinguish odor concentrations even after extended experience. Additional analyses attempt to distinguish hypothesized circuit mechanisms for adaptation but are inconclusive. A larger point that runs through the manuscript is that overall spiking activity has an inconsistent relationship with behavior and that the structure of population activity may be the more appropriate feature to consider.

To my knowledge, the dissociation of effects of odor concentration and adaptation on olfactory system population codes was not previously demonstrated. This is a significant contribution that improves on any simple model based on overall spiking activity. The primary result is most strikingly supported by visualization of a principal components analysis in Figure 4. However, there are some weaknesses in the data and analyses that limit confidence in the overall conclusions.

1. Behavioral work interpreted to demonstrate discrimination of different odor concentrations yields inconsistent results. Only two of the four odorants follow the pattern that is emphasized in the text (Figure 1F). Though it's a priori unlikely that animals are incapable of distinguishing odor concentrations at any stage in adaptation, the evidence presented is not sufficient to reach this conclusion.
2. While conclusions center on concepts related to the combination of activated neurons or the "active ensemble", this specific level of description is not directly demonstrated in any part of the results. We see individual neural responses and dimensional reduction analyses, but we are unable to assess to what extent the activated ensemble is maintained across experience.
3. There is little information about the variance or statistical strength of results described at the population level. While the PCA presents a compelling picture, the central point that concentration changes and adaptation alter population responses across separable dimensions is not demonstrated quantitatively. The correlation analysis that might partially address this question is presented to be visually interpreted with no additional testing.
4. Results are often presented separately for each odor stimulus or for separate datasets including two odor stimuli. An effort should be made to characterize patterns of results across all odor stimuli and their statistical reliability. This concern arises throughout all data presentations.
5. The relevance of the inconclusive analysis of inferred adaptation mechanisms in Figure 2d-f and the single experiment including a complex mixture in Figure 7 to the motivating questions for this study are unclear.
6. Throughout the description of the results, typical standards for statistical reporting (sample size, error bars, etc.) are not followed. This prevents readers from assessing effect sizes and undermines the ability to assign a confidence to any particular conclusion.

Reviewer #2 (Public Review):

Summary:
The authors' main goal was to evaluate how both behavioral responses to odor, and their early sensory representations are modified by repeated exposure to odor, asking whether the process of adaptation is equivalent to reducing the concentration of an odor. They open with behavioral experiments that actually establish that repeated odor presentation increases the likelihood of evoking a behavioral response in their experimental subjects - locusts. They then examine neural activity patterns at the second layer of the olfactory circuit. At the population level, repeated odor exposure reduces total spike counts, but at the level of individual cells there seems to be no consistent guiding principle that describes the adaptation-related changes, and therefore no single mechanism could be identified.

Both population vector analysis and pattern correlation analysis indicate that odor intensity information is preserved through the adaptation process. They make the closely related point that responses to an odor in the adapted state are distinct from responses to lower concentration of the same odor. These analyses are appropriate, but the point could be strengthened by explicitly using some type of classification analysis to quantify the adaptation effects. e.g. a confusion matrix might show if there is a gradual shift in odor representations, or whether there are trials where representations change abruptly.

Strengths:
One strength is that the work has both behavioral read-out of odor perception and electrophysiological characterization of the sensory inputs and how both change over repeated stimulus presentations. It is particularly interesting that behavioral responses increase while neuronal responses generally decrease. Although the behavioral effect could occur fully downstream of the sensory responses the authors measure, at least those sensory responses retain the core features needed to drive behavior despite being highly adapted.

Weaknesses:
Ultimately no clear conceptual framework arises to understand how PN responses change during adaptation. Neither the mechanism (vesicle depletion versus changes in lateral inhibition) nor even a qualitative description of those changes. Perhaps this is because much of the analysis is focused on the entire population response, while perhaps different mechanisms operate on different cells making it difficult to understand things at the single PN level.

From the x-axis scale in Fig 2e,f it appeared to me that they do not observe many strong PN responses to these stimuli, everything being < 10 spikes/sec. So perhaps a clearer effect would be observed if they managed to find the stronger responding PNs than captured in this dataset.

Reviewer #3 (Public Review):

Summary:
How does the brain distinguish stimulus intensity reduction from response reductions due to adaptation? Ling et al study whether and how the locust olfactory system encodes stimulus intensity and repetition differently. They show that these stimulus manipulations have distinguishable effects on population dynamics.

Strengths:
1. Provides a potential strategy with which the brain can distinguish intensity decrease from adaptation. -- while both conditions reduce overall spike counts, intensity decrease can also changes which neurons are activated and adaptation only changes the response magnitude without changing the active ensemble.
2. By interleaving a non-repeated odor, they show that these changes are odor-specific and not a non-specific effect.
3. Describes how proboscis orientation response (POR) changes with stimulus repetition., Unlike the spike counts, POR increases in probability with stimulus. The data portray the variability across subjects in a clear way.

Weaknesses:
1. Behavior
a. While the "learning curve" of the POR is nicely described, the behavior itself receives very little description. What are the kinematics of the movement, and do these vary with repetition? Is the POR all-or-nothing or does it vary trial to trial?

b. What are the reaction times? This can constrain what time window is relevant in the neural responses. E.g., if the reaction time is 500 ms, then only the first 500 ms of the ensemble response deserves close scrutiny. Later spikes cannot contribute.

c. The behavioral methods are lacking some key information. While references are given to previous work, the reader should not be obligated to look at other papers to answer basic questions: how was the response measured? Video tracking? Hand scored?

d. Can we be sure that this is an odor response? Although airflow out of the olfactometer is ongoing throughout the experiment, opening and closing valves usually creates pressure jumps that are likely to activate mechanosensors in the antennae.

e. What is the baseline rate of PORs in the absence of stimuli?

e.What can you say about the purpose of the POR? I lack an intuition for why a fly would wiggle the maxillary palps. This is a question that is probably impossible to answer definitively, but even a speculative explanation would help the reader better understand.

2. Physiology
a. Does stimulus repetition affect "spontaneous" activity (i.e., firing in the interstimulus interval? To study this question, in Figures 2b and c, it would be valuable to display more of the pre-stimulus period, and a quantification of the stability or lability of the inter-stimulus activity.

b. When does the response change stabilize? While the authors compare repetition 1 to repetition 25, from the rasters it appears that the changes have largely stabilized after the 3rd or 4th repetition. In Figure 5, there is a clear difference between repetition 1-3 or so and the rest. Are successive repetitions more similar than more temporally-separated repetitions (e.g., is rep 13 more similar to 14 than to 17?). I was not able to judge this based on the dendrograms of Figure 5. If the responses do stabilize at it appears, it would be more informative to focus on the dynamics of the first few repetitions.

c. How do temporal dynamics change? Locust PNs have richly varied temporal dynamics, but how these may be affected is not clear. The across-population average is poorly suited to capture this feature of the activity. For example, the PNs often have an early transient response, and these appear to be timed differently across the population. These structures will be obscured in a cross population average. Looking at the rasters, it looks like the initial transient changes its timing (e.g., PN40 responses move earlier; PN33 responses move later.). Quantification of latency to first spike after stimulus may make a useful measure of the dynamics.

d. How legitimate is the link between POR and physiology? While their changes can show a nice correlation, the fact the data were taken from separate animals makes them less compelling than they would be otherwise. How feasible is it to capture POR and physiology in the same prep? This would be most helpful, but I suspect may be too technically challenging to be within scope.

Author Response

We thank the reviewers for their useful and constructive comments. In this provisional response, we will address a few of the major issues and plan to submit a detailed, point-by-point response along with the revised manuscript.

1. Robustness of activated combination of neurons (the ‘activated ensemble’).

The reviewers have asked for additional analyses and visualization of the group of neurons activated and a classification analysis to illustrate the point that the activated set of neurons would allow discrimination between different concentrations even after the spiking activity reduced significantly in the later trials. We relied on visualization using PCA (Manuscript Fig. 4) and quantification using correlation analysis (Manuscript Fig. 5a and Manuscript Supplementary Figure 2). But this point can be easily amplified further to support our conclusions and address a major concern raised by the reviewers.

Visualization of neural responses across trials and odorants: As recommended, we followed the procedures used in Stopfer et al., 2003 (Fig. 6c) and Miura et al., 2012(Fig. 3C) to image neural responses across recorded PNs as a matrix (Author response image 1).

Author response image 1.

Author response image 1: Spike counts averaged over the entire 4s odor presentation window across all recorded neurons are shown as a function of trial number (columns). The sorting is same across different panels. Note that there are 80 neurons whose response was monitored for hexanol and octanol responses (Dataset 1; first row of panels), and 81 neurons whose response was monitored for isoamyl acetate and benzaldehyde (Dataset 2; second row of panels). As can be noted, across the 25 trials the pattern of activation remains consistent. Also, the activated combination of neurons varied robustly with odor identity and intensity.

Classification analysis: To illustrate that there is enough information to recognize an odorant and discriminate between different intensities, we performed a leave-one-trial-out classification analysis. The left-out trial was assigned the class label of its nearest neighbor (using correlation distance metric). The results from this classification analysis are shown below in Author response image 2. As a control, we shuffled the odor class labels and repeated the leave-one-trial-out classification analysis.

Author response image 2.

Author response image 2: Results from classification analysis are shown for the two datasets: hexanol–octanol at different concentrations (dataset 1; 80 PNs), and isoamyl acetate and benzaldehyde (dataset 2; 81 PNs). We did a leave-onetrial-out validation. The true odor label is shown along the x-axis and the predicted odor label is shown along the yaxis. As can be noted, the class labels for every single trial were correctly predicted in both datasets. The result after class labels were shuffled is also shown for comparison. These results strongly support our conclusion that odor intensity information is preserved and odor concentration can be recognized independent of adaptation.

Correlation with the first trial:

We had shown the correlation across odorants and concentrations as a function of the trial (manuscript Figure 5A). To complement these analyses, here we focus on the correlations with the response evoked in the first trial of each odorant at high and low concentrations and plot this information as a function of trial number (Author response image 3, 4). As can be noted, the correlation across different trials of a given odorant at specific concentrations remains much higher than all other conditions.

Author response image 3.

Author response image 3: (top-left) Correlation between 80-dimensional neural responses (averaged over the entire 4s odor presentation window) with the first trial of hexanol at high intensity (hex-H; 1% v/v) is plotted as a function of trial number. (top-right) similar plots but correlation computed with neural responses evoked during the first trial of octanol at high intensity (oct-H; 1% v/v). (bottom-left) similar plots but correlation computed with neural responses evoked in the first trial of hexanol at low intensity (hex-L; 1% v/v). (bottom-right) similar plots but correlation computed with neural responses evoked in the first trial of octanol at low intensity (oct-L; 1% v/v).

Author response image 4.

Author response image 4: (top-left) Correlation between 81-dimensional neural responses (averaged over the entire 4s odor presentation window) with the first trial of isoamyl acetate at high intensity (iaa-H; 1% v/v) is plotted as a function of trial number. (top-right) similar plots but correlation computed with neural responses evoked in the first trial of benzaldehyde at a high intensity (bza-H; 1% v/v). (bottom-left) similar plots but correlation computed with neural responses evoked in the first trial of isoamyl acetate at low intensity (iaa-L; 1% v/v). (bottom-right) similar plots but correlation computed with neural responses evoked in the first trial of benzaldehyde at low intensity (bza-L; 1% v/v).

Behavioral significance and dynamics: The reviewers had wondered about the relevance of the behavior to the organism. The maxillary palps are sensory organs close to the mouth parts that are used to grab food and help with the feeding process. In our previous studies, we had shown that these palpopening responses are innately triggered by many ‘appetitive odorants.’ However, the probability of palp opening varied across different odorants (Chandak and Raman, 2023). Some odorants evoked higher palp-opening responses and others reduced the probability of palp-opening response (below the median value across odorants). Since all other parameters (such as the clicking sound of valves, and mechanical cues due to airflow during odor presentation), are the same across these different odorants, these observed differences in palp-opening response probability are attributed to the identity of the odorants presented.

Author response image 5.

Author response image 5: Preference indices were calculated for all odors tested and are shown as a bar plot (n = 26 locusts). Blue bars indicate odors classified as appetitive, gray bars indicate neutral odors and red bars indicate unappetitive odors. Locusts with a significant deviation from the median response (one-sided binomial test, P < 0.1, were classified as either being appetitive or unappetitive; *P < 0.1, **P < 0.05, ***P < 0.01). Error bars indicate s.e.m. [Replotted Fig 1.c from Chandak and Raman, 2023].

We had also shown that we could train locusts to have stereo-typed palp-opening responses using the classical conditioning approach (odor – odor-conditioned stimulus and food reward – unconditioned stimulus; Video: https://static- content.springer.com/esm/art%3A10.1038%2Fncomms7953/MediaObjects/41467_2015_BFncomms7953 _MOESM483_ESM.mov; Saha et al., 2015). The dynamics of those conditioned palp-opening responses have been well characterized.

We will use similar tracking procedures to monitor and quantify the dynamics of innate palp-opening responses as well. We will add supplementary videos to fully capture this behavior.

Early vs. late neural responses:

Since behavioral responses are more likely to start as soon as the odorant is presented, the reviewers wondered whether there are differences in the observed findings if we focus only on the early neural activity (as it might be more important to triggering behavior). Note that the median response time for conditioned palp-opening responses is less than 750 ms (Saha et al., 2015, Chandak and Raman, 2023). Hence, we divided the neural dataset and analyzed the neural response patterns during these early (0-750 ms after onset) and late (2-4 s after odor onset) time windows. In both these epochs, we found that the total spike counts across neurons reduced as a function of trial number or repetition and the combination of neuron activated remained robust (Author response images 6-11). Hence, we conclude that while the neural responses in different time windows would be important for shaping other parameters of behavioral response dynamics, the overall behavioral response probability that we used in our analysis had a similar relationship with early, late, or total neural activity during the entire odor presentation (i.e. time-window of the neural response did not matter for the analyses presented in the manuscript).

Author response image 6.

Author response image 6: Total spike counts reduced as a function of trial number. This reduction was observed for the total spike counts during the entire odor presentation window and during both the early (0-750 ms) and late (2-4 s) response time windows. Dataset 1: 80 PNs, hexanol, and octanol odorants.

Author response image 7.

Author response image 7: Total spike counts reduced as a function of trial number. This reduction was observed for the total spike counts during the entire odor presentation window and during both the early (0-750 ms) and late (2-4 s) response time windows. Dataset 2: 81 PNs, isoamyl acetate, and benzaldehyde odorants.

Author response image 8.

Author response image 8: Similar plots as in Figures 3 and 4 but analyzing 80-dimensional spike count vectors calculated using only the first 750 ms of odor-evoked response. Note that the correlation with the odor evoked response in the first trial remains high across trials. But between different odorants or different intensities of the same odorant, the response correlation drops significantly. Dataset 1: 80 PNs, hexanol, and octanol odorants.

Author response image 9.

Author response image 9: Similar plots as in Figures 3 and 4 but analyzing 80-dimensional spike count vectors calculated using only the last 2 seconds of odor-evoked response. Note that the correlation with the odor evoked response in the first trial remains high across trials. But between different odorants or different intensities of the same odorant, the response correlation drops significantly. Dataset 1: 80 PNs, hexanol, and octanol odorants.

Author response image 10.

Author response image 10: Similar plots as in Figures 3 and 4 but analyzing 80-dimensional spike count vectors calculated using only the first 750 ms of odor-evoked response. Note that the correlation with the odor evoked response in the first trial remains high across trials. But between different odorants or different intensities of the same odorant, the response correlation drops significantly. Dataset 2: 81 PNs, isoamyl acetate, and benzaldehyde odorants.

Author response image 11.

Author response image 11: Similar plots as in Figures 3 and 4 but analyzing 80-dimensional spike count vectors calculated using only the last 2 seconds of odor-evoked response. Note that the correlation with the odor evoked response in the first trial remains high across trials. But between different odorants or different intensities of the same odorant, the response correlation drops significantly. Dataset 2: 81 PNs, isoamyl acetate, and benzaldehyde odorants.

Other Statistical Tests:

The reviewers felt that in many analyses, we did not include error bars to indicate the sample size, SEM, or SD. We will fix this by adding the sample size information to each panel and as appropriate. However, we would also like to point out that many of the analyses are done in a trial-by-trial fashion (e.g. Manuscript Figures 3 – 6). For these analyses, it would not be possible to add SEM or SD. One condition (hex -H or iaa-H) was repeated in each dataset, and we have added them in the results shown in this response letter to demonstrate repeatability. We will strive our best to add these statistics as would be appropriate, but this cannot be done for the trial-by-trial analyses.

References:

Stopfer M, Jayaraman V, Laurent G. Intensity versus identity coding in an olfactory system. Neuron. 2003 Sep 11;39(6):991-1004. doi: 10.1016/j.neuron.2003.08.011. PMID: 12971898.

Miura K, Mainen ZF, Uchida N. Odor representations in olfactory cortex: distributed rate coding and decorrelated population activity. Neuron. 2012 Jun 21;74(6):1087-98. doi: 10.1016/j.neuron.2012.04.021. PMID: 22726838; PMCID: PMC3383608.

Chandak, R., Raman, B. Neural manifolds for odor-driven innate and acquired appetitive preferences. Nat Commun 14, 4719 (2023). https://doi.org/10.1038/s41467-023-40443-2

Saha, D., Li, C., Peterson, S. et al. Behavioural correlates of combinatorial versus temporal features of odour codes. Nat Commun 6, 6953 (2015). https://doi.org/10.1038/ncomms7953