Learning is a fundamental source of individuality

Reviewer #1 (Public review):

"Learning is a fundamental source of individuality," by Manna and colleagues, interrogates different sources of variation in individual behavior. The authors place individual flies in a Y-shaped arena, which is a common design in the field, and illuminate the arms of the Y with blue versus green light. They track the color preference of individual animals and also perform operant conditioning, meaning that they teach the fly to avoid a particular color/arm by generating a foot shock when the fly enters that arm. There are a number of things that are impressive about this setup: The authors are able to collect data on thousands of individual flies of many different strain backgrounds, and they demonstrate a strong change in color preference after conditioning. This is nice, because in past papers, visual learning ability has been modest and difficult to study. To put a number on it, in this paper, animals on average don't show a color preference at the start of the assay, spending around 30% of their time in the one arm illuminated green, and the remaining time in the two arms illuminated blue. After conditioning, the average animal spends only 23% of its time in the green arm.

The authors run 64 animals through the assay for each of 88 wild-type strains (maybe? see Major Point 1 below) and see considerable strain-specific (genetic) variation in the change in time spent in the shocked color after conditioning. Some strains show no learning, while others spend <10% of their time in the shocked color after conditioning. They also, I believe, see that some strains have more variability across individuals, which would suggest that some strains have stronger canalization at the development or circuit function level than others, i.e., some genotypes produce more consistent copies of the individual, others less consistent copies. (Or, some genotypes produce robust circuits, and others produce noisy circuits.)

Finally, the authors argue statistically that learning itself increases variability in individual performance. This makes a lot of sense to me intuitively. Learning changes the physical/chemical properties of circuits in the brain, and because it evolves over time and interacts with environmental variables, it seems like it should send different animals down different channels. Or, at a conceptual level, if I learn to play the piano and my sister doesn't (because of some genetic difference between us or something stochastic), this learning experience will cause all sorts of other differences in our behavior as time passes. I also think the authors do have enough data to be able to make this finding. However, the presentation of the argument in this portion of the paper is hard for me to understand, and I am not an expert in statistics, so the strength of the result is difficult for me to evaluate.

Major points

(1) It's difficult to track through the paper the number of animals tested for different assays. At the beginning, it says N=5632, which works out to 64 flies for each of the 88 DGRP strains. 64 happens to be the number of parallel Y arenas they have. Later in the methods, there's a description of more variation within the set of 64 for each strain, two different parent sets per strain, different sexes, conditioned and unconditioned. And, while the results text focuses on the color learning, the methods discuss additional assays (place learning, multi-day learning).

Given the numbers, does each run of the 64 mazes include all the tested flies of one strain, or are flies of many strains included in each batch? Do different flies do different assays (color, place, multi-day), or do they all do all the assays? Perhaps there is a table including this information already in the supplement, but I recommend making it much clearer in the main results text and methods. While the dataset is large, if it is split over many conditions and/or if batch and genotype confound each other, this will affect the robustness of the results and how strong the conclusions can be.

(2) The data presentation in Figure 1 is elegant and easy to follow, but getting into Figure 2 and subsequently, I get lost in the statistics and have trouble understanding what is being measured. My understanding of the big picture is that while genetics and individual randomness contribute a lot to behavior, the evidence for learning as an amplifier of individuality is that variance in behavior among animals of the same strain increases over time in the conditioned group (i.e., the group that is doing the most learning, or a specific kind of learning), but not in the control group. This idea is illustrated in the flattening distributions in the cartoons in Figure 1A. The authors should include graphs of the real data that use the same format as in that cartoon. Instead, the graphs present "residuals," and I don't know what those are. I suspect it's "variation left over after accounting for effects of strain and individual stochasticity." I see the residuals being tracked per strain over time in Figure 2H, but I don't see the change over time in other graphs. I'm looking for something simple like, "variation within the strain at the beginning of learning and at later time points in learning." (But I'm not sure exactly what instantaneous measurement would be the focus in longitudinal analyses of learning behavior.)

(3) Figure 3 is a cool stab at tracking down the precise mechanism by which a stochastic environment interacts with learning to send individuals along different behavioral routes. But again, like in Figure 2, I don't have the sophisticated understanding of statistics to understand exactly what the graphs are telling me, or how they relate to the underlying measurements. I'm relying on the results text alone to reach a conceptual understanding, and just taking the graphs on trust.

So, overall, the authors have a very nice body of work here, and with the potential to add a new facet to our understanding of the origins of diversity in animal behavior. In addition to the interpretations they focus on here, this dataset also represents an advance in studying visual associative learning in general, and quite an amazing ability to make longitudinal measurements of many behavioral decisions within the same animals. Improving the data presentation to make it easier to follow for a larger swathe of researchers, especially in figures 2 and 3, will increase its potential impact.

Reviewer #2 (Public review):

Summary:

The authors set out to test the extent to which differences in learning capacity and experience contribute to behavioural variation in a genetically identical population under identical environmental conditions.

Strengths:

The authors developed and used a scaled-up version of a simple two-choice behavioural paradigm, allowing them to test thousands of individuals across multiple genotypes. They then deployed clever and powerful statistical analysis methods and provided compelling evidence for a role of variability in learning in the expression of behavioural variation.

Weaknesses:

There are no major weaknesses, although some level of longitudinal analysis to strengthen the evidence for a strict definition of individuality would be a welcome extension of a future study. In addition, it would have been very interesting, although understandably beyond the current scope, to delineate a potential source of learning variability in the brain.

Author response:

Reviewer 1:

Clarification of sample sizes, assay structure, and experimental design.

Reviewer 1 noted that the number of animals tested across strains, assays, sexes, parent sets, conditioned and unconditioned groups, and longitudinal conditions is difficult to track through the manuscript. Given the extent of the experimental and data processing procedures such as filtering for inactive or injured flies, we agree that a summary table and/or a visual schematic of the full experimental setup would be helpful.

Importantly, the vast majority of individuals was used for the main experiment where we conditioned the flies to avoid the green arm, and where the colors of the arms were fixed throughout the assat. A smaller number of flies were tested in the validation experiments (such as different types of conditioning). In each experiment, 64 flies were always set up per genotype and their behaviour was measured in parallel. Usually, around ~60 flies passed the filtering step before analysis (filtering due to inactivity or injured flies). Among those 60-ish flies per genotype the distribution of flies of different sex or flies raised in different replicate vials was balanced. Different individual flies were tested across different assays, except in the multiday experiment, where each individual was tested across four different assays.

We will add a supplementary summary that includes how many flies were tested across assays, how individuals, males, females, replicates and genotype were distributed across batches (and in the multiday experiments how they were distributed across experiments), and how many flies were filtered out from the final analysis. 

Clearer presentation of the statistical argument that learning amplifies individuality.

Reviewer 1 also noted that the presentation of the statistical analyses, particularly in Figure 2, was difficult to follow (e.g. what is residual individuality, how is it tracked over time, and why not replace it with something simpler like variance?).

Our experimental design combines multiple, replicated environments and genotypes. For example, genetically identical flies from genotype A, are raised under identical developmental environments that are replicated two times in two vials. The same is true for genotype B. Individuals from both genotypes are then tested under different conditions, i.e. control and conditioned. 

As we saw, combinations of these factors can change both the means and variance of distributions of individual behaviours in both genotype- or environment-specific manner. Normally, variance would be a good estimate for expressed individuality within a genotype, and comparison of variances would be a good estimate of change in individuality due to some factor (genotype, replicate, or type of conditioning).

However, we saw that the resulting shape of the data in these experiments, (the shape of the distributions) was incompatible with the classical definition of extent of individuality measured by variance. While it would be more intuitive to track variance over time, we found that this measure obfuscates some obvious changes in the normal shape of the distributions of individual behaviours, as can be visually observed for example between conditioned and control experiments. This is why we moved to develop the measure of residual individuality. Residual individuality aims to measure exactly this dimension of individuality that is missed by measures of variance. We will add a schematic presentation of residual individuality in Figure 2 to explain more explicitly and visually what is being measured here, and what residual individuality represents. This should shed more light on how these analyses support the conclusion that learning increases behavioural variability among individuals in both Figure 2 and Figure 3. The schematic should provide more intuition on how to interpret the data to those unfamiliar with some of the statistics. Besides the schematics, we will also add more intuitive visualizations of the behaviour data in the supplementary, including representations of how within-strain distributions of behaviour change before and during learning or in control condition for all strains, so that the reader may inspect them in more detail.

Improved explanation of Figure 3 and the link between statistical outputs and behavioural measurements.

Reviewer 1 also noted that the analyses in Figure 3 are difficult to interpret without relying heavily on the Results text. Hopefully the added schematic in Figure 2 that explains what Divergence represents should address this note and make the interpretation of Figure 3 easier. Indeed, upon reflection, we agree that the label “Divergence” is quite vague. The “Divergence” in fact shows again residual individuality, and how it changes with every made decision in the case where we compare distributions of flies that start at green versus the blue arm. We further subset the distributions by clustering flies that share the same individual initial color bias or similar learning score and measure residual individuality for them as well. Here, value 0 means the two distributions have the same shape, and higher values mean the shapes are more different. We will rename Divergence to “Residual individuality Start” to make it clear that this is conceptually the same type of measurement, and revise the figure legends accordingly so that they match the new schematic in Figure 2. This should hopefully clarify what the figures show. We will also add a schematic to depict how change in the shape of the distribution with each decision can affect residual individuality.

Reviewer 2:

Clarification of the term “deterministic” when referring to genetic sources of variation.

Reviewer 2 noted that describing genotype as a deterministic source of variation could be confusing, since gene expression and downstream cellular phenotypes are themselves noisy and stochastic. Indeed, gene expression as a phenotype is noisy, but also at the core it is a result of G x E (albeit the environment at the molecular scale). What we meant to emphasize here is that an individual’s genotype can be considered a fixed variable that determines phenotype expression across environments. The environment also determines the phenotype, again, in concert with genotype, but it will always vary over time. We agree with the reviewer that the wording should be made stricter to avoid confusion.

We changed this sentence from “In every individual, behaviour is shaped by deterministic, genetic factors and by environmental events throughout lifetime, which may be stochastic and can occur at the molecular, cellular, organismal and even population scales.” to “In every individual, behaviour is shaped by fixed genetic factors and by variable environmental events throughout lifetime, which may be stochastic and can occur at the molecular, cellular, organismal and even population scales.” 

Longitudinal analysis and neural sources of learning variability.

Reviewer 2 suggested that additional longitudinal analysis could further strengthen the evidence for individuality, and that identifying neural sources of learning variability would be an interesting future direction. We appreciate these suggestions and very much agree with them. But as it was pointed out by the reviewer, this was beyond the scope of this study. Nonetheless, it may be good to note that we have in fact already started this (ongoing and quite extensive) experimental endeavour to identify neural sources of individuality, which we hope will be soon available as a follow-up study.

Within the current study we were able to track behaviour longitudinally within a 20-minute experiment, and in one case over multiple days, though for only a smaller subset of flies. Broader conclusions on how behaviour would change over longer timeframes (except those already included in the manuscript) could not be made with the current dataset. We have added a figure in the supplement where the reader can visually explore the temporal changes to the distributions of behaviour. More extensive study to see how individuality evolves over longer time frames is indeed planned for the future.

We thank the reviewers again for their thoughtful and constructive comments. We believe that addressing these points improved the manuscript.