Whole-brain neural substrates of behavioral variability in the larval zebrafish

Reviewer #1 (Public Review):

Summary:

In this paper, Manley and Vaziri investigate whole-brain neural activity underlying behavioural variability in zebrafish larvae. They combine whole brain (single cell level) calcium imaging during the presentation of visual stimuli, triggering either approach or avoidance, and carry out whole brain population analyses to identify whole brain population patterns responsible for behavioural variability. They show that similar visual inputs can trigger large variability in behavioural responses. Though visual neurons are also variable across trials, they demonstrate that this neural variability does not degrade population stimulus decodability. Instead, they find that the neural variability across trials is in orthogonal population dimensions to stimulus encoding and is correlated with motor output (e.g. tail vigor). They then show that behavioural variability across trials is largely captured by a brain-wide population state prior to the trial beginning, which biases choice - especially on ambiguous stimulus trials. This study suggests that parts of stimulus-driven behaviour can be captured by brain-wide population states that bias choice, independently of stimulus encoding.

Strengths:

-The strength of the paper principally resides in the whole brain cellular level imaging in a well-known but variable behaviour.

- The analyses are reasonable and largely answer the questions the authors ask.

- Overall the conclusions are well warranted.

Weaknesses:

A more in-depth exploration of some of the findings could be provided, such as:

- Given that thousands of neurons are recorded across the brain a more detailed parcelation of where the neurons contribute to different population coding dimensions would be useful to better understand the circuits involved in different computations.

- Given that the behaviour on average can be predicted by stimulus type, how does the stimulus override the brain-wide choice bias on some trials? In other words, a better link between the findings in Figures 2 and 3 would be useful for better understanding how the behaviour ultimately arises.

- What other motor outputs do the noise dimensions correlate with?

The dataset that the authors have collected is immensely valuable to the field, and the initial insights they have drawn are interesting and provide a good starting ground for a more expanded understanding of why a particular action is determined outside of the parameters experimenters set for their subjects.

Reviewer #2 (Public Review):

Overview

In this work, Manley and Vaziri investigate the neural basis for variability in the way an animal responds to visual stimuli evoking prey-capture or predator-avoidance decisions. This is an interesting problem and the authors have generated a potentially rich and relevant data set. To do so, the authors deployed Fourier light field microscopy (Flfm) of larval zebrafish, improving upon prior designs and image processing schemes to enable volumetric imaging of calcium signals in the brain at up to 10 Hz. They then examined associations between neural activity and tail movement to identify populations primarily related to the visual stimulus, responsiveness, or turn direction - moreover, they found that the activity of the latter two populations appears to predict upcoming responsiveness or turn direction even before the stimulus is presented. While these findings may be valuable for future more mechanistic studies, issues with resolution, rigor of analysis, clarity of presentation, and depth of connection to the prior literature significantly dampen enthusiasm.

Imaging

- Resolution: It is difficult to tell from the displayed images how good the imaging resolution is in the brain. Given scattering and lensing, it is important for data interpretation to have an understanding of how much PSF degrades with depth.

- Depth: In the methods it is indicated that the imaging depth was 280 microns, but from the images of Figure 1 it appears data was collected only up to 150 microns. This suggests regions like the hypothalamus, which may be important for controlling variation in internal states relevant to the behaviors being studied, were not included.

- Flfm data processing: It is important for data interpretation that the authors are clearer about how the raw images were processed. The de-noising process specifically needs to be explained in greater detail. What are the characteristics of the noise being removed? How is time-varying signal being distinguished from noise? Please provide a supplemental with images and algorithm specifics for each key step.

- Merging: It is noted that nearby pixels with a correlation greater than 0.7 were merged. Why was this done? Is this largely due to cross-contamination due to a drop in resolution? How common was this occurrence? What was the distribution of pixel volumes after aggregation? Should we interpret this to mean that a 'neuron' in this data set is really a small cluster of 10-20 neurons? This of course has great bearing on how we think about variability in the response shown later.

- Bleaching: Please give the time constants used in the fit for assessing bleaching.

Analysis

- Slow calcium dynamics: It does not appear that the authors properly account for the slow dynamics of calcium-sensing in their analysis. Nuclear-localized GCaMP6s will likely have a kernel with a multiple-second decay time constant for many of the cells being studied. The value used needs to be given and the authors should account for variability in this kernel time across cell types. Moreover, by not deconvolving their signals, the authors allow for contamination of their signal at any given time with a signal from multiple seconds prior. For example, in Figure 4A (left turns), it appears that much of the activity in the first half of the time-warped stimulus window began before stimulus presentation - without properly accounting for the kernel, we don't know if the stimulus-associated activity reported is really stimulus-associated firing or a mix of stimulus and pre-stimulus firing. This also suggests that in some cases the signals from the prior trial may contaminate the current trial.

- Partial Least Squares (PLS) regression: The steps taken to identify stimulus coding and noise dimensions are not sufficiently clear. Please provide a mathematical description.

- No response: It is not clear from the methods description if cases where the animal has no tail response are being lumped with cases where the animal decides to swim forward and thus has a large absolute but small mean tail curvature. These should be treated separately.

Results

- Behavioral variability: Related to Figure 2, within- and across-subject variability are confounded. Please disambiguate. It may also be informative on a per-fish basis to examine associations between reaction time and body movement.

- Data presentation clarity: All figure panels need scale bars - for example, in Figure 3A there is no indication of timescale (or time of stimulus presentation). Figure 3I should also show the time series of the w_opt projection.

- Pixel locations: Given the poor quality of the brain images, it is difficult to tell the location of highlighted pixels relative to brain anatomy. In addition, given that the midbrain consists of much more than the tectum, it is not appropriate to put all highlighted pixels from the midbrain under the category of tectum. To aid in data interpretation and better connect this work with the literature, it is recommended that the authors register their data sets to standard brain atlases and determine if there is any clustering of relevant pixels in regions previously associated with prey-capture or predator-avoidance behavior.

Interpretation

- W_opt and e_1 orthogonality: The statement that these two vectors, determined from analysis of the fluorescence data, are orthogonal, actually brings into question the idea that true signal and leading noise vectors in firing-rate state-space are orthogonal. First, the current analysis is confounding signals across different time periods - one could assume linearity all the way through the transformations, but this would only work if earlier sources of activation were being accounted for. Second, the transformation between firing rate and fluorescence is most likely not linear for GCaMP6s in most of the cells recorded. Thus, one would expect a change in the relationship between these vectors as one maps from fluorescence to firing rate.

- Sources of variability: The authors do not take into account a fairly obvious source of variability in trial-to-trial response - eye position. We know that prey capture responsiveness is dependent on eye position during stimulus (see Figure 4 of PMID: 22203793). We also expect that neurons fairly early in the visual pathway with relatively narrow receptive fields will show variable responses to visual stimuli as the degree of overlap with the receptive field varies with eye movement. There can also be small eye-tracking movements ahead of the decision to engage in prey capture (Figure 1D, PMID: 31591961) that can serve as a drive to initiate movements in a particular direction. Given these possibilities indicating that the behavioral measure of interest is gaze, and the fact that eye movements were apparently monitored, it is surprising that the authors did not include eye movements in the analysis and interpretation of their data.

Reviewer #3 (Public Review):

Summary:

In this study, Manley and Vaziri designed and built a Fourier light-field microscope (fLFM) inspired by previous implementations but improved and exclusively from commercially available components so others can more easily reproduce the design. They combined this with the design of novel algorithms to efficiently extract whole-brain activity from larval zebrafish brains.

This new microscope was applied to the question of the origin of behavioral variability. In an assay in which larval zebrafish are exposed to visual dots of various sizes, the fish respond by turning left or right or not responding at all. Neural activity was decomposed into an activity that encodes the stimulus reliably across trials, a 'noise' mode that varies across trials, and a mode that predicts tail movements. A series of analyses showed that trial-to-trial variability was largely orthogonal to activity patterns that encoded the stimulus and that these noise modes were related to the larvae's behavior.

To identify the origins of behavioral variability, classifiers were fit to the neural data to predict whether the larvae turned left or right or did not respond. A set of neurons that were highly distributed across the brain could be used to classify and predict behavior. These neurons could also predict spontaneous behavior that was not induced by stimuli above chance levels. The work concludes with findings on the distributed nature of single-trial decision-making and behavioral variability.

Strengths:

The design of the new fLFM microscope is a significant advance in light-field and computational microscopy, and the open-source design and software are promising to bring this technology into the hands of many neuroscientists.

The study addresses a series of important questions in systems neuroscience related to sensory coding, trial-to-trial variability in sensory responses, and trial-to-trial variability in behavior. The study combines microscopy, behavior, dynamics, and analysis and produces a well-integrated analysis of brain dynamics for visual processing and behavior. The analyses are generally thoughtful and of high quality. This study also produces many follow-up questions and opportunities, such as using the methods to look at individual brain regions more carefully, applying multiple stimuli, investigating finer tail movements and how these are encoded in the brain, and the connectivity that gives rise to the observed activity. Answering questions about variability in neural activity in the entire brain and its relationship to behavior is important to neuroscience and this study has done that to an interesting and rigorous degree.

Points of improvement and weaknesses:

The results on noise modes may be a bit less surprising than they are portrayed. The orthogonality between neural activity patterns encoding the sensory stimulus and the noise modes should be interpreted within the confounds of orthogonality in high-dimensional spaces. In higher dimensional spaces, it becomes more likely that two random vectors are almost orthogonal. Since the neural activity measurements performed in this study are quite high dimensional, a more explicit discussion is warranted about the small chance that the modes are not almost orthogonal.

The conclusion that sparsely distributed sets of neurons produce behavioral variability needs more investigation because the way the results are shown could lead to some misinterpretations. The prediction of behavior from classifiers applied to neural activity is interesting, but the results are insufficiently presented for two reasons.

(1) The neurons that contribute to the classifiers (Figures 4H and J) form a sufficient set of neurons that predict behavior, but this does not mean that neurons outside of that set cannot be used to predict behavior. Lasso regularization was used to create the classifiers and this induces sparsity. This means that if many neurons predict behavior but they do so similarly, the classifier may select only a few of them. This is not a problem in itself but it means that the distributions of neurons across the brain (Figures 4H and J) may appear sparser and more distributed than the full set of neurons that contribute to producing the behavior. This ought to be discussed better to avoid misinterpretation of the brain distribution results, and an alternative analysis that avoids the confound could help clarify.

(2) The distribution of neurons is shown in an overly coarse manner in only a flattened brain seen from the top, and the brain is divided into four coarse regions (telencephalon, tectum, cerebellum, hindbrain). This makes it difficult to assess where the neurons are and whether those four coarse divisions are representative or whether the neurons are in other non-labeled deeper regions. For these two reasons, some of the statements about the distribution of neurons across the brain would benefit from a more thorough investigation.