Fast burst fraction transients convey information independent of the firing rate

Reviewer #1 (Public Review):

Summary:

In this paper, the authors' study aimed to test existing theories on the role of bursting in learning and attention. They find evidence for both. It is not clear how these two can be reconciled, but this is one of the first studies to explicitly test recent theories of spike multiplexing in the brain. This will pave the way for future investigations, both experimental and theoretical.

Strengths:

(1) A key strength of this study is the fact that it aims to test existing theories of spike multiplexing, finding support for both attention-like and learning-like signals.

(2) The task setup is of particular interest to brain-machine interfaces, and how such setups trigger learning and attention mechanisms.

Weaknesses:

(1) The fact that the teaching signal is an (artificial) stimulation of the primary sensory cortex, makes it unclear how applicable are these results to a more general understanding of learning and attention in the brain.

(2) It would have been useful to more directly compare the results obtained here with existing burst-dependent computational models of learning and attention. This is particularly important since there appears to be an interaction between learning and sharpening signals.

(3) There are inherent limitations in our current ability to read out bursting and non-bursting signals, this is a brave first attempt, but at this point, it is unclear how can one robustly read out this information from noisy data.

Reviewer #2 (Public Review):

Naud et al investigate whether single spikes and bursts encode different information in behavior. To do this, they reanalyze juxtasomal recordings of deep-layer cortical neurons from behaving rats collected in two previous studies by Doron et al. Rats were trained (in a Go-NoGo design) to lick a spout for a water reward in response to electrical microstimulation of the primary somatosensory cortex, which rats quickly learn to do in a single day. Juxtasomal recordings near the site of micro stimuli are then divided up into single spikes ("events") versus high-frequency bursts ("bursts"). Training results in the appearance of bursts, which do not seem to correlate with the rate of events, suggesting that bursts and events carry different information. While the fraction of bursts is elevated during Hit trials, errors appear to uniquely trigger additional bursts. The distribution of burst times appears to shift from long after the stimulus (early in training) to shortly after the stimulus (later in training). Bursts of layer 5 pyramidal neurons in particular are associated with apical tuft activity that could enhance plasticity. The observed increased bursting is therefore suggestive of a potential mechanism by which errors engage plasticity.

This paper has substantial strengths: the experiments appear to be well performed, the dataset is substantial, and the questions and phenomena are interesting.

The exclusion of fast-spike (inhibitory) data, which the experiments seem to have generated, is a weakness as these data could have provided an important control. If the bursts here reflect apical dendrite activity, the same phenomena might be absent in inhibitory cells as they lack apical tufts.

Another weakness is the need to better control movement, which could be an alternative explanation to the top-down modulation of apicals that the authors suspect. For example, the bursts on error trials could be due to the animals moving more when an error occurs. Layer 5 of the somatosensory cortex has increased activity during whisking or body movements. If the mouse fidgets out of frustration that the reward has not occurred or whisks more, bursts are highly likely due to less exotic purely bottom-up inputs.

Reviewer #3 (Public Review):

Summary:

The burst fraction neural code has conceptual interest but has been little examined in vivo. This study examines and compares the burst fraction, the standard firing rate (firing rate) code, and the related event fraction (event rate) code using published data from an experiment where rats learned to lick after detecting electrical microstimulation in the somatosensory (barrel) cortex. Analyzing single-neuron spiking responses, the study reports that the burst fraction identifies more and different neurons showing the effects of training than the firing rate. The study further claims that the burst fraction (1) most promptly responded to false-negative detection errors, (2) during further training of trained animals (from 80% to 90% accuracy, over five days), correlates with behavioral accuracy, and (3) by shifting earlier to align with the (relatively constant) event rate modulation, leads to the observed sharpened firing rate response during this further training. The study concludes that 'a fine-grained separation of spike timing patterns [into burst fraction, firing rate, and event rate] reveals two signals,' an error signal and a sharpening signal.

Strengths:

The burst fraction is shown to discern more (and somewhat different) cells showing significant responses in trained animals and also to reveal a larger absolute difference in the fraction of responsive cells between naïve and trained animals. The Poisson model analysis particularly convincingly shows that the firing rate alone cannot explain either the spiking pattern or the prevalence of burst fraction-ON cells, thereby furnishing strong evidence that the burst fraction conveys independent information from the firing rate. The demonstration of error signals on miss trials in all three neural codes (burst fraction, firing rate, event rate) is interesting. It is also interesting to see that neural responses broadly shift earlier for animals even during further training in an already 'expert' stage and that the burst fraction correlates with further accuracy increases.

Weaknesses:

The evidence is inadequate for the burst fraction as responding more promptly to missed trials.

This key claim seems to rest solely on the timing of the first bins in Figure 3B showing statistically significant differences. This reasoning implicitly draws inferences from the lack of statistical differences, which cannot support a positive claim in general. Specifically, here, the burst fraction is calculated with a division, which can magnify small differences and impact the power of statistical tests. If I trace back from the first bin showing significant differences to the first bin the signal starts rising, the timing seems to be comparable for all three neural codes (~1.6 s).

Pertinently, what is the statistical test used in Figure 3B? A parametric test may be inappropriate for the burst fraction, a ratio that like does not fulfill the normality assumption. An inappropriate test would compound the problem of concluding from the lack of (early) significant differences.

The evidence that burst fraction is responsible for sharpening is opaque due to insufficient statistical reporting. Specifically, it seems there is a correlation between firing rate and accuracy that is reported as non-significant.

Changes in the reaction times (or other movement parameters) over-training may confound the correlation of the burst fraction to the accuracy and firing rate sharpening during further training. Lack of control for changes in movement over training weakens the results.

The claim of independence of burst fraction and event rate/firing rate information is too strong. The authors show a significant negative correlation between burst fraction and firing rate (2D).

The claim that there is no 'functional reorganization' beyond day two is too strong. Although this claim is not a core one to the study, it derives from an absence of statistical significance, especially problematic here as the effect sizes are large. For example, the Spearman correlation is 0.67/0.87 for the analyses with burst fraction. With only five data points, even strong effects may not achieve statistical significance, making negative conclusions problematic. Further, how are the p-values calculated (if using a parametric test, are the assumptions met), and why should these analyses use Spearman's correlation when analogous analyses in Figure 4E, F use Pearson's r?

Does the burst fraction correlate with accuracy in cross-training?

If the burst fraction correlates with accuracy, it should be expected to do so also when the animals progress from the naïve to the trained stage. Moreover, the correlation in Figure 4E can benefit from strengthening as it is now supported by only five points, is driven by only three 'clusters,' and only represents a narrow range of accuracies. If the data is available for this analysis, it should be done to test and potentially strengthen the main claim of the study.

The text and figures contain numerous ambiguities that need to be clarified. These do not include obvious typos, only elements that affect conceptual understanding.

- Some key terms in the main claims are never defined. For example, in the title, it is unclear what 'fast' and 'transients' mean. The abstract uses, but the main text never defines, 'demultiplexing,' 'a *conjunctive* burst code,' 'sparse and succinct [sic],' and 'correlated more *globally*.'

- Some paper components are un(der)explained and, sometimes, apparently internally inconsistent. For example, in Figure 1I, the fraction of firing rate-ON cells does not look like the 6% shown in Figure 1J, left. In Figure 2E-G, what is the total cell number, 279, in Figure 2G legend, why is it different from the 153 total cells in Figure 2E legend, and what is the 'n = 5' within Figure 2G? All n numbers should be explained in general; more examples include the 245 in Figure 3C and the 49 in Figure 3B. In Figure 3C, what is the top horizontal bar (I assume significant differences)? About catch trials, the Figure 3D legend says rewards are given on licks, but the text says licking was not rewarded; which is the case? Figure 4B legend says 'firing rate (left), burst fraction (middle) and event rate (right),' but the plot colors imply a different order.

- The abstract states, 'The alignment of bursting and event rate modulation [...] was strongly associated [sic] behavioral accuracy.' It seems to me it is not the alignment of burst fraction and event rate but rather burst fraction per se that correlates with behavioral accuracy (Figure 4E right). At least, the latter correlation is the only one tested.