Author response:
Reviewer #1 (Public review):
Summary:
The paper by Tolossa et al. presents classification studies that aim to predict the anatomical location of a neuron from the statistics of its in-vivo firing pattern. They study two types of statistics (ISI distribution, PSTH) and try to predict the location at different resolutions (region, subregion, cortical layer).
Strengths:
This paper provides a systematic quantification of the single-neuron firing vs location relationship.
The quality of the classification setup seems high.
The paper uncovers that, at the single neuron level, the firing pattern of a neuron carries some information on the neuron's anatomical location, although the predictive accuracy is not high enough to rely on this relationship in most cases.
Thank you for your thoughtful feedback. The level of predictive accuracy offered by our current approach, while far above chance, is insufficient for electrode localization in most cases. Although, we speculate that our results represent a lower limit on possible performance—future improvements are almost certain as larger datasets are generated, more diverse features of neural activity are employed, and more advanced ML tools are implemented. We note that the current performance indicates a far more reliable embedding of anatomy in spiking than precedented by the modest statistical significance previously described in the literature. It would have been impossible to achieve this without the tremendous resources provided by the Allen Institute. In our revision, we will clarify that major performance improvements are both possible and probable.
Weaknesses:
As the authors mention in the Discussion, it is not clear whether the observed differences in firing are epiphenomenal. If the anatomical location information is useful to the neuron, to what extent can this be inferred from the vicinity of the synaptic site, based on the neurotransmitter and neuromodulator identities? Why would the neuron need to dynamically update its prediction of the anatomical location of its pre-synaptic partner based on activity when that location is static, and if that information is genetically encoded in synaptic proteins, etc (e.g., the type of the synaptic site)? Note that the neuron does not need to classify all possible locations to guess the location of its pre-synaptic partner because it may only receive input from a subset of locations. If an argument on activity-based estimation being more advantageous to the neuron than synaptic site-based estimation cannot be made, I believe limiting the scope of the paper (e.g., in the Introduction) to an epiphenomenal observation and its quantification will improve the scientific quality.
Summarily, in response to the two reviewers, we will minimize our discussion of this question in the revision. However, given that our results are either epiphenomenal or functional, we feel that it is important to indicate these possibilities, even if this indication is succinct and conservative.
In pursuit of a more concise revision, we will not expand our discussion to accommodate this interesting conversation with the reviewer, but we are excited to briefly offer our perspective here.
Regarding the epiphenomenal nature of our observations: this is a complex question that would be challenging but not impossible to validate experimentally. It has been previously established that neurons, especially those that integrate inputs from a variety of regions and are involved in diverse functions, could benefit from mechanisms for dynamically parsing inputs (Gutig, Sompolinsky 2006). Neurotransmitter and neuromodulator identities may indeed convey some information about presynaptic neuron location (e.g., NE may originate from the locus coeruleus). However, hypothetically, the binding of a neurotransmitter only bears on the postsynaptic neuron via ionic current, or second messenger activity. Postsynaptic neurons do not consume or otherwise endocytose the neurotransmitter, thus the ability of a neuron to “know” the presynaptic identity is a function of induced postsynaptic activity. Certainly, there are multiple streams of information that can provide insight into anatomical location all taking the ultimate form of neural activity and membrane dynamics. This would be broadly consistent with (for example) reward prediction error which is evident in dopamine release, firing rates, spiking patterns, and oscillatory rhythms.
We could imagine a possible role for the embedding of location in spiking patterns. It is important to note that many neurons in neighboring areas share common neurotransmitters (e.g., glutamate, GABA). Neurons receiving input from multiple regions with similar neurotransmitter profiles could benefit from additional information in the spiking patterns for distinguishing input sources, especially for multimodal integration. For instance, an inferior parietal lobule neuron or microcircuit could be downstream from both auditory cortex (listening) and Broca’s area (speaking). Imagine an individual is in a crowded coffee shop waiting for their drink order to be called while speaking to their friend. In this scenario, it may be important to recognize region-specific activity and thus selectively attend to it. Thus, it is unlikely that neurons actively update a “location prediction,” but rather that location-related information is passively embedded in spike patterning and this might be dynamically leveraged in computation. We emphasize that this is a simplified conceptual example and not a hypothesis that we test in the paper. This conversation, however, is a wonderful example of the thought experiments that we hope will grow from this type of work.
Reviewer #2 (Public review):
Summary:
In this manuscript, Tolossa et al. analyze Inter-spike intervals from various freely available datasets from the Allen Institute and from a dataset from Steinmetz et al. They show that they can modestly decode between gross brain regions (Visual vs. Hippocampus vs. Thalamus), and modestly separate sub-areas within brain regions (DG vs. CA1 or various visual brain areas).
Strengths:
The paper is reasonably well written, and the definitions are quite well done. For example, the authors clearly explained transductive vs. inductive inference in their decoders. E.g., transductive learning allows the decoder to learn features from each animal, whereas inductive inference focuses on withheld animals and prioritizes the learning of generalizable features.
Thank you!
Weaknesses:
However, even with some of these positive aspects, I still found the manuscript to be a laundry list of results, where some results are overly explained and not particularly compelling or interesting, whereas interesting results are not strongly described or emphasized. The overall problem is that the study is not cohesive, and the authors need to either come up with a tool or demonstrate a scientific finding. The current version attempts to split the middle and thus is not as impactful as it could be
In our revision, we will endeavor to present our results in line with your suggestions. Thank you for the careful and thorough feedback that will improve the readability of our manuscript. We strove to be complete in establishing the logic leading to our ultimate finding—that a robust code for anatomical location can be extracted from single neuron spike trains, but not from more traditional descriptions of neural activity. Our detection of this code, albeit not perfect in performance, is, in most cases, both far above chance levels and is robust to animal identity and laboratory of origin. Our presentation of these results is cohesive in as much as we sequentially establish a series of results that build towards a concluding set of experiments. We start by establishing a baseline via standard measurements and then explore more challenging problems through more complex models that build toward our final test. Based on your feedback, we will contract and expand elements of this sequence.
While our findings raise the possibility of developing a computational tool for electrode localization, pending additional features and/or datasets, our current focus is on establishing the neurobiological principle of anatomical embedding in spike trains. The purpose of briefly mentioning a possible application is that we hope to encourage those engaged in machine-learning on multi-modal neural data that this problem is tractable, yet still open. Based on your feedback, we will clarify that the focus of our current work is not an introduction of a new tool.
