A Neuron Is Not a Signal: It Is a Conversation Between Three Clocks

What if the most important fact about a neuron is not what it is, but when you ask?

A neuron can fire within milliseconds, alter its ion channels over minutes, change its gene expression over hours, and shift its role within a circuit over days. Each of these descriptions can be accurate at the same time. Yet neuroscience has often treated them as separate objects: electrical recordings for activity, transcriptomic profiles for identity, and spatial maps for context.

The deeper problem is not a shortage of data. It is that we have been collecting descriptions of the same cell in different languages and then mistaking each language for the cell itself.

The emerging combination of deep learning, single cell transcriptomics, spatial transcriptomics, and single neuron physiology suggests a more useful view: neuronal activation is not a single observable property. It is an inference about a changing biological state, reconstructed from measurements that operate on different clocks.

That shift has consequences far beyond better classification. It changes what counts as evidence, how experiments should be designed, and how we interpret the relationship between a neuron’s identity and its behavior.

The same neuron tells different stories at different speeds

Imagine trying to understand a city using three instruments. One records traffic light changes every millisecond. Another surveys the kinds of businesses operating in each neighborhood. A third maps where people live and how neighborhoods connect. None of these instruments is the city. Yet together they reveal something no single measurement can provide.

Neurons pose a similar problem. Physiology captures what a cell is doing now. Transcriptomics captures molecular programs that support, reflect, or remember what it has been doing. Spatial information captures the neighborhood in which its behavior has meaning.

Electrical activity is fast and direct. A spike, membrane potential change, or synaptic response tells us about the cell’s immediate functional state. But such measurements are often difficult to scale across many cells, especially when the goal is to understand an entire tissue or brain region.

Transcriptomic measurements offer a different kind of reach. They can reveal gene expression patterns across thousands of cells and distinguish populations that appear similar under a microscope. But a transcriptomic profile is not a live video of neural activity. It is closer to a biochemical report, shaped by recent history, cell type, stress, developmental state, and the technical process used to measure it.

This creates a fundamental interpretive tension. A gene expression signature associated with activation may indicate that a neuron recently responded to a stimulus. It may also reflect its baseline identity, its position in a circuit, or a broader cellular response to environmental change. The signal is meaningful, but it is not self interpreting.

A transcriptome is not a photograph of neuronal activity. It is a molecular trace whose meaning depends on time, context, and comparison.

Deep learning becomes valuable precisely at this boundary. Its role is not simply to replace a missing measurement. It is to learn relationships between patterns that are too complex, distributed, or nonlinear for simple rules. A model can be trained on neurons for which molecular profiles and known activity states are available, then use those learned relationships to estimate activation in cells where direct physiological recording is absent.

But this power also exposes a danger. If a model learns that a particular cell type usually occurs in an active region, it may infer activation from identity or location rather than from a true activity related molecular pattern. The prediction can be statistically accurate while biologically misleading.

The central question therefore becomes: is the model detecting activity, or is it detecting the neighborhood in which activity tends to occur?

The hidden variable problem

The most useful way to think about neuronal activation is as a hidden variable. We cannot always observe it directly at the scale we want. Instead, we observe consequences: electrical events, gene expression, protein changes, anatomical location, and connectivity.

This resembles weather forecasting. Temperature, humidity, pressure, wind, and cloud cover are distinct measurements. A good forecaster does not confuse any one of them with weather itself. Instead, the forecast estimates a latent atmospheric state from their joint pattern.

Neural data require the same discipline. Let a neuron’s underlying functional state be represented by a hidden variable, which we can call its activation state. Physiological measurements provide fast evidence about that state. Transcriptomic measurements provide slower, molecularly filtered evidence. Spatial data provide priors about what kinds of states are plausible in a given location.

The three forms of evidence should not be treated as interchangeable. They should be treated as complementary constraints.

A practical framework is to ask three questions for every prediction:

1. What does this measurement respond to quickly?
2. What does it preserve from the cell’s past?
3. What alternative explanation could produce the same pattern?

For physiology, the first answer is strong: electrical activity changes quickly. The second answer is weaker, because many transient events leave little lasting molecular trace. For transcriptomics, the first answer is slower and more selective, while the second is stronger: gene expression can retain evidence of recent or sustained cellular programs. Spatial context is different again. It does not necessarily record what the cell did, but it may explain why the cell is likely to respond in a certain way.

This framework prevents a common category error: treating a measurement of cellular readiness as a measurement of cellular action. A neuron can express the machinery associated with a response without currently firing. Conversely, it can fire in a way that does not produce a distinctive transcriptomic signature under the sampling conditions.

The distinction is especially important when comparing cell types. Suppose two neuronal populations show different expression of activity related genes after the same stimulus. The difference could mean that one population was more active. It could also mean that the two populations translate electrical activity into transcription differently. In that case, identical physiological events might produce different molecular readouts.

This is why integrating physiology and transcriptomics at the level of the same neuron is so powerful. It allows researchers to ask not merely whether two measurements correlate across populations, but whether a particular molecular profile corresponds to a particular physiological phenotype within individual cells.

The unit of explanation changes from the average cell type to the cell that has both a molecular history and a measured behavior.

Deep learning should be a translator, not an oracle

Machine learning is often introduced as a way to extract patterns from high dimensional data. That description is accurate but incomplete. In this setting, deep learning functions most usefully as a translator between biological measurement systems.

A transcriptomic profile may contain thousands of measured features. Many are weakly informative on their own, but collectively they can encode a state. A deep model can learn combinations of genes, interactions among cell programs, and nonlinear relationships between molecular states and observed activation. It can also incorporate spatial information, allowing the prediction to account for the arrangement of neighboring cells rather than treating every cell as an isolated point.

The analogy is language translation. A word rarely has a single meaning outside a sentence. Its interpretation depends on neighboring words, grammar, and context. Similarly, a gene rarely serves as a universal activation marker. Its meaning depends on the cell type, the tissue state, the stimulus, and the broader pattern of expression.

A model that treats individual genes as independent switches will miss this compositional structure. A model that learns the full context may do better, but it must still be tested against confounding.

There are at least four ways a prediction can appear biologically impressive while failing to measure the intended phenomenon:

• Identity leakage: the model uses genes that distinguish cell types, even when those genes do not indicate activation.
• Spatial leakage: the model uses location as a shortcut because active cells are concentrated in known regions.
• Batch leakage: the model learns technical differences between experiments instead of biological differences.
• Temporal ambiguity: the model predicts a molecular aftermath, then reports it as if it were a direct record of instantaneous activity.

These are not merely statistical nuisances. They are failures of biological interpretation.

The remedy is not to reject complex models. It is to give them more demanding tests. A model designed to infer activation should be evaluated on cells from new animals, new experiments, new regions, and, where possible, new perturbations. It should be asked to predict physiology in settings where cell identity and location no longer provide easy shortcuts.

Interpretability also needs a biological definition. A list of genes with high model importance is not automatically an explanation. A stronger explanation identifies whether the model’s prediction is driven by a coherent molecular program, whether that program generalizes across cell classes, and whether perturbing the relevant biological process changes the predicted state.

The goal is not to make the model simple. The goal is to make its complexity accountable.

The overlooked value of disagreement

When physiology and transcriptomics disagree, the first instinct is often to decide which measurement is wrong. That may be a mistake. Their disagreement can reveal the temporal structure of neural computation.

Consider a neuron that shows strong electrical responses but weak expression of canonical activity related genes. Several interpretations are possible. The activity may have been too brief to trigger transcription. The neuron may use a different molecular coupling mechanism. The sampling interval may have missed the relevant gene response. Or the cell may be electrically active but transcriptionally buffered.

Now consider the opposite case: a neuron displays a strong molecular activation signature but little recorded firing. Perhaps the recording occurred after the key event. Perhaps the molecular program reflects prolonged activity that has already subsided. Perhaps neighboring signals or neuromodulators altered transcription without producing large spikes in the measured cell.

In both cases, disagreement is not simply noise. It is a clue about latency, persistence, and transformation.

A useful conceptual model is to treat each measurement as a filter applied to the hidden neural state. Physiology is a high speed filter with fine temporal resolution. Transcriptomics is a slower filter that integrates signals over time and transforms them through gene regulatory networks. Spatial context is a contextual filter that changes the prior probability of different states.

If two filters disagree, researchers can ask what kind of signal would pass through one but not the other. This turns multimodal integration into a method for discovering biological dynamics.

For example, repeated brief bursts of activity may produce a different transcriptomic outcome from one sustained episode, even if the total number of spikes is similar. Physiology can distinguish the patterns immediately. Transcriptomics may reveal that the cell’s molecular response depends not only on how much activity occurred, but on its temporal arrangement.

This is a broader lesson for biology: measurement disagreement can contain information about transformation between levels of organization. The gap between an electrical event and a gene expression program is not an inconvenience to be eliminated. It is the mechanism by which the cell converts experience into a durable state.

From better maps to better experiments

The practical payoff of this synthesis is not just a more detailed atlas. It is a different experimental strategy.

First, experiments should be designed around the question’s time scale. If the goal is to identify immediate responsiveness, physiological measurements are indispensable. If the goal is to understand how repeated experience changes cellular state, transcriptomics becomes more informative. If the goal is to explain why neighboring cells respond differently, spatial context must be included from the beginning.

Second, researchers should collect paired measurements whenever possible. A large dataset of transcriptomes and a separate dataset of physiological recordings can reveal population level relationships, but paired measurements expose the variability that averages conceal. The same molecular class may contain cells with different response dynamics, and the same physiological response may arise from different molecular routes.

Third, validation should test transfer, not just fit. A model that performs well on held out cells from the same experiment may still fail when the tissue, stimulus, or preparation changes. Strong validation asks whether the learned relationship survives changes in biological context.

Fourth, predictions should be reported with uncertainty and alternatives. A probability of activation is not the same as a declaration of activation. The most informative output may include the leading prediction, its confidence, and the main competing explanations, such as cell identity or spatial location.

Finally, integrated datasets should be used to generate perturbation experiments. If a model identifies a molecular program associated with a physiological response, the next step is not merely to celebrate the association. It is to alter that program and test whether the predicted response changes. Correlation maps the bridge. Perturbation tests whether the bridge carries weight.

Key Takeaways

• Treat activation as a hidden state, not a single marker. Infer it from complementary evidence, while preserving the distinction between immediate activity and molecular aftermath.
• Match measurement to time scale. Use physiology for rapid dynamics, transcriptomics for persistent cellular programs, and spatial data for contextual interpretation.
• Test for shortcuts. Ask whether predictions depend on cell identity, location, experimental batch, or sampling time rather than on activation itself.
• Use disagreement as data. Divergence between molecular and electrical readouts can reveal latency, persistence, and cell specific transformations.
• Validate across contexts and perturb the mechanism. A robust model should transfer to new experiments and produce hypotheses that can be experimentally tested.

The most ambitious promise of this field is often described as the ability to infer neuronal activity from molecular data at large scale. That is important, but it is not the deepest transformation.

The deeper transformation is conceptual. A neuron is not adequately represented by a firing trace, a gene expression vector, or a point on a spatial map. It is a process that moves between these forms. Its electrical activity can become a transcriptional program. Its transcriptional program can alter future excitability. Its location can shape which inputs it receives and which molecular responses are possible.

To understand the cell, we must therefore follow the conversions, not merely collect the endpoints.

The future of single neuron analysis will belong to the questions that ask not only what a cell is doing, but how an event becomes a lasting cellular identity.

Once we adopt that view, multimodal neuroscience stops looking like the accumulation of incompatible datasets. It becomes an attempt to reconstruct a living conversation between three clocks: the speed of electricity, the persistence of molecules, and the organizing force of context. The neuron is not any one of these clocks. It is the changing relationship among them.