Tools that infer cell-to-cell communication from gene expression treat each connection as a fixed number. They score how strongly a "sender" cell type expresses a signaling molecule and how strongly a "receiver" type expresses its sensor, and that score never changes. In the brain this misses something basic. When a neuron fires several times in quick succession, many of its connections weaken within milliseconds because the presynaptic terminal runs out of ready-to-release vesicles. This is called short-term synaptic depression. Other connections instead strengthen with repeated firing, which is called facilitation. Which behavior a connection shows depends on both the sending and the receiving cell. As a result, a communication map that is accurate at low firing rates can be wrong by a factor of two during a burst.
We propose to learn a mapping from the gene expression of a presynaptic cell type and a postsynaptic cell type to the parameters of a standard short-term plasticity model. We then use those predicted parameters to make transcriptome-based neuron-to-neuron communication scores depend on firing frequency. The test that matters is whether the predictions hold for connection types that were never recorded electrophysiologically. If they do, then single-cell and spatial atlases, which cover far more cell-type pairs than paired recordings ever will, could say not only who talks to whom but at which firing rates that conversation survives.
Core Claim
Short-term plasticity is a property of a cell pair, not of a sender alone. The supplied material makes this concrete. In mouse motor cortex, outputs from VIP interneurons (a class of inhibitory cell) are uniformly depressing, with a paired-pulse ratio near 0.5 at 30 Hz. The paired-pulse ratio is the size of the second response divided by the first. Excitatory inputs onto VIP cells, by contrast, span ratios of about 0.5 to 2.5, and the target cell sets the rule (The short-term plasticity of VIP interneurons in motor cortex). That paper explicitly names missing transcriptomic data as the barrier to explaining this variation, and it fits a three-parameter Tsodyks-Markram model. This is a phenomenological model of vesicle depletion and recovery with an initial release probability $U$, a recovery time constant $\tau_D$, and a facilitation time constant $\tau_F$. It reports that these three parameters capture slice data without overfitting.
Our claim is that a low-capacity model which pairs presynaptic gene modules with postsynaptic gene modules can predict these three parameters for held-out connection types better than nearest-neighbor lookup from recorded types. We expect it to do so in the way that experiment predicts: postsynaptic genes should matter for excitatory-onto-VIP connections and barely matter for VIP outputs. The predicted parameters then turn a static communication score into a frequency-dependent one. For the pure-depression case, steady-state efficacy during regular firing at interspike interval $\Delta$ is
$$ E(\Delta) = U \cdot \frac{1 - e^{-\Delta/\tau_D}}{1 - (1-U)\,e^{-\Delta/\tau_D}}, $$
Here $E$ is the per-spike response relative to a fully recovered synapse. The expression follows from equating vesicle depletion per spike, a fraction $U$ of what remains, with exponential recovery between spikes. The facilitating extension adds $\tau_F$ and is used in practice.
Relation to Prior Work
The closest existing tools motivate the gap but do not close it. CellChat scores ligand-receptor pairs with a mass-action model over curated multi-subunit complexes. COMMOT adds spatial distance costs and competition among ligands that share receptors, using optimal transport. NeuronChat handles neurotransmitters that are not peptides with AND/OR rules over synthesis enzymes and vesicular transporters. Of these, NeuronChat is our direct substrate, and it reaches an AUROC of 0.832 against monosynaptic tracing in visual cortex. All three output one time-invariant strength per edge. We keep NeuronChat's static score as the "is there a connection" term and multiply it by a learned, pair-specific dynamic term. We do not claim to improve connectivity detection.
The supervision and the biological constraints come from electrophysiology. The VIP study above supplies the external test set and the pre-versus-post asymmetry we must reproduce. The Allen Synaptic Physiology Dataset, as used in Interactions between long- and short-term synaptic plasticity transform temporal neural representations into spatial, supplies paired recordings across cortical connection types. That paper validated simulated plasticity distributions against those recordings but never tied them to gene expression. We use the same data for the opposite direction, going from genes to parameters.
Two Patch-seq studies constrain the design. Patch-seq records electrophysiology and sequences the same cell. Diversity of Interneurons in the Dorsal Striatum Revealed by Single-Cell RNA Sequencing and PatchSeq shows that Pvalb forms a continuous gradient within Pthlh cells that tracks an electrophysiological continuum. This argues for predicting from continuous expression rather than from cluster labels. Integrating EM and Patch-seq data shows strong target specificity among Sst subtypes. For example, MET-8 places 62.3% of its synapses on L4 IT cells. This tells us that a Euclidean-distance prior like COMMOT's is the wrong way to decide which pairs to score. Neither study measured synaptic short-term dynamics, which is the quantity we target.
Research Program
The first step is building the training table. We map each recorded connection type in the Synaptic Physiology Dataset, defined by presynaptic and postsynaptic class and layer, to transcriptomic types in a matched mouse cortical single-cell reference. Each pair receives a type-level expression profile on each side. Where Patch-seq positions are available, it also receives a within-type position along continuous latent axes. Targets are Tsodyks-Markram fits ($U$, $\tau_D$, $\tau_F$), refit uniformly with the same Bayesian procedure used in the VIP study, so that heterogeneous fitting does not contaminate the labels.
We expect only tens of distinct connection types. That number dictates the model class. We therefore restrict genes to synaptic gene families: vesicle fusion and priming machinery, calcium sensors, presynaptic autoreceptors and neuromodulator receptors, and postsynaptic receptor subunits. Expression is summarized into a small number of modules per side. The predictor is a regularized bilinear form, in which each parameter is predicted from pre-module, post-module, and pre-times-post terms. This is the ligand-receptor pairing idea applied to kinetics rather than to presence. Hierarchical shrinkage pools across layers. Every evaluation is leave-one-connection-type-out, and a stricter leave-one-presynaptic-class-out variant is reported as well.
Integration into communication maps is a secondary deliverable. We multiply NeuronChat edge scores by $E(\Delta)$ at chosen frequencies, apply the result to MERFISH cortex sections that NeuronChat already analyzed, and impute synaptic genes missing from the imaging panel from the single-cell reference. Imputation is a known source of artifacts, so this step is presented as descriptive. Electrophysiological checks gate it.
Baselines
The naive reference is the global mean of each parameter across all training connection types. Every prediction is scored against it. A model that cannot beat the mean on held-out types has learned nothing transferable.
The strongest competitor is nearest-neighbor lookup in transcriptome space. A held-out pair receives the measured parameters of the recorded pair whose concatenated pre and post expression profile is most similar. This is the sensible form of "assign literature constants per cell class," extended to pairs no one recorded. It is strong because neighboring types often share synaptic phenotypes, and it needs no fitting. It sees the same reference expression and the same training connections, and its neighborhood size and distance metric get the same cross-validated tuning budget as our model.
The second baseline is a uniform-depression NeuronChat. Every edge gets the same fitted $U$ and $\tau_D$. This tests whether any pair-specific signal is needed for frequency-dependent maps.
The key ablation is a presynaptic-only version of our model, with the post and interaction terms removed, at matched parameter count. A second ablation replaces continuous expression with one-hot type labels, which tests the continuum argument.
On the core question, no reported result found on Lacuna for predicting short-term plasticity parameters from transcriptomes, so there is no published number to beat. The closest anchors are the measured ratios themselves: about 0.5 for VIP outputs and 0.5 to 2.5 for excitatory inputs onto VIP cells at 30 Hz. For connectivity, NeuronChat's AUROC of 0.832 is something our weighting should leave unchanged rather than improve. Beating the strongest baseline means two things. First, a lower held-out error in log paired-pulse ratio and higher Spearman correlation than nearest-neighbor lookup, with bootstrap intervals over connection types that exclude zero. Second, a better classification of depressing versus facilitating connections on held-out types.
Why not just…?
Why not just keep static scores, as CellChat and NeuronChat do? For questions about whether a connection exists, static scores are fine, and we do not compete there. For questions about signaling during sustained activity, the VIP data show a twofold drop between the first and second spike at 30 Hz, so a static score is wrong by construction. The uniform-depression baseline measures how much of that error a single global correction fixes. If uniform depression matches our model on held-out pairs, the pair-specific claim fails and we will report that.
Why not just look up recorded constants per cell class? This is the real threat. If held-out types are all close to recorded ones, nearest-neighbor lookup may tie our model. The leave-one-presynaptic-class-out split is designed to separate the two. Lookup has no near neighbor there, while a model that has learned which gene modules shift release probability can still extrapolate. A second decisive test is the external motor-cortex VIP data, a region absent from training. If lookup still ties, we will narrow the claim to "transcriptomes do not add beyond type identity." That would itself answer the question the VIP study left open.
Why not just regress parameters on a few hand-picked release-machinery genes? This is a reasonable prior, and it is effectively our presynaptic-only ablation with fewer genes. The VIP finding that the target cell sets plasticity for excitatory inputs predicts that it will fail specifically on those connections. Our per-class error breakdown checks this directly.
Decisive Checks
Three results would establish the idea, and we commit to them in advance. The first is held-out generalization: lower log paired-pulse-ratio error than nearest-neighbor lookup and the presynaptic-only ablation under leave-one-presynaptic-class-out cross-validation on the Synaptic Physiology Dataset. The second is the asymmetry test on the VIP motor-cortex recordings, which cover 60 VIP→Martinotti, 41 VIP→basket, and 25 excitatory→VIP connections. The model must predict uniform depression near 0.5 for both VIP outputs, with little sensitivity to postsynaptic expression. It must predict a broad, target-dependent spread for excitatory inputs, and the learned postsynaptic terms must carry that spread. The third is a continuum test. Where Patch-seq positions are available, predicted parameters should vary smoothly along within-type gradients, as electrophysiology does along the Pthlh/Pvalb axis, rather than jumping at cluster boundaries.
One caveat applies to all labels. They come from slices bathed in elevated calcium, which inflates release probability and biases recordings toward depression. We therefore treat absolute $U$ values as slice-specific and emphasize rank and sign agreement. Frequency-dependent MERFISH maps are released only if the first two checks pass.
References
- The short-term plasticity of VIP interneurons in motor cortex
- Inference and analysis of cell-cell communication using CellChat
- Screening cell–cell communication in spatial transcriptomics via collective optimal transport
- Inferring neuron-neuron communications from single-cell transcriptomics through NeuronChat
- Interactions between long- and short-term synaptic plasticity transform temporal neural representations into spatial
- Diversity of Interneurons in the Dorsal Striatum Revealed by Single-Cell RNA Sequencing and PatchSeq
- Integrating EM and Patch-seq data: Synaptic connectivity and target specificity of predicted Sst transcriptomic types