Logged Light Histories Close the RNA–Protein Prediction Gap

Every model that predicts a cell's protein levels from its RNA is missing an input. Systematic proteome-wide measurements show that mRNA abundance tracks protein abundance only 40–60% of the time, leaving roughly half of protein-level variance unexplained by transcripts (A high-stringency blueprint of the human proteome). Some of that missing variance is surely history: what the cell has recently been doing. The clearest evidence comes from memory research, where the kinase PKMζ stays elevated in engram neurons for a month after learning while its transcript Prkcz shows zero differential expression (Persistently Increased Expression of PKMzeta and Unbiased Gene Expression Profiles Identify Hippocampal Molecular Traces of a Long‐Term Active Place Avoidance Memory and "Shadow" Proteins). Present any current RNA-to-protein model with those neurons' transcripts and it will predict baseline protein. The information simply is not in the RNA.

The obvious missing input is causal stimulation history, and optogenetics — light-gated actuators that let you switch a defined cell population on and off with millisecond timing — is the only tool that delivers it with per-cell, per-moment precision and full reversibility. Yet no perturbation resource pairs optical control with single-cell multi-omic readout: scPerturb: Harmonized Single-Cell Perturbation Data harmonizes 44 datasets, all CRISPR or chemical, with zero optogenetic time courses, and single-cell optical platforms stop at a single fluorescent reporter readout (TOWARDS PRECISION OPTOGENETIC CONTROL OF MAMMALIAN CELL SIGNALING PATHWAYS BASED ON MICROFLUIDICS). This proposal builds the missing instrument-model pair: a closed-loop single-cell optogenetic platform that logs the exact light dose delivered to every cell, paired with single-cell transcript and protein measurements, used to train and test a protein-prediction model conditioned on that logged history. The payoff is twofold: a causal, per-cell accounting of how much of the transcript–protein gap is explained by stimulation history, and the first perturbation-dynamic RNA-to-protein model.

Core Claim

The claim is that under acute, controlled stimulation a substantial and quantifiable fraction of single-cell protein-level change is (i) invisible in the transcriptome — the "shadow protein" regime — yet (ii) predictable from a per-cell record of delivered stimulation. Formally, we model protein abundance $y_i$ for cell $i$ as

$$y_i = f(z_i, h_i) + \varepsilon_i,$$

where $z_i$ is a latent representation of the cell's RNA counts (as in existing models) and $h_i$ is an encoding of its logged light history — dose, timing, and pulse structure. Existing models are the special case $h_i = \emptyset$, and the PKMζ result shows that case has a hard ceiling. Our experiments are designed to measure how much ceiling is lifted by $h_i$, and to rule out the boring alternative that RNA simply carries the signal if you model it well enough.

Research Program

The program has three components: an instrument, a dataset, and a model.

The instrument extends the microfluidic closed-loop platform of TOWARDS PRECISION OPTOGENETIC CONTROL OF MAMMALIAN CELL SIGNALING PATHWAYS BASED ON MICROFLUIDICS, which already achieves 2.49 μm spatial resolution for cell-by-cell optical actuation and real-time computer-vision tracking of individual trapped cells. We keep its trap array, segmentation, and Digital Light Processing mask projection, and add two things: a per-cell stimulation ledger (every pulse, its duration, and its projected intensity, tied to a tracked cell identity), and molecular readout beyond the mCherry reporter — barcoded single-cell RNA capture plus in-situ immunofluorescence for a protein panel at multiple time points (minutes to hours post-stimulation). This matters because identical nominal pulses reach cells differently: opsin expression, trafficking, and dendritic morphology make the effective dose vary cell to cell, so optical stimulation recruits neurons unevenly (Optogenetic stimulation recruits cortical neurons in a morphology-dependent manner). In a light cone over cortex, that per-cell dose is unmeasurable; in a trap array, it is logged exactly. Fast optical control of protein-regulatory machinery is established in principle — light-gated control of nuclear protein export has been demonstrated (Optogenetic control of nuclear protein export) — so actuation at the relevant timescales is feasible.

The dataset is the first optogenetic entry in the perturbation-resource sense of scPerturb: Harmonized Single-Cell Perturbation Data. We will follow its experimental rules of thumb — at least 300 cells per perturbation condition and about 1,000 UMI counts per cell — and adopt its Energy Distance and permutation E-test as our primary statistic for measuring how far a stimulated cell cloud moves from its unstimulated control in protein space. Design: neuronal cells (a PKMζ-relevant line) carrying an excitatory opsin, stimulated under a grid of pulse counts, frequencies, and recovery intervals, with unstimulated trapped cells on the same chip as matched controls; readouts at early (minutes, post-translational) and late (hours, transcriptional) time points.

The model is a conditioning extension of the two-stage architecture in scChord: A Probabilistic Manifold Rectification Framework for RNA-to-Protein Translation: the same noise-aware latent for sparse RNA counts (a Variational Autoencoder with a zero-inflated negative binomial decoder), with a small light-history encoder whose output is concatenated into the conditional flow that transports RNA embeddings to protein space. scChord is the right scaffold because it already absorbs single-cell technical noise, which is essential when the biological signal (acute protein shifts) can be smaller than the noise floor; our addition is deliberately minimal — the conditioning pathway — so that any gain is attributable to it and nothing else.

Baselines

The naive reference point is a do-nothing model: predict each cell's protein levels as the unperturbed population mean for that condition. It sets the floor for per-protein error and, importantly, its error on the E-distance metric shows how large the stimulation effect is at all — if light does not move the protein cloud, the whole premise fails here and we report that honestly.

The strongest existing competitor is the RNA-only generative translation model. We will run two: scChord, retrained from scratch on our dataset with its published architecture and its hyperparameter search budget, and A pre-trained large generative model for translating single-cell transcriptome to proteome (scTranslator), fine-tuned on our data — its gene positional encoding lets it operate on arbitrary gene panels, so the comparison is fair even with our targeted panel. Both get the same training cells, the same protein panel, the same compute, and the same tuning budget as our model, differing only in that they receive RNA and not light histories. On absolute performance the file gives us one anchor: scTranslator reports a 10.7% increase in cosine similarity over sciPENN and cTP-net on gold-standard datasets, and maintains similarity above 0.83 in unaligned gene settings; no reported result found on Lacuna for scChord's exact RMSE or Pearson correlation on GSE100866, so we will treat scTranslator's reported operating range as the reference point to clear. Beating the baselines means, concretely, lower per-protein RMSE and higher Pearson correlation on held-out cells, and — the decisive criterion — lower error on held-out stimulation regimes (new pulse frequencies and doses the model never saw), where an RNA-only model should fail badly because the transcript profile of two different light programs can be nearly identical while proteins differ.

The key ablation is our own model with light histories replaced by shuffled histories from other cells in the same experiment. If the conditioning variable carries real information, shuffling must collapse performance exactly to the RNA-only baseline; if it does not, whatever gains we observed came from added capacity, not from the histories, and the core claim is dead. A second ablation removes the time-resolved structure of $h_i$ and keeps only integrated dose, which tells us whether timing or amount is the informative part.

Why not just…?

Why not just add optogenetic datasets to scPerturb and leave the modeling to others? This is a real threat to the framing, and the honest answer is that the dataset is half the contribution and we will release it in scPerturb's harmonized format. But the resource alone cannot answer the question. scPerturb's E-statistics quantify how far a perturbed cell cloud moved; they do not predict per-cell protein states, and they cannot decompose an observed protein shift into a transcript-mediated part and a history-mediated part. That decomposition requires the conditioning model, and the PKMζ result predicts the decomposition is non-trivial: it is precisely the case where the cloud moved and the transcripts did not (Persistently Increased Expression of PKMzeta...).

Why not just stimulate in vivo and read out with bulk proteomics? Logistically simpler, and it would be the natural follow-up. But bulk averaging is the third confound this proposal exists to remove: rare responding cells are diluted exactly as rare stem-like tamoxifen-resistant cells are invisible in bulk breast-cancer profiles (Integrative analyses of gene expression and DNA methylation profiles in breast cancer cell line models of tamoxifen-resistance indicate a potential role of cells with stem-like properties), and existing deconvolution cannot help because both Count Bridges enable Modeling and Deconvolving Transcriptomic Data and Cross-modal transfer learning for mapping bulk transcriptomes at cellular level operate in RNA space only. Worse, in vivo the conditioning variable is confounded: a light cone illuminates thousands of cells at unknown per-cell effective dose, and morphology-dependent recruitment makes that dose heterogeneous in an unmeasurable way. Our in-vitro result would be the necessary pilot that justifies the in-vivo version.

Why not just fine-tune a bigger static model (scTranslator) on stimulated cells? This is the strongest alternative, because a model fine-tuned on data from a given stimulation condition can learn the condition-average protein shift without any per-cell conditioning. It would plausibly match us on population-level metrics. Where it must fail is per-cell prediction within a stimulated condition and across unseen regimes: if two cells in the same dish received different light programs and their transcripts are nearly identical, an RNA-only model predicts the same proteins for both no matter how large it is. Our shuffled-history ablation and the held-out-regime evaluation are designed precisely to expose this — if fine-tuned scTranslator matches us there, the RNA does carry the signal and we will report that the shadow-protein regime is narrower than the PKMζ case suggested.

Proposed Experiments and Decisive Checks

Experiment one measures the shadow fraction. Across the stimulation grid, we compute the E-distance between each stimulated protein cloud and its matched control, then ask how much of that shift a retrained RNA-only scChord predicts from the accompanying RNA data. The residual — protein shift with no transcript-predictable component — is the causal shadow fraction. The PKMζ result licenses the expectation that it is nonzero at early time points, where rapid phosphorylation-type responses dominate (Signalome-wide mapping of the NFκB pathway in T-cells reveals novel targets for immunotherapy); at late time points transcription catches up and the fraction should shrink, which is itself a testable signature.

Experiment two is the out-of-distribution test, taking its logic from perturbation neuroscience: Biologically informed cortical models predict optogenetic perturbations showed that models with the right inductive bias hold up under optical perturbations they never saw (7% average error in predicting hit-probability changes, versus 48.5% for unconstrained networks), while the wrong ones collapse. We evaluate all models on held-out light regimes — interpolated frequencies and extrapolated doses. The conditioning model should extrapolate in the dose dimension because integrated dose is in $h_i$; the RNA-only baselines have no path to.

Experiment three closes the loop on the PKMζ story specifically: whether direct optical activation of the circuit-like engram, which the TRAP2 study tagged but never stimulated, is sufficient to drive protein-level elevation without transcript change. If our platform reproduces a PKMζ-like signature under controlled light, the proposal delivers what that study listed as open: causal evidence that optical actuation alone writes a shadow protein trace, at single-cell resolution rather than bulk microdissection, and with a predictive model attached.