Executable Gene-Circuit Models Checked Against RNA Velocity

Language and vision models describe a physical scene well but cannot predict what it will do next. The clearest fix found so far is to stop asking the model for the answer. Instead, the model writes a small simulation of the scene, runs it in a physics engine, compares the result with what was observed, and revises the simulation. Single-cell biology has the same problem in a different form. Models that predict how cells respond to a drug or a gene knockout tend to output a safe population average, and they collapse on perturbations they have not seen. The two recent "virtual cell world models" in this space reason in text over knowledge graphs; neither runs a simulation that could be checked against data. This note proposes carrying the physics recipe across. A language model writes executable gene-circuit code, a simulator turns it into a population of virtual cells, and the code is accepted only if it reproduces both the measured expression and the measured direction of change of each cell, known as RNA velocity.

The central bet is that verification, not language reasoning, is what transfers from physical common sense. scRNA-seq (single-cell RNA sequencing, which counts the messenger RNA in thousands of individual cells) destroys each cell it measures, so the data are static snapshots. RNA velocity partly escapes this limit. The ratio of newly made, still-unspliced RNA to mature, spliced RNA gives each cell a local arrow saying where its expression is heading. A mechanistic model that predicts unspliced and spliced RNA separately can be tested against these arrows as well as against average expression. A model that has only learned to output averages should fail that second test. In physics terms, velocity plays the role of the dynamic field that keeps a model from settling on a static answer.

Core Claim

We claim that an analysis-by-synthesis loop improves prediction of unseen genetic and chemical perturbations, especially of large changes in key genes, over both statistical predictors and text-based reasoning models. Each candidate is a small executable kinetic model of the genes around a perturbation, and it is accepted only if its simulated cells match the observed distribution of expression and RNA velocity. We further claim that the language model's specific contribution is proposing circuit structure where curated knowledge graphs are incomplete. This second claim is falsifiable, and we design an ablation to test it directly.

Prior Work and Where This Sits

Two physics papers supply the method. LLMPhy has a language model write MuJoCo code for an observed scene and refine hidden parameters such as friction and damping through an iterative optimization trace. Code as Worlds generalizes this into a propose–execute–render–verify loop over scene composition, dynamics and appearance, and trains the proposer with reinforcement learning. We reuse the loop but replace rigid bodies with stochastic transcription–splicing–degradation kinetics. We also replace the rendered video with RNA velocity.

LLMPhy also constrains us. Its gain over plain Bayesian optimization was small (62.0% vs 59.6% mean IoU on TraySim), which suggests the language model adds little when its job is only parameter tuning. We therefore give it a different job, choosing circuit structure, and we test whether that job helps.

Scene Dynamic Fields motivates the velocity signal. That paper shows that multimodal models attend to the last video frame and miss fluid motion until explicit velocity fields are supplied. Hodge Decomposition of Single-Cell RNA Velocity supplies the field analysis. It splits a cell velocity field into three parts: an irrotational part that tracks differentiation, a rotational part that isolates the cell cycle, and a harmonic part that marks fate branch points. That work is a descriptive tool applied after the fact; we use its components as verification targets for a predictive model.

On the biology side, VCWorld and PertReasonQA are the closest work.

  • VCWorld retrieves analogous cases from a biological knowledge graph and has a language model reason step by step to predict up, down or unchanged.
  • PertReasonQA conditions regulatory paths on the cell's baseline state and trains a 4B model with GRPO, a reinforcement-learning method that scores groups of sampled answers against each other.

Both produce categorical text answers, and neither checks a hypothesis against data by running it. Ours produces a simulated cell population and is checked by execution.

Three further papers set the bar any new method must clear:

  • PertEval-scFM found that large single-cell foundation models often fail to beat a simple MLP.
  • SCALE deliberately avoids modeling continuous trajectories because flows are unstable on sparse data. We address this by imposing velocity only locally, on small circuits, rather than learning a genome-wide flow.
  • CellVerse shows that generalist language models given cells as text sit near 50–55% on drug response.

Research Program

For each perturbation, we first select a module of 20 to 100 genes. It contains the perturbed gene or drug targets, their knowledge-graph neighbors, and genes that vary strongly in control cells.

The proposer, an open language model, then writes Python that defines the circuit. The circuit specifies which genes regulate which, whether each regulation activates or represses, and the kinetic form of each gene's dynamics. Each gene has unspliced and spliced RNA, with transcription driven by Hill functions of its regulators (smooth switch-like response curves). The circuit also specifies how the perturbation enters: a knockdown scales one transcription rate, and a drug modifies its target's activity.

A stochastic simulator runs the circuit from cells sampled from the measured unperturbed population and returns simulated perturbed cells with both RNA species. A verifier scores the candidate on three things:

  • the distance between simulated and observed perturbed expression distributions on module genes (energy distance);
  • agreement between simulated and observed velocity directions per cell;
  • where a dataset has strong cyclic or branching structure, agreement of the Hodge components.

Parameters inside a fixed structure are fitted numerically by gradient-free or gradient-based optimization. The language model sees the verifier's error report and edits the structure. On training perturbations, the proposer is then fine-tuned with GRPO, using the verifier score on held-out cells as the reward.

The reward design responds to the shortcut finding in Can Vision Language Models Learn Intuitive Physics from Interaction?. That paper found 94–97% in-distribution accuracy with no real-world transfer. A reward computed by executing a mechanism against held-out cells should be harder to game than a reward for matching a label, and our experiments test whether it is.

A synthetic testbed accompanies the real data. Ground-truth circuits are drawn from curated topologies, simulated with the same kinetics, and given scrambled gene names. On this testbed we can measure edge recovery directly, and the language model cannot answer from memory.

Baselines

The naive reference is to predict the unperturbed control population, together with the mean baseline from PertEval-scFM. An additive linear model and the MLP on raw expression that matched foundation models in that benchmark sit one step above it. We train all of these on the same splits.

The strongest statistical competitor depends on the dataset:

  • Genetic screens (Norman, Replogle): GEARS, which encodes gene interaction graphs. On the Norman screen it reports AUSPC (area under the SPECTRA performance curve, a measure of how prediction error grows as test perturbations become less similar to training ones) of about 0.008, against about 0.045 for scGPT, per PertEval-scFM. In that benchmark's reporting, lower is better.
  • Drugs on Tahoe-100M: STATE and SCALE. SCALE reports a 12.02% PDCorr gain and a 10.66% DE Overlap gain over STATE.

For reasoning models, the bars are as follows. PertReasonLM reports 0.736 balanced accuracy on PertReasonQA, against GEARS at 0.356 and STATE at 0.411. VCWorld reports 0.70–0.76 differential-expression accuracy on GeneTAK.

Fairness rules apply to all comparisons. Every method receives the same knowledge-graph sources: the graph databases named by VCWorld and the interaction databases named by PertReasonQA. Every method uses the same SPECTRA or cluster-based out-of-distribution splits. Every method receives the same hyperparameter search budget in GPU-hours. The proposer's backbone is matched in size to PertReasonLM. We convert our simulated cells to up, down or unchanged labels using each benchmark's own thresholds.

Beating the field means two things. First, lower AUSPC than GEARS on the Norman SPECTRA splits. Second, higher balanced accuracy than 0.736 on PertReasonQA's out-of-distribution clusters, or within noise of it while also producing continuous distributions that those models do not produce.

Three ablations isolate the key ingredients:

  • Velocity off: the verifier sees expression only.
  • KG structure + numeric fit: the circuit is read directly from the knowledge graph, and only parameters are fitted. The language model never edits structure.
  • SFT instead of GRPO: the proposer is trained by supervised fine-tuning on accepted circuits rather than reinforcement learning on verifier reward.

Why not just…?

Why not just pre-train a larger single-cell foundation model? PertEval-scFM found zero-shot foundation models barely beating the mean baseline and degrading under distribution shift. PhysBench finds the same pattern for scale in physics. We still include the best foundation model on identical splits rather than relying on that argument.

Why not just prompt a strong language model to reason step by step? In CellVerse, text-only models sat at 50–55% on drug response. GRASP found that step-by-step prompting does not fix dynamic physical reasoning. Grounded text reasoning does work better: VCWorld and PertReasonLM exceed 0.7. That makes PertReasonLM a real threat, and it is a headline baseline rather than a straw man.

Why not just take the circuit from the knowledge graph and fit its parameters numerically? This is the most serious alternative. LLMPhy's narrow margin over Bayesian optimization suggests it may match the full loop. The KG + numeric-fit ablation tests it directly. We stratify results by whether the perturbation's true or responding regulators are missing from the knowledge graph; this is measurable exactly on the synthetic testbed and approximately on real data by holding out edges.

If structure editing does not help on the missing-edge stratum, we drop the language model from the claim. The contribution then becomes velocity-verified mechanistic modules, which the velocity-off ablation still tests.

Why not just fine-tune on observed perturbation outcomes? Supervised training on outcomes is exactly the setting where Schulze Buschoff et al. found shortcut learning. The SFT ablation, evaluated on the farthest SPECTRA splits, shows whether execution-based reward changes this.

Proposed Experiments

The first decisive check runs on the synthetic testbed. With scrambled gene names, does the loop recover more true edges and predict held-out perturbations better than KG + numeric fit, and does the velocity term reduce wrong-sign edges that still match the means? If velocity does not separate circuits that expression alone cannot, the main premise fails.

The second check runs on the Norman and Replogle Perturb-seq screens, using SPECTRA splits and reporting AUSPC against GEARS. Velocity here is computed from unspliced and spliced counts, which requires raw reads. We also report error on the top differentially expressed genes per perturbation, to test the claim that execution-based verification avoids the mean-effect trap. These screens are in K562 cells, where velocity is dominated by the cell cycle. We therefore also report agreement on the rotational Hodge component, as a check that perturbed circuits do not break cycling unless the data show it.

The third check runs on PertReasonQA and GeneTAK. Converted to categorical labels, our predictions are compared against 0.736 balanced accuracy and 0.70–0.76 DE accuracy respectively.

The fourth check is a generalization test without perturbation labels, on the endocrine pancreas and hippocampus velocity datasets. Circuits fitted on early cells should reproduce the harmonic, branch-marking component of later cells.

A memorization control appears in every real-data experiment: we rerun with gene identifiers scrambled, so any gain that disappears is attributed to recalled literature rather than to verification.