Speculate-then-Verify Simulation of Combinatorial Gene Knockouts

To learn what a gene does, biologists disable it and watch what breaks. But the interesting questions have outgrown that recipe: a gene's role depends on the cell type it acts in, and on which other genes are disabled alongside it. Conditional knockout biology — deleting a gene only in one designated cell lineage, for example deleting MyD88 only in pain-sensing neurons and finding that chronic neuropathic pain is blunted while acute pain is untouched (TLR signaling adaptor protein MyD88...) — was invented precisely because global deletion collapses these distinctions into systemic failure. The problem is scale: mapping even pairwise gene interactions empirically required 4,186 tested combinations for just 92 genes (Predicting transcriptional outcomes of novel multigene perturbations with GEARS), and each perturbation needs on the order of hundreds of cells to separate signal from noise (scPerturb: Harmonized Single-Cell Perturbation Data). So the field increasingly asks machines to predict knockout outcomes in silico. Today's predictors, however, are one-shot: they emit an answer in a single feed-forward pass and have no internal sense of when that answer is wrong.

Computer architects solved a structurally identical problem decades ago. In two-pass speculative execution, a processor runs memory operations optimistically in a fast front-end pass, then re-executes them in strict order in a verifying back-end pass, replaying only the rare instructions whose two answers disagree. Slackened Memory Dependence Enforcement showed that the fast speculative pass gets the right value about 98% of the time and only 2.4% of loads need replay, so the design lands within 2–5% of an idealized always-verify machine at a fraction of its cost (Slackened Memory Dependence Enforcement). This proposal transplants that architecture — a cheap speculative pass, a principled verification pass, and targeted replay — onto in silico prediction of combinatorial, cell-type-conditioned gene knockouts.

Core Claim

The residual error of cheap, additive knockout prediction is concentrated in a small, detectable subset of predictions, so a two-pass architecture can recover most of the accuracy of an expensive model at a small fraction of its cost. Concretely: pass one predicts the transcriptional effect of a multi-gene knockout by additively composing single-gene effects, exactly as current simulators do; pass two checks every predicted gene against learned dependency invariants on a regulatory knowledge graph, in direct analogy to how thread-level speculation detects read-after-write conflicts by piggybacking on cache-coherence invalidations (The Potential for Using Thread-Level Data Speculation to Facilitate Automatic Parallelization); flagged predictions are recomputed by an expensive pairwise model and committed in dependency order, like the in-order epoch commit of STAMPede (The STAMPede approach to thread-level speculation). The second claim, which is where the conditional-knockout direction earns its place: the verifier's invariants are relations between regulators and targets that hold across cell types, while the speculative pass carries all the cell-type-specificity. That split is what should let the system simulate a knockout conditioned on a lineage it was never trained on — the in silico counterpart of a Cre driver — which single-pass models cannot do without retraining (Predicting transcriptional outcomes of novel multigene perturbations with GEARS).

Research Program

Three components get built. The speculative front end is deliberately boring: a GEARS-style graph neural network that composes single-gene knockout embeddings additively and conditions on the basal expression state of the target population. This mirrors the L0 cache in SMDE, which serves values opportunistically with no correctness guarantee and no ordering enforcement (Slackened Memory Dependence Enforcement); the point is not front-end cleverness but making the front end cheap and swappable.

The verification pass is the new ingredient. It learns, from large libraries of single-gene perturbations — which exist at genome-wide scale, for example roughly 10,000 perturbations across more than 2.5 million cells in Replogle et al. (scPerturb: Harmonized Single-Cell Perturbation Data) — a set of dependency relations between regulators and targets on the same co-expression and Gene Ontology graphs GEARS uses as priors. For a gene $j$ whose knockout-combination effect is predicted as a log fold-change $\hat{\Delta}_j$, and whose regulators $i \in R(j)$ carry graph weights $w_{ij}$, the verifier raises a conflict flag

$$v_j = \mathbb{1}\left[\left|\hat{\Delta}_j - \textstyle\sum_{i \in R(j)} w_{ij}\,\hat{\Delta}_i\right| > \tau_j\right],$$

where $\tau_j$ is a per-gene threshold calibrated on held-out single-gene data. Intuitively: if a transcription factor is predicted to be strongly down but its documented target is predicted to go up, the additive composition has produced something the regulatory graph says cannot happen — the biological analog of a read-after-write violation. Flagged genes enter a small replay buffer, a direct analog of the write buffers that isolate uncommitted state in speculative multiprocessors (Embedded multi-core computing and applications, Architectural support for scalable speculative parallelization in shared-memory multiprocessors) and of EnfoRSer's choice of speculative buffering with rollback over idempotent deferral, which it benchmarked as the cheaper of the two isolation strategies at 27% versus 32% overhead (Hybrid Static: Dynamic Analysis for Statically Bounded Region Serializability).

The replay path then recomputes only flagged genes with an expensive pairwise model, resolving the buffer in topological order of the dependency graph so that commits never invalidate already-committed genes. Two refinements matter. First, a small pre-routing queue in the spirit of SMDE's Fuzzy Disambiguation Queue — which prevented 75–80% of replay-triggering violations — sends genes straight to the expensive path when they are knowably ambiguous, mainly genes with weak connectivity in the prior graphs, exactly where GEARS admits its accuracy drops (Predicting transcriptional outcomes of novel multigene perturbations with GEARS). Second, the conditioning mechanism handles partial penetrance: real conditional knockouts are incomplete — MyD88 positivity in target neurons fell from about 52% to roughly 24%, not to zero (TLR signaling adaptor protein MyD88...) — and Perturb-seq measurements likewise mix perturbed and unperturbed cells, so the system must predict and verify effects parameterized by a perturbation fraction, not assume a clean null.

A note on what this is not. The "attention knockout" literature ablates attention heads inside transformers to map which components carry facts or grammar (Causal Circuit Verification via Attention Knockout, e.g. Mor Geva et al., Olga Kovaleva et al.). That work borrows the knockout metaphor for network weights and evaluates ablations post hoc. Here the knockout is the object being predicted, and the two-pass machinery is the predictor; we reuse the architecture lesson (ablate cheaply, verify, replay) rather than the metaphor.

Baselines

The do-nothing reference is the additive predictor: the predicted effect of a two-gene knockout is simply the sum of the two individually predicted effects, with no interaction modeling. This is the floor every published method clears, and it isolates how much of the task is trivially additive. A related naive reference is plain linear regression from basal state to perturbed state; scDEBART reports that large single-cell foundation models surprisingly often fail to outperform it (scDEBART: Predicting in silico Single-Cell Perturbation Responses via Large-Scale Differential Expression Learning), but no reported result was found on Lacuna for a linear-regression baseline's numerical performance on the Norman combinatorial task, so we will establish that number ourselves under identical splits.

The strongest existing competitor is GEARS (Predicting transcriptional outcomes of novel multigene perturbations with GEARS), the graph-neural simulator whose numbers define the current bar on exactly our benchmark: a 53% improvement over prior models on zero-shot two-gene perturbations where neither gene was seen individually during training, 40% higher precision than existing approaches on four genetic-interaction subtypes, and a 30–50% mean-squared-error improvement over CPA on the most relevant genes. GEARS is the right chief competitor because our speculative front end is essentially its compositional module, so any gain is attributable to the verification pass rather than backbone capacity. It will be run with its own published architecture and hyperparameter protocol on identical Norman et al. and Replogle et al. splits, with our front end given the same training data, the same prior graphs, and an equal hyperparameter search budget. The second strong competitor is scDEBART (scDEBART: Predicting in silico Single-Cell Perturbation Responses via Large-Scale Differential Expression Learning), pretrained on 6.28 million differential expression profiles and reporting a mean DEG enrichment factor of 11.96 versus 1.74–2.99 for GEARS-class baselines; we will run its pretrained model on the same splits and, critically, also wrap our verifier around its backbone, since the verification pass is deliberately model-agnostic.

Beating these means: on the Norman two-gene held-out task, lower mean squared error on the top differentially expressed genes and higher precision on non-additive interaction subtypes than GEARS under identical splits, with the margin exceeding split-to-split variance across five resamples; and a lower error-per-FLOP frontier than scDEBART, since raw accuracy alone at unbounded compute would not establish the two-pass claim. The key-ingredient ablation is the same system with the verification pass switched off, which reduces it to a single-pass additive model and isolates what the second pass contributes; its mirror, an always-verify variant that recomputes every gene with the expensive model, isolates what speculation contributes over blanket verification.

Why not just…?

Why not just train a bigger single-pass model? This is the most serious objection, and the honest answer is that scDEBART's enrichment factor of 11.96 shows capacity and pretraining scale do buy a lot (scDEBART). But the same paper concedes that foundation models often fail to beat linear regression because they learn static co-expression rather than causal regulatory dynamics, and it operates in one-shot inference with no mechanism to flag when a prediction violates regulatory structure. The decisive experiment is built into the proposal: wrap our verifier around scDEBART's own backbone. If the two-pass structure adds nothing on top of the strongest available single-pass model, the idea fails, and we will report it as such. GEARS' own admission that additive composition degrades under epistasis and weak graph connectivity (GEARS) is prior evidence that capacity alone leaves exactly the gap we target.

Why not just always run the expensive model on everything? This is the biological analog of conservative synchronization — give up speculation, verify everything — and if verification were free it would be the right answer. The systems literature quantifies why it usually is not: idealized fully-ordered designs are the expensive ceiling that speculative designs approach at a few percent of cost, and EnfoRSer measured the conservative idempotent variant as strictly costlier than speculative rollback (Slackened Memory Dependence Enforcement, Hybrid Static: Dynamic Analysis...). In our setting the expensive path cannot even be trained everywhere, because pairwise combinations are what the experiments cannot supply. Still, we treat this alternative as a live threat rather than a strawman: the compute–accuracy frontier experiment below includes the always-verify point explicitly, and if it matches two-pass accuracy at acceptable cost, the speculative structure is unjustified.

Why not just run more CRISPR screens instead of simulating? Empirical measurement is the gold standard, but the arithmetic is unforgiving: 4,186 combinations for 92 genes (GEARS), a minimum of about 300 cells per perturbation and 1,000 UMI counts per cell to clear noise (scPerturb), and, worse, dissociated screens miss effects that depend on neighboring cells, which spatial Perturb-FISH assays showed can even flip sign with local cell density (Simultaneous CRISPR screening and spatial transcriptomics...). In vivo conditional knockouts scale even worse, one transgenic line at a time, as the MED1 macrophage study illustrates for a single gene in a single injury model (MED1 Deficiency in Macrophages Accelerates Intimal Hyperplasia...). Simulation does not replace these experiments; it ranks which of the untestably many combinations deserve them, and our final experiment tests whether the verifier's conflict flags are good rankers.

Proposed Experiments

Experiment one runs on the Norman et al. combinatorial dataset of 131 two-gene CRISPR perturbations in K562 cells, the benchmark on which GEARS reports its 53% zero-shot improvement (GEARS). We hold out the same combinations, train on single-gene data, and report mean squared error on the top differentially expressed genes, precision on the four interaction subtypes, and the fraction of genes flagged for replay. The central hypothesis, borrowed from SMDE's 2.4% replay rate, is that flagged predictions are a small minority that contains most of the additive model's error, and that accuracy on flagged predictions after replay approaches the expensive model's.

Experiment two tests the conditional-knockout claim on Replogle et al. genome-wide data across the RPE-1 and K562 lineages, over 340,000 cells as used by GEARS: train the speculative front end on one lineage, transfer only the verifier's regulatory invariants, and predict knockouts in the other conditioned on basal expression. Success is defined as error close to a model trained in-domain, where GEARS must be retrained per cell type. Experiment three stress-tests penetrance by mixing perturbed and unperturbed expression profiles at controlled fractions, asking whether the verifier flags degraded predictions and whether conditioning on the true fraction recovers them — the computational counterpart of Cre leakage, for which the MyD88 study's 52%-to-24% reduction is the reference behavior (TLR signaling adaptor protein MyD88...).

Experiment four traces the compute–accuracy frontier by sweeping the flag threshold $\tau_j$, plotting error against total inference FLOPs, with the additive predictor and the always-verify variant as the two endpoints; the two-pass claim is confirmed only if the curve dominates its middle. Experiment five closes the loop empirically: use conflict flags to rank untested combinations by predicted non-additivity, and check the ranking against the 92-gene, 4,186-combination fitness screen that GEARS used for biological confirmation (GEARS), and against Perturb-FISH's observation that neighborhood context changes outcomes (Simultaneous CRISPR screening and spatial transcriptomics...).

The decisive checks are three, and any one can kill the idea. If flagged predictions are not substantially enriched for true non-additive interactions, the conflict detector is decoration and the verifier fails as a measurement instrument. If always-verify matches two-pass accuracy at comparable cost on the frontier, speculation buys nothing here and the architectural transplant is pointless. And if wrapping the verifier around the strongest single-pass backbone — GEARS and scDEBART alike — improves neither, then the gap this proposal targets was never real.