GrnPFN learns from simulated gene networks: it predicts transcriptional interactions before any combination is measured

Combined perturbations elicit transcriptional responses that depart from the additive expectation, and GrnPFN makes that residual itself the prediction target, defined as the pair response minus the sum of the two single-perturbation effects, so interaction prediction is scored separately from the larger additive part. Earlier combination predictors mainly predicted the full expression response, and in the K562 combinatorial screen adding single-perturbation effects already captured much of the accuracy for full responses, leaving the nonadditive component without a dedicated target. With no measured pairs in the target screen, GrnPFN's pooled cosine similarity exceeded gradient boosting trained on the other four screens by 0.049 (97.5% CI, 0.038–0.060), with the largest gains in K562, A172 and U87MG and similar performance in THP-1 and T98G.

The model first learns how single effects relate to interactions from simulated data, so at application it can predict zero-shot from single-perturbation profiles alone or take measured pairs as context without retraining. Regression models fitted directly to measured pairs in the target screen have little evidence from which to learn the nonadditive component when combinations are sparsely sampled. Pretraining used 140,000 simulated screens from random gene-regulatory networks with unmeasured pathway nodes, perturbable transcription factors, saturating program nodes and measured genes, spanning CRISPR activation, knockdown and drug combinations; at 15% of training pairs measured, GrnPFN exceeded kernel ridge fitted to the same pairs by a pooled 0.069 (95% CI, 0.048–0.090).

The low-budget advantage extends to perturbation members absent from measured combinations and to a gene–drug design absent from pretraining. Sparse budgets leave some perturbations entirely out of the measured combinations, while the transfer baselines were trained on all pairs from other screens. At the 15% budget GrnPFN exceeded kernel ridge by 0.091 on unseen pairs (95% CI, 0.059–0.122) and by 0.069 on seen pairs (0.046–0.095); in the A375 CRISPRi plus vemurafenib screen, not used in pretraining, zero-shot cosine similarity was 0.421 versus 0.319 for gradient boosting trained on the five main screens, a difference of 0.102 (95% CI, 0.090–0.113).

Evaluation covers CRISPR activation in K562, Cas13 knockdown in THP-1, trametinib drug combinations in A172, T98G and U87MG, and a CRISPRi plus vemurafenib screen in A375; model development used one cell partition of the first four screens, the four later partitions reused the same cells, and U87MG was excluded from model selection but shared its compound panel and anchor design with two development screens. Pair folds retained substantial overlap in members or compounds, so the evaluation mainly concerns unmeasured combinations of perturbations with known single effects.

The authors note that transcriptional nonadditivity does not necessarily indicate a specific molecular mechanism or imply synergy on other phenotypes such as cell fitness; the A375 drug–gene screen supported zero-shot transfer but accuracy declined as context increased, and fine-tuning gave consistent gains of 0.024–0.080 in A172 and 0.044–0.078 in T98G but inconsistent effects in K562, THP-1 and U87MG. Dependence-aware intervals address only part of the dependence from shared members or compounds and do not cover every shared control or anchor effect, and the four cell partitions are not treated as independent biological replicates; the authors state that prospective evaluation in independent biological experiments and additional biological contexts is still needed.