Occam predicts single-cell perturbation effects with a plain MLP, beating retrained complex models across cell lines and gene pairs
Synopsis
Occam hands both single-cell representation learning and perturbation-effect prediction to multilayer perceptrons: swap-in denoising reconstructs replaced gene values, then a frozen encoder is modulated by gene-network, text and protein-sequence features. In HepG2, where a perturbation and a cell line are each seen in training but never together, Occam agrees with reference delta profiles better than experimental reproducibility does; on 111 held-out double perturbations it tracks measured responses more closely than summing the two measured single-gene effects.
Interpretation
Single-cell perturbation-effect prediction can now be done with a multilayer perceptron rather than pretrained transformer embeddings or complex generative predictors. Reusable cell representations have mostly come from transformer models, yet pretrained embeddings and complex predictors do not consistently outperform simple alternatives for perturbation-effect prediction. Occam-90M, pretrained on roughly 1.4 million DiverseRNA profiles, scores 39.31 on the aggregate perturbation-information representation benchmark and 53.72 on scIB cell integration, the highest among evaluated models and baselines; it leads on RPE1, HepG2 and Pancreas, while TxFM-B leads on Jurkat.
When a perturbation and a cell line are each seen but their combination is not, Occam recovers perturbation-specific response patterns better than four retrained competitors and a transfer baseline. The transfer baseline copies the same perturbation's measured effect from other training cell lines without learning how that effect changes with cell line. On the 611 RHJ/HepG2 test perturbations, Occam's mean Pearson exceeds experimental reproducibility (0.3999 versus 0.3722) and its L1 discrimination is 0.761 versus 0.668 for the next-best STATE; its mean absolute error is 0.0269, behind TxPert's 0.0228.
Double perturbations no longer have to be approximated by adding two measured single-gene effects, because Occam applies the single-gene predictor in both orders and averages the results. The additive baseline adds the two measured single-gene effects directly as the reference for the double-perturbation response. On the 111 held-out Norman double perturbations, Occam reaches Pearson 0.8526 and mean absolute error 0.01225, versus 0.8333 and 0.01465 for the additive baseline; its Pearson advantage is 0.0264 on weakly additive doubles and 0.0133 on strongly additive ones.
Perspective
The demonstrated scope is an encoder pretrained on DiverseRNA and evaluated frozen on three cell lines absent from that corpus, plus perturbation prediction on RHJ HepG2, K562 and H1 and on Norman K562 double perturbations, where neither K562 nor H1 took part in predictor training. The intended reader is a researcher working on single-cell perturbation modeling and benchmarking.
The reported results mainly assess population-level effects and do not establish accurate prediction of how an individual cell responds or how responses vary between cells; Norman covers only double perturbations whose constituent single-gene responses were observed in training, and unseen target genes, perturbations of more than two genes and chemical perturbations are not evaluated. K562 and H1 differ in both cell line and measurement conditions, so their comparison cannot isolate either factor; on held-out cell lines the transfer baseline has higher Pearson than Occam, and no model leads on every metric.
