Occam paper shows simple MLPs predicting genetic perturbation effects across cell contexts
The authors report that a multilayer perceptron combining denoised cell representations with gene, text and protein features improves perturbation-effect prediction, though expression-level accuracy remains difficult.
In plain words
Researchers behind the Occam project describe a neural network that predicts what happens to a cell when genes are switched off or altered. The model learns cell representations by denoising data and adds information about gene interactions, scientific text and protein sequences. The authors say it handles new gene-and-cell-type combinations better than retrained competitors, but accurate prediction of expression levels is still hard.
What was announced
In a paper posted to arXiv, the Occam authors present a multilayer perceptron framework for representation learning and perturbation-effect prediction. The model combines cell representations learned through denoising with gene interaction, text and protein-sequence features, and the authors report the highest aggregate representation scores for perturbation information and cell-type preservation with batch integration.
The evaluation covers cross-context few-shot settings, where a perturbation and a cell context are each seen during training but never together. According to the abstract, Occam captures measured response patterns for these unobserved combinations better than four retrained competitors or direct transfer of a perturbation's measured effects from other training cell lines, and without fine-tuning it best distinguishes perturbations in held-out cell lines across seen and unseen assays. For double perturbations targeting two genes, the model composes learned single-gene responses, reducing expression error compared with competing models and summed single-gene effects, with gains on both strongly and weakly additive responses. The authors caution that predicting expression levels accurately in these settings remains challenging.
What it does not show
- The evidence is the paper's abstract only; the full methods, datasets and detailed results are not available in the supplied passages.
- All performance claims come from the authors themselves and have not been independently confirmed.
- The abstract does not specify model sizes, training data volumes, or which specific assays and cell lines were evaluated.
For security and IT teams
- Scope
- The supplied source is the arXiv abstract only, so claims reflect the authors' summary rather than the full paper.
Sources
Source-support review completed 19 Nov 2026. This story was drafted with AI assistance and checked against retained source passages. Automated checking is not independent verification.
Coverage: arXiv: abstract.
Updates and corrections
No corrections so far.
First published 19 Nov 2026
New to the subject? AI from the Ground Up starts from the beginning.