Loading…
Cell microscopy foundation model with a 1.9-billion-parameter masked autoencoder producing embeddings that stay consistent across screening batches.
No providers recorded yet. Browse all providers
In a genome-scale microscopy screen, most perturbations do nothing visible. Knock out a gene the cell does not need and the well looks like an untreated control; kill the cell outright and the well looks like every other dead-cell well. Train a masked autoencoder on a random sample of such a screen and most of its capacity goes to modelling the unperturbed phenotype and a few common failure modes, rather than the subtle differences that separate one perturbation from another. Phenom-2 is built on the finding that removing those uninformative wells buys more than collecting more of them.
Phenom-2 — called MAE-G/8 in its paper — is a 1.86-billion-parameter vision transformer masked autoencoder from Recursion and Valence Labs, published at ICML 2025. It is the largest member of Recursion's Phenom family of Cell Painting encoders, whose smallest and only openly released member is OpenPhenom-S/16. It is used strictly as a frozen feature extractor: images go in, embeddings come out, and every downstream benchmark is a linear probe or a cosine similarity computed over those embeddings.
The paper's second contribution concerns where in the network the embedding is read from. Probing all 48 transformer blocks shows that intermediate blocks carry more biologically linear features than the final block that vision models conventionally use — a result that holds for microscopy-pretrained and natural-image-pretrained transformers alike. Phenom-2 is therefore deployed trimmed, at block 38, which improves the representation and cuts inference cost at the same time.
The backbone is a ViT-G/8 — 48 blocks, model dimension 1664, 1.86 billion parameters — consuming 256×256×6-pixel Cell Painting crops at an 8×8 patch size. Pretraining used the standard masked autoencoder objective at a 75% mask ratio with an added Fourier-domain reconstruction term, for 500 epochs over more than 8 billion crops from Phenoprints-16M: 256 H100 GPUs for over a week, roughly 48,000 GPU-hours, with the Lion optimizer at learning rate 3e-5 and a global batch size of 8192.
Whole-genome benchmarks aggregate embeddings over 17,063 genes at six guides each in HUVEC cells from the RxRx3 screen. Trimmed Phenom-2 recalls 45.4% of known gene-gene relationships across CORUM, hu.MAP, Reactome, Signor and StringDB, against 44.4% for the previously published ViT-L/8 MAE trained on the uncurated RPI-93M corpus, and raises replicate-consistency statistics from 0.52 to 0.63 (Kolmogorov-Smirnov) and 12.3 to 18.2 (Cramér-von Mises). On external JUMP-CP data it leads on all four relationship databases tested, and on the RxRx3-core compound-gene benchmark it improves the z-score over a random baseline from 3.77 to 5.38. Linear separability of genetic perturbations improves about 60% over the earlier model's final-block features.
Phenom-2 targets phenotypic drug discovery groups running compound and CRISPR screens at industrial scale. Embedding a whole screen once yields a space in which perturbations cluster by mechanism of action, uncharacterized genes are assigned to complexes and pathways by their nearest neighbours, and a compound's likely target is nominated zero-shot from the cosine similarity between its embedding and those of gene knockouts. Because the encoder is frozen and every readout is linear, one pass over the images supports many analyses without retraining. The model itself is internal to Recursion, distributed neither as weights nor as an inference endpoint.
Phenom-2 extends the scaling relationship between training compute and biological relationship recall, first established for microscopy masked autoencoders at the hundred-million-parameter scale, into the multi-billion-parameter regime. It pairs that with two lessons that travel beyond microscopy: curating experimental training data by measured effect size beats collecting more of it, and the conventional final-layer readout is often the wrong place to take features from a self-supervised vision transformer applied to scientific data. The caveats are equally concrete. Neither the weights nor Phenoprints-16M is public; what Recursion released alongside the paper is the shared masked autoencoder architecture code and the far smaller OpenPhenom-S/16 checkpoint. And the whole-genome results measure recall against curated relationship databases, not prospective wet-lab validation.
Much of this page is generated or calculated automatically. Flag anything that looks off and we will re-run it.