Loading…
Microscopy foundation model for high-content screening, embedding genome-scale CRISPR knockout and compound perturbations from Cell Painting images.
No providers recorded yet. Browse all providers
Turning a high-content screen into numbers has traditionally meant either a segmentation and feature-extraction pipeline such as CellProfiler, or a classifier trained to predict which perturbation was applied to each well, whose internal representation becomes the image feature. That second route — weakly supervised learning — wins on small, carefully curated screens with strong phenotypes, but it depends on perturbation labels and its representations stop improving past a few million training images.
Phenom-1 is the alternative that keeps scaling. Built by Recursion with Valence Labs and presented as a spotlight paper at CVPR 2024, it is a vision transformer trained as a masked autoencoder: 75% of the image patches are hidden and the network learns to reconstruct them, using no perturbation labels at any point. Its pretraining corpus, RPI-93M, is Recursion's internal set of 92,764,542 microscopy images spanning 3,957,400 perturbations. Phenom-1 is the first generation of the Phenom family, whose openly released member is the much smaller OpenPhenom-S/16 and whose billion-parameter successor is Phenom-2.
The encoder is used frozen. Crops go in, embeddings come out, and every downstream readout — gene-gene relationship recall, zero-shot compound-gene interaction ranking, morphology feature prediction — is a cosine similarity or a linear model over those embeddings, with batch effects removed by Typical Variation Normalization rather than by retraining.
The encoder is a ViT-L with 8x8 patches: 24 transformer blocks, roughly 304 million encoder parameters plus a 25 million parameter decoder, about 330 million in total. It consumes 256x256x6 crops with channel-wise self-standardization and emits a 1,024-dimensional embedding, the average of the final-layer patch embeddings. Pretraining ran for about 40 epochs, some 3.5 billion crops, on 128 A100 GPUs for over 20,000 GPU-hours, using the Lion optimizer with batches of up to 16,384 crops at a 75% mask ratio.
On RxRx3 CRISPR knockouts, recall of known relationships among the top and bottom 5% of cosine similarities reaches 0.622 on CORUM, 0.443 on hu.MAP, 0.267 on Reactome and 0.484 on StringDB after Typical Variation Normalization and chromosome-arm bias correction — an 11.5% relative gain on hu.MAP over the strongest weakly supervised model tested. At a 5% false positive rate on StringDB, recall is 0.80, 0.23 and 0.17 on the PoC-124, MoA-300 and DG-1640 gene sets. Applied to JUMP-CP, produced by different laboratories under a different channel structure, it retrieves 0.78 of perturbations after its stains are mapped onto the pretraining channel set and one channel is duplicated to fill the model's six expected inputs. Linear regression on its embeddings predicts 955 CellProfiler features with a median coefficient of determination of 0.737 on intensity features, against 0.297 for a weakly supervised DenseNet-161.
Phenom-1 was built to run Recursion's own map of biology: embed every well of a genome-scale CRISPR screen or a compound library, then read relationships out of the resulting space. Uncharacterized genes are assigned to complexes and pathways by their nearest neighbours, compounds group by mechanism of action, and a compound's likely target is nominated from its cosine similarity to gene-knockout embeddings, with no task-specific training. Because the encoder is frozen, one inference pass supports many such analyses. Recursion describes the model as in production for its internal teams and close partners; it is distributed neither as weights nor as an inference endpoint.
Phenom-1 established that the scaling behaviour of self-supervised vision models on natural images also holds for cellular microscopy, measured not by reconstruction quality but by how well the embeddings recover biology that was never in the training signal. That set the direction for the family: Phenom-2 extended it into the multi-billion-parameter regime, and OpenPhenom-S/16 carried the architecture, though not this checkpoint, into public hands. The limitations are equally clear. Neither the weights nor RPI-93M is available outside Recursion; what accompanies the paper is the shared masked-autoencoder architecture code. The 8x8 patching that helps accuracy makes inference costly, and the gains are recall against curated relationship databases rather than prospective experimental validation.
Much of this page is generated or calculated automatically. Flag anything that looks off and we will re-run it.