Loading…
Cell Painting microscopy foundation model, a channel-agnostic masked autoencoder producing morphological embeddings for zero-shot phenotypic analysis.
OpenPhenom-S/16 is a publicly released foundation model for high-content microscopy developed by Recursion. It applies a Channel-Agnostic Masked Autoencoder (CA-MAE) architecture to Cell Painting images, generating compact morphological embeddings that capture the phenotypic state of cells without requiring any labeled training data. The underlying research was presented as a spotlight paper at CVPR 2024.
The model reached researchers outside Recursion twice, under two names. It first appeared as Phenom-Beta on 9 January 2024, hosted on NVIDIA's BioNeMo cloud APIs — the first third-party model that platform carried — with access gated behind a BioNeMo beta application and Recursion's non-commercial terms. In November 2024 Recursion published downloadable weights as OpenPhenom-S/16 on HuggingFace and Google Cloud Vertex AI Model Garden — weights the repository's own commit history names phenom-beta.
Both releases answer the same challenge in phenomics: microscopy datasets are acquired under varied experimental conditions with different fluorescence channel configurations, making it difficult to train a single model that generalizes across assays. Conventional vision models stack channels as fixed-depth tensors, requiring a consistent channel count at inference. OpenPhenom-S/16 instead processes each fluorescence channel independently through patch tokenization and then fuses information across channels via cross-attention, enabling inference on images with any number or ordering of channels. It is the publicly accessible member of Recursion's broader Phenom family, whose proprietary larger models — Phenom-1 and Phenom-2 — are trained on internal datasets of tens of millions of wells.
OpenPhenom-S/16 is built on a Vision Transformer Small backbone with 16x16 pixel patch size (ViT-S/16). The embedding encoder — the part most users run — totals approximately 22 million parameters; the full released CA-MAE checkpoint is roughly 178 million, because it also ships the six channel decoders the model card offers for predicting new channels or stains. The key architectural innovation is channelwise cross-attention: rather than stacking fluorescence channels into a single multi-channel input tensor, the model processes each channel's patch tokens independently and then applies cross-attention across channels to build a contextualized representation. Input images are 256x256 pixels in uint8 format; each image produces a single 384-dimensional embedding.
The released weights were pretrained on three publicly accessible Cell Painting collections: RxRx3 (Recursion's public high-content screening dataset with six fluorescence channels), JUMP-CP overexpression, and JUMP-CP gene-knockouts (the Joint Undertaking for Morphological Profiling datasets, acquired across multiple laboratories under varied conditions). Their differing channel configurations directly motivated the channel-agnostic design. The earlier checkpoint served on BioNeMo was trained on RxRx3 alone — roughly 2.2 million images of HUVEC cells spanning about 17,000 genetic knockouts and 1,674 chemical entities — on Recursion's BioHive-1 supercomputer, built on the NVIDIA DGX SuperPOD reference architecture; the openly published checkpoint adds the two JUMP-CP datasets and a deeper decoder. Benchmarks reported at CVPR 2024 show that ViT-based masked autoencoders outperform weakly supervised classifiers by up to 11.5% relative improvement in recalling known biological relationships curated from public interaction databases, with CA-MAEs generalizing effectively to held-out JUMP-CP conditions.
OpenPhenom-S/16 targets researchers in high-content screening and phenomics who need general-purpose morphological representations. Use cases include profiling compound or genetic perturbation screens to cluster agents by phenotypic similarity and identify mechanism-of-action groups; compound-gene interaction prediction using embedding cosine similarity for zero-shot target identification; and cross-assay transfer to images acquired with different microscopes, staining conditions, or channel configurations.
OpenPhenom-S/16 is one of the first openly released foundation models specifically designed for Cell Painting microscopy, filling a gap between proprietary pharmaceutical-scale models and general-purpose computer vision. Its CVPR 2024 spotlight recognition indicates peer validation of the channel-agnostic masked autoencoder approach, and its earlier life as Phenom-Beta showed that an industrial phenomics model could be distributed through a third-party inference platform — a route the open weights have since replaced. The model demonstrates a scalable pretraining paradigm: performance improves predictably with both model size and dataset scale, as validated by Recursion's internal Phenom-1 (ViT-L/8, 3.5 billion image crops) and Phenom-2 models. Limitations include a non-commercial-only license restricting industrial use, a model card that places brightfield and H&E imaging out of scope, and reduced representational capacity relative to Recursion's proprietary larger models. Image preprocessing — illumination correction, channel normalization, and resizing to 256x256 — remains the user's responsibility.
Much of this page is generated or calculated automatically. Flag anything that looks off and we will re-run it.