Genejepa

GeneJEPA: a self-supervised foundation model for single-cell RNA-seq that learns to predict representations of masked gene sets from visible context (Joint-Embedding Predictive Architecture). Perceiver-style encoder over unordered gene sets, continuous-value tokenizer (gene identity + Fourier features for expression), EMA teacher targets with variance-covariance regularization. Trained on the Tahoe-100M atlas. Input: single-cell gene expression (gene sets). Output: transferable cell embeddings (annotation, drug response, perturbation reasoning). Supports test-time scaling by rereading more genes with fixed latent compute.

Use in DarwinOmics

How to use it

Use this model when the submitted analysis depends on embeddings, predictions, scoring, generation, or model-backed comparison. Model output is treated as a reproducible analysis step, not as independent validation of a biological claim.