OmicsFM brings proteomics into the foundation model era
While foundation models have been shown to learn biological representations from large transcriptomic atlases, it remained unknown whether proteomics data allow the same. We here therefore introduce OmicsFM, a modality-agnostic transformer pretrained through masked abundance reconstruction on an unprecedented proteomics data corpus of 48,837 quality-filtered proteomics profiles from 1,397 reprocessed PRIDE projects. Interestingly, despite training on 14- to 93-fold fewer profiles than matched bulk- and single-cell transcriptomic models, respectively, our proteomics model rivals both. On held-out projects, OmicsFM attention networks recovered more molecular relationships than co-expression methods and existing single-cell foundation models across nine reference databases that reveal pathway-level organization. Sample-level embeddings preserved biological structure across independent studies, and its representations transferred successfully to cell-type classification, gene-essentiality prediction, and perturbation-response prediction, while consistently outperforming task-specific models. Moreover, our results show that proteomics and transcriptomics representations capture complementary biology. OmicsFM thus firmly establishes the possibility of training highly performant proteomics-based foundation models, and their importance in modelling and uncovering fundamental biology.