A biological foundation model is pretrained to represent biological entities, processes, or data across one or more modalities. The goal is a unified model that relates DNA sequences, protein structures, cellular phenotypes, tissue organization, and organism-level physiology. section 13.4.4
Multimodal pretraining can align representations of the same gene across DNA sequence, RNA expression, and protein structure. Knowledge learned from one modality can then inform predictions in another, supporting transfer across biological tasks. section 13.4.4
Examples span ESM-2 for proteins, MegaMolBART for small molecules, Phenom for cellular imaging, LucaOne for genes, ProCyon across five biological knowledge domains, and Evo 2 across genome-scale sequences. Applications include protein function prediction, drug-target interaction prediction, clinical outcome prediction, variant-effect prediction, and genetic-circuit engineering. section 13.4.4
