A foundational model is broadly pretrained on massive unlabeled data, then reused or adapted for multiple downstream tasks through transferable representations. In drug discovery, this approach reduces dependence on large task-specific labeled datasets. section 13.2.4
Foundational models support several biological data types and reuse patterns:
- ESM-2 learns protein representations from UniProt sequences that support secondary-structure prediction, localization, protein interactions, and variant-effect prediction. section 13.2.4
- MegaMolBART encodes small-molecule structure into embeddings that can correlate with properties such as lipophilicity without explicit property labels. section 13.2.4
- Multimodal systems such as LucaOne and ProCyon align information across sequences, structures, expression, literature, knowledge graphs, and annotations. section 13.4.4
Variant-effect prediction with ESM-2
Training a variant-effect predictor from scratch may require thousands of experimentally measured variants. Fine-tuning ESM-2 instead starts from protein principles learned across millions of sequences, so fewer task-specific labels are needed. section 13.2.4
Foundational models are components of multimodal drug-discovery platforms rather than complete replacements for experiments or experts. Their value depends on orchestrating predictions across biological scales, validating results experimentally, and maintaining human oversight. section 13.5





























