IBM MAMMAL Beats AlphaFold 3 on Antibody Binding and Predicts Cancer Drug
IBM Research's MAMMAL (Molecular Aligned Multimodal Architecture and Language) is the first foundation model trained simultaneously on chemistry, genetics, and proteins. It beat AlphaFold 3 on 5 of 7 antibody-binding targets and improved CDR-H3 antibody sequence generation by 19%. In a wet-lab validation against 805 solid-tumor cancer cell types, MAMMAL correctly ranked carfilzomib—an existing blood-cancer drug long overlooked for solid tumors—as the top candidate, with 95% rank preservation in physical experiment.
Why It Matters
Multimodal biology foundation models can now repurpose existing approved drugs for new indications via sequence inference alone—compressing the 10-15-year, $1B-per-drug discovery pipeline dramatically.