MAMMAL AI outperforms AlphaFold3
- IBM researchers published MAMMAL in Nature this week — a multimodal drug-discovery model that links proteins, antibodies, small molecules, and gene-expression data in one system. - The headline comparison is narrower than “beats AlphaFold 3” — on an antibody-antigen benchmark, fine-tuned MAMMAL beat AlphaFold 3 confidence scores on 5 of 7 targets. - That matters because drug discovery is fragmented today; a single model spanning target biology, molecules, and assays could make early screening faster.
Protein AI just got a new contender — but the real story is not that AlphaFold 3 got “beaten” across biology. It’s that IBM researchers built a broader drug-discovery model, called MAMMAL, and showed that on one important antibody-binding task it did better than AlphaFold 3’s confidence scores. That result landed in Nature this week. The stakes are obvious: if one model can move across proteins, molecules, and cell readouts, it could connect steps that are still handled by separate tools today. (nature.com) ### What is MAMMAL, exactly? MAMMAL stands for Molecular Aligned Multi-Modal Architecture and Language. Basically, it is a foundation model for drug discovery rather than a pure structure predictor. It was pretrained on about 2 billion samples spanning proteins, antibodies, small molecules, and omics data, then tested on tasks like property prediction, generation, and transcriptomic response. In the paper’s main result set, it reached state-of-the-art(nature.com) ### So did it really beat AlphaFold 3? Yes — but in a specific, limited way. The paper says fine-tuned MAMMAL prediction scores significantly outperformed AlphaFold 3 confidence scores, used as a proxy for binding likelihood, in 5 of 7 antigen targets in an antibody-antigen benchmark. That is not the same as showing MAMMAL is a better general structure model than AlphaFold 3. It is a head-to-head on a narrower classification-style binding problem. (n([nature.com)## Why is that narrower than it sounds? AlphaFold 3 was built to predict structures and interactions across proteins, DNA, RNA, ligands, ions, and chemical modifications. DeepMind framed the big jump as molecular interaction modeling, with at least a 50% improvement over existing methods for protein interactions with other molecule types, and even larger gains in some categories. MAMMAL is aiming at a different bottleneck — stitching multiple biomedi(nature.com)covery questions. (blog.google) ### Why would a multimodal model help drug discovery? Because early drug work is messy. One team studies a target protein. Another screens molecules. Another looks at antibody binding. Another checks what happens to gene expression in cells. Those steps are linked in the real world, but the software stack is usually fragmented. MAMMAL’s pitch is that one aligned model can move across those representations instead of handing work off between disconnected specialist tools. (arxiv.org) ### What was the strongest detail in the paper? Two numbers matter. First, the scale — roughly 2 billion training samples across modalities. Second, the benchmark spread — 11 downstream tasks, with 9 new state-of-the-art results. The AlphaFold 3 comparison is the attention-grabber, but the bigger claim is that one unified architecture can stay competitive across a lot of the pipeline without being rebuilt for every task. (nature.com)tch is that “outperforms AlphaFold 3” is easy to oversell. The paper compares MAMMAL against AlphaFold 3 confidence scores as a binding-likelihood proxy in one benchmark family, not against AlphaFold 3 as a universal biology engine. And because this is a broad foundation-model claim, outside groups will need to reproduce the gains and test whether they hold up on real industrial datasets, not just curated benchmarks. The go(nature.com)icly, which makes that verification possible. (nature.com) ### Why does this matter now? The field is shifting from single-purpose biology models to systems that can reason across modalities. AlphaFold 3 pushed structure prediction into richer molecular interactions. MAMMAL is pushing in a parallel direction — from prediction toward an integrated discovery stack. If that works, the win is not just better folds. It is fewer handoffs between tools, faster triage of candidates, and a shorter path from biological question to testable drug idea. (blog.google) ### Bottom line? MAMMAL did not dethrone AlphaFold 3 across the board. But it did post a credible, specific win in antibody-antigen binding and made a bigger argument: the next useful biology AIs may be the ones that connect structure, chemistry, and cell biology in one place. (nature.com)
