MAMMAL

MAMMAL (Molecular Aligned Multi-Modal Architecture and Language) is a foundation model for cross-modal learning in drug discovery, described in a paper in npj Drug Discovery. It represents small molecules, proteins, antibodies and gene-expression profiles in a unified sequence framework, using a structured prompt syntax and an architecture that embeds numerical values through continuous projections. It was pre-trained on about 2 billion samples covering denoising, infilling and binding-related tasks.

The model supports classification, regression and generative tasks on cross-modal inputs. According to the paper, fine-tuned MAMMAL reaches state-of-the-art results on nine of eleven drug discovery benchmarks and competitive results on the other two. Tasks include antibody-antigen binding, TCR-epitope binding, targeted antibody design (CDR infilling), toxicity and cell-type annotation. In an antibody-antigen binder versus non-binder benchmark, its scores significantly outperformed AlphaFold 3 confidence scores (used as a proxy) on five of seven antigens. AF3 evaluation used a downsampled test set because of its computational cost.

The authors state the framework and pretrained models are publicly available for research. The provided material gives no repository, model hub link or license, so the exact access route and terms are unconfirmed here. It is aimed at computational biology and pharmaceutical researchers who fine-tune it for specific tasks. Results are benchmark-based, and task-specific fine-tuning is required.