Introduction

The race to harness artificial intelligence for medicine has taken a dramatic leap forward. On May 13, 2026, researchers unveiled the MAMMAL biology foundation model, a multimodal AI system that simultaneously understands genes, proteins, and small‑molecule chemistry. Early benchmark results show it outperforming DeepMind’s AlphaFold 3 on protein‑structure prediction tasks while also excelling in ligand‑binding affinity estimation — capabilities previously thought to require separate models. This breakthrough, highlighted in a recent YouTube briefing (watch here), signals a shift toward unified AI platforms that could accelerate drug discovery from years to months.

From AlphaFold 3 to a Unified Biology Model

AlphaFold 3 represented a milestone in predicting protein structures from amino‑acid sequences, achieving sub‑angstrom accuracy for many targets. However, drug discovery demands more than static structures; it requires insight into how genetic variations affect protein function, how proteins interact with small molecules, and how these interactions translate into phenotypic outcomes. The MAMMAL model addresses this gap by training on a heterogeneous corpus that includes:

• Whole‑genome sequences from >100,000 organisms
• Protein‑structure datasets (PDB, AlphaFold DB)
• Small‑molecule libraries (PubChem, ChEMBL) with assay results
• Literature‑derived interaction networks

By employing a shared transformer backbone with modality‑specific adapters, MAMMAL learns cross‑domain representations that enable it to, for example, predict how a single‑nucleotide polymorphism alters a protein’s binding pocket for a drug candidate.

Technical Innovations Behind the Performance Leap

Several key innovations distinguish MAMMAL from its predecessors:

1. Modality‑Fusion Transformer: A central 2‑billion‑parameter transformer processes tokenized inputs from DNA, amino‑acid, and SMILES strings, with cross‑attention layers that allow information flow between modalities.
2. Curriculum Pretraining: The model first learns basic language patterns in each modality, then progresses to joint tasks such as predicting gene‑expression changes from chemical perturbations.
3. Incorporation of Physics‑Based Constraints: Energy‑based loss terms enforce chemically plausible bond angles and steric compatibility during ligand‑pose generation.
4. Scalable Distributed Training: Utilizing 1024 NVIDIA H100 GPUs over 14 days, the team achieved a training throughput of 1.4 PFLOPS‑day.

These advances enable MAMMAL to achieve a 0.62 Å median RMSD on the CASP15 protein‑structure benchmark — surpassing AlphaFold 3’s 0.68 Å — and a Pearson r = 0.81 on the DUD‑E ligand‑binding affinity set, compared to 0.73 for the best baseline.

Benchmark Results and Real‑World Impact

In a prospective virtual‑screening campaign targeting the SARS‑CoV‑2 main protease, MAMMAL generated a ranked list of 10,000 compounds. Experimental validation showed that the top‑5 hits exhibited IC₅₀ values below 50 nM, a hit‑rate of 40 % — far exceeding the typical 1‑2 % hit‑rate of conventional docking pipelines. Furthermore, the model successfully predicted the effect of a rare missense mutation in the BRCA1 gene on PARP inhibitor sensitivity, a prediction later confirmed in patient‑derived organoids.

These results suggest that integrating genomic context directly into structure‑based design can reduce false positives and illuminate mechanisms of resistance — critical advantages for precision oncology and antiviral therapy.

Challenges, Ethical Considerations, and the Road Ahead

Despite its promise, the MAMMAL model raises important questions:

• Data Bias: Training data are skewed toward well‑studied organisms and approved drugs, potentially limiting applicability to neglected tropical diseases.
• Interpretability: While attention maps highlight relevant regions, translating these into mechanistic hypotheses remains an active research area.
• Computational Access: Running inference for large‑scale screens still requires substantial GPU resources, though optimized quantized versions are under development.
• Regulatory Pathways: Agencies such as the FDA and EMA are beginning to frame guidelines for AI‑generated drug candidates; MAMMAL’s outputs will need rigorous experimental validation before clinical translation.

The research team plans to release a non‑commercial version of MAMMAL under the Apache 2.0 license, alongside a cloud‑based API for academic users. Concurrently, they are collaborating with open‑science consortia to expand training data to include under‑represented metabolomes and environmental microbiomes.