Biotech Israel United States AI Pharma Healthcare Analysis Assay Validation Drug Discovery AI Infrastructure Open Models
May 2026

Negative Labels Beat Pose Proxies

MAMMAL suggests the scarce layer in AI drug discovery may be task-aligned biological supervision, not another larger structure engine alone.

MAMMAL shifts the strategic bottleneck from owning the biggest molecular model to owning the cross-modal labels, negative examples, and assay loops that teach models which biomedical predictions matter.

MAMMAL, a multimodal biomedical foundation model from IBM Research-Israel, IBM TJ Watson, and Technion, is introduced in an npj Drug Discovery paper as a 0.5B-parameter model trained on 2 billion samples spanning proteins, antibodies, small molecules, interactions, and gene-expression profiles [1].

The tension is whether scarce advantage in AI drug discovery sits in ever-larger structure engines, or in the negative labels and assay-aligned prompts that teach a model what biological failure looks like.

2B
Pretraining Samples
MAMMAL was pretrained on public datasets spanning protein and antibody sequences, small molecules, interactions, and single-cell gene-expression data.
0.5B
Model Parameters
The paper introduces MAMMAL as a 0.5B-parameter multimodal biomedical foundation model.
5/7
MAMMAL AF3 Wins
MAMMAL outperformed AF3-derived confidence-score proxies on five of seven antibody or nanobody binding targets.
UNLOCK ACCESS

Want to read the full Signal?

Sign up to access the complete analysis for Negative Labels Beat Pose Proxies.