Mol-JEPA: Molecular Foundation Model Learns from 14 Modalities Simultaneously
rbhar90 · x · 2026-08-26
Boehringer Ingelheim and collaborators released Mol-JEPA, a molecular foundation model using a multimodal Joint Embedding Predictive Architecture. It learns from 14 modalities including structure, cell painting, binding affinity, ADMET, and quantum chemistry. By using modality masking and a SIGReg regularizer to prevent representation collapse, Mol-JEPA addresses the underfitting issues of single-view models. It shows significant Out-of-Distribution (OOD) gains on small ADME datasets, according to the paper on arXiv.
More from Research
- New book released: A practical guide to post-training, SFT, and RL — Hesamation · 2026-08-26
- AgentFlow lets small models outperform large ones via specialized roles — TheZachMueller · 2026-08-26
- Yoav Goldberg: actionable interpretability research is the lesser form — yoavgo · 2026-08-26
- Are We Just Lucky That Transformers Work as Both LM and Agent? — MillionInt · 2026-08-26
- Oral history of Lynn Conway, Mead-Conway pioneer, published — kevinakwok · 2026-08-26
- Mixedbread on retrieval scaling laws: co-designing models and vector DBs — lateinteraction · 2026-08-26