De novo design of peptide binders to conformationally diverse targets with contrastive language modeling
Problem
Designing molecules that bind to 'undruggable' disease proteins is a major challenge in drug discovery. PepPrCLIP was created to design short target-binding peptides using only the target protein's amino acid sequence.
Why it matters
Binders to undruggable disease proteins could enable new therapies for hard-to-treat conditions.
What the AI did
PepPrCLIP generates peptide candidates via Gaussian perturbation of the ESM-2 protein language model's latent space, then screens them with a CLIP-based contrastive architecture; top-ranked peptides were experimentally validated as inhibitors and E3-ligase fusions.
Results
Functionally potent binding and degradation of conformationally diverse disease targets stated; quantitative metrics not stated in source (abstract truncated).
Impact
AI-designed peptides were experimentally validated to bind and degrade disease targets.
Record
| Venue | Science Advances |
|---|---|
| Published | 2025 |
| Work type | article |
| Authors | Suhaas Bhat, Kalyan Palepu, Lauren Hong, Yiwei Mao, Tianzheng Ye, Rema Iyer, Lin Zhao, Tianlai Chen et al. |
| Citations | 64 (OpenAlex, at import time) |
| References | 49 |
| Model or method | PepPrCLIP (ESM-2 + CLIP contrastive learning) |
| Organization | not stated in source |
| Primary source | https://doi.org/10.1126/sciadv.adr8638 |
| Open access | gold |
| Keywords | peptide design, ESM-2, CLIP, protein binders, generative AI, drug discovery |
Importance rating
significant — thousands to millions affected
Targeting undruggable disease proteins is a formidable drug-discovery challenge; AI-designed peptides were experimentally validated to bind and degrade disease targets in vitro, but no scale stated.