In my last snippet, I speculated that AI tools recently published in Science and Nature (the “logos” method and an RFdiffusion-based method) could be used to design a protein targeting the intrinsically disordered and toxic N-terminal region of the prion protein. Yet within a week, Nature Biotechnology published a tool that might be even better suited to the task. PepMLM (“Peptide binder design algorithm via Masked Language Modeling”) is a protein language model that designs peptides that will bind to a target protein. PepMLM leverages amino acid sequence only, with no structural information (by contrast, RFdiffusion and logos operate in 3D structural space), potentially making it better able to handle conformational disorder. In addition, peptides have several advantages over larger proteins, including more ‘drug-like’ properties.
The researchers showed that PepMLM-generated peptides bind to disease-relevant targets and, when fused to E3 ubiquitin ligases, degrade Huntington’s disease-driving proteins in a cellular model. The team is now extending the model to (1) account for post-translational modifications in the target, (2) enable motif-specific binding, (3) avoid off-target effects, and (4) optimize therapeutic properties. It’s a fast-moving field that promises to yield binders for traditionally “undruggable” targets.
Target sequence-conditioned design of peptide binders using masked language modeling
In Nature Biotechnology, 13 August 2025
From the group of Pranam Chatterjee at the University of Pennsylvania
Snippet by Katrina Woolcock
Image credit: Figure 3 from Chen et al. cited above (CC BY-NC-ND 4.0).