moPPIt

moPPIt (motif-specific Protein-Protein Interaction targeting) is a de novo peptide binder design framework developed in the Pranam Chatterjee lab, then at Duke University. It addresses a persistent gap in protein engineering: designing binders that target a specific motif on a protein — such as a disordered segment or a conserved epitope — rather than a well-folded, structurally resolved pocket. Many disease-relevant targets, including intrinsically disordered regions, lack stable three-dimensional structure, which makes conventional structure-based binder design difficult or impossible.

The framework operates entirely from sequence, removing the requirement for a high-quality target structure. It pairs two components: BindEvaluator, a transformer that interpolates protein language model embeddings to predict binding-site residues, and a Multi-Objective-Guided Discrete Flow Matching generator that produces peptide binder sequences directed at the chosen motif. BindEvaluator reaches an AUC of 0.97 for binding-site prediction, providing the target signal that steers generation.

First posted to bioRxiv in 2024, moPPIt is notable for in-vitro validation across several biologically meaningful targets, demonstrating that sequence-only, motif-directed design can yield functional binders.

#Key Features

• Motif-specific targeting: Designs binders directed at a chosen sequence motif — disordered regions or conserved epitopes — rather than only structured pockets.
• Sequence-only operation: Requires no experimental or predicted target structure, enabling design against intrinsically disordered and otherwise intractable targets.
• High-accuracy site prediction: BindEvaluator, a transformer over interpolated protein language model embeddings, predicts binding-site residues at an AUC of 0.97.
• Multi-objective discrete flow matching: A guided discrete flow-matching generator balances multiple design objectives to produce de novo peptide binder sequences.
• Experimentally validated: In-vitro tests span NCAM1, the β-catenin disordered region, the GM-CSF receptor, and a CAR Treg AGR2t target.

#Technical Details

moPPIt decomposes binder design into prediction and generation. BindEvaluator is a transformer that interpolates embeddings from a protein language model to score which residues on a target are likely binding sites, achieving an AUC of 0.97. The generative stage uses Multi-Objective-Guided Discrete Flow Matching — a discrete generative process over amino-acid sequences steered by multiple objectives — to synthesize peptide binders aimed at the predicted or specified motif. Because both stages consume sequence rather than structure, the pipeline applies to disordered regions and conserved epitopes that structure-based methods struggle to address. The authors report in-vitro validation against NCAM1, the intrinsically disordered region of β-catenin, the GM-CSF receptor, and a CAR Treg AGR2t target.

#Applications

moPPIt is intended for protein engineers and therapeutic discovery teams who need binders against targets that resist structure-based design — particularly intrinsically disordered proteins and specific conserved epitopes implicated in disease. By working from sequence alone and accepting a user-specified motif, it lets researchers direct binder generation to a precise interaction surface, which is valuable for modulating protein-protein interactions, building cell-engineering reagents (as in the CAR Treg example), and prototyping peptide therapeutics prior to experimental screening.

#Impact

moPPIt extends de novo binder design into the large and therapeutically important space of disordered and motif-defined targets, where dominant structure-based approaches have limited reach. The combination of an accurate sequence-based binding-site predictor with a multi-objective discrete flow-matching generator, backed by in-vitro validation across several targets, makes it a notable contribution to the peptide and protein design literature. Code is available at programmablebio/moppit and trained checkpoints are hosted on Hugging Face (ChatterjeeLab/moPPIt), though access is gated behind academic, non-commercial terms that limit unrestricted reuse.