Peptriever: Large-Scale Protein–Peptide Binding Search with a Transformer Bi-Encoder

Peptriever: Protein–Peptide Binding Search at Scale

I co-authored Peptriever, a peer-reviewed protein–peptide binding search engine (Bioinformatics 2024, 40(5), btae303) that ranks candidate bindings at large scale.

The framing problem

Protein binding was being studied one pair at a time: given this protein and that peptide, will they bind? Models answered that question well. But it isn’t the question drug discovery actually asks, which is what else would this bind to, and how selectively? Answering that by brute force means running a pairwise model across millions of combinations - which doesn’t scale.

Borrowing a pattern from search

Search engines and recommender systems hit the same wall years earlier, and solved it in two stages: narrow millions of candidates down to a shortlist cheaply, then score only the shortlist in detail. Protein binding already had the second stage. The first one was missing, and nobody was building it.

Peptriever is that missing stage - a bi-encoder: two transformer towers that embed proteins and peptides into a shared vector space, so finding candidates becomes a nearest-neighbour lookup instead of millions of pairwise predictions. It was trained on a home gaming machine.

Results

The evaluation results demonstrate that our approach is competitive with the best models for this task, including the renowned AlphaFold-Multimer, while also enabling large-scale search, which the pairwise models could not do.

Key Achievements

• Performance Parity: Competitive with state-of-the-art models in binding prediction accuracy
• Scalability: Enabling search across databases of millions of proteins
• Speed: Reducing analysis time from months to minutes
• Accessibility: Making advanced protein analysis available to researchers worldwide

Why it matters

Searching at scale changes which questions are askable. Not just “do these two bind?” but “what else does this bind to, and how selectively?” - which is what you need when checking a candidate compound for off-target effects.

Open source

Training code and model weights are released under a permissive MIT license.

Access the Technology

Research Papers

• Published Paper - Peer-reviewed research in Bioinformatics
• Preprint - Early access to our findings

Open Source Implementation

• Codebase - Complete implementation on GitHub
• Model Weights - Pre-trained models on Hugging Face

Live Demo

• Official App Page - Production-ready protein binding search

Interactive Demo

Experience the power of AI-powered protein binding search firsthand:

---

Peptriever was independent research. Connect on LinkedIn if you’d like to discuss it.

Publications: Google Scholar · ORCID.