About the Platform

Research Motivation

Accurate prediction of MHC peptide binding is foundational to rational vaccine design, neoantigen discovery, and T-cell epitope mapping. Wet-lab binding assays are slow and expensive, and existing computational predictors often generalize poorly to under-represented alleles or peptide lengths outside the canonical range. This platform was developed to provide fast, sequence-only binding predictions — across both MHC Class I and Class II — for the peptide lengths researchers actually need to screen when designing vaccine candidates.

BioAnchorFormer is a compact transformer encoder over peptide sequences, covering the full 8–11-mer range typically required for MHC Class I epitope screening. Each residue is embedded and combined with a positional encoding, then processed through self-attention layers before a feed-forward scoring head produces a single binding probability. The architecture is intentionally lightweight so that inference over thousands of candidate peptides — generated from a single input protein — remains fast enough for interactive use.

MHC Class I binding is dominated by a small number of anchor positions — most commonly position 2 (P2) and the C-terminal residue (PΩ) — that sit deep in the peptide-binding groove. BioAnchorFormer’s attention mechanism is designed to let the model learn to emphasize these anchor positions relative to the rest of the peptide, rather than treating every residue with equal weight, mirroring the structural biology of the MHC-I binding groove.

The MHC-II Predictor is a BiLSTM-based sequence model covering the 13–17-mer range typically required for MHC Class II epitope screening. Unlike BioAnchorFormer, it is a multi-label model rather than a single binary classifier: BLB1-binding and BLB2-binding are independent biochemical events — a peptide can bind neither, either, or both — so the model produces two independent sigmoid probabilities, each compared against its own tuned decision threshold, rather than one 4-way softmax.

Those two independent yes/no decisions are combined into one of four display labels, with all underlying probabilities always shown alongside the label:

Proteome Scan is not a separate model — it runs BioAnchorFormer, the MHC-II Predictor, or both across every protein in a multi-FASTA file in one batch job, then surfaces the best individual peptides overall and which proteins are richest in epitopes. This is intended for screening a whole pathogen proteome at once, rather than one protein at a time.

Scientific Contributions

• An anchor-aware attention formulation tailored to MHC-I peptide binding.
• A unified BioAnchorFormer model covering the full 8–11-mer peptide length range typically required for MHC-I epitope screening.
• A multi-label MHC-II Predictor modeling BLB1 and BLB2 binding as independent events, rather than forcing them into mutually exclusive classes.
• An open, research-grade platform — including whole-proteome batch screening — enabling reproducible sequence-to-prediction workflows for vaccine and immunology labs.