NeoGen-BC

Peptide-HLA binding predictors answer one question well: will this peptide sit in an MHC groove? A peptide can bind tightly and still provoke no T cell, and the converse bites harder in breast cancer. Several MHC class II epitopes carrying the ESR1 Y537S, ESR1 D538G and PIK3CA H1047R substitutions have been shown experimentally to raise CD4+ responses, yet NetMHCIIpan 4.3 places them outside the 10% rank threshold that conventional pipelines use as their cutoff. Screening on affinity alone therefore discards the very epitopes an off-the-shelf vaccine would want.

NeoGen-BC, from Yuan Ze University with clinical collaborators at Taipei Veterans General Hospital, is built around that gap. It ranks peptides by intrinsic immunogenicity rather than by predicted binding, representing each one with frozen embeddings from the ProtTrans ProtT5 encoder and pulling that representation toward known immunogenic epitopes through a retrieval step before a compact convolutional-recurrent classifier decides. A second, generative leg then steers ProtGPT2 to write new peptides in the same neighbourhood, so the framework proposes candidates as well as scoring them.

The design target is deliberately narrow and deliberately public. Most neoantigen pipelines chase patient-private mutations, which is inherently unscalable; NeoGen-BC restricts itself to recurrent driver mutations in ESR1 and PIK3CA, whose neoepitopes are shared across many hormone receptor-positive and triple-negative tumours and are therefore addressable by a single vaccine construct.

#Key Features

• Immunogenicity as the objective: The classifier is trained on epitopes labelled by observed immune response, not by measured or predicted affinity, which is what lets it recover epitopes that binding tools rank as non-binders.
• Retrieval-augmented peptide embeddings: For each query peptide the five nearest sequences in a curated epitope database are located by L2 distance in embedding space, averaged, and blended into the query representation before classification.
• Multi-window convolution with a recurrent read-out: Parallel convolutional branches with different kernel widths each feed a bidirectional LSTM, so short motifs and longer-range residue dependencies are summarised separately and then concatenated.
• Sliding-window mutation scanning: Mutant protein sequences are tiled into overlapping windows, which enumerates every candidate MHC-II epitope spanning a driver substitution rather than relying on a predictor's own core-offset call.
• Generation followed by staged filtering: ProtGPT2 output is narrowed by perplexity and biophysical profiling, then by MHC-II binding checks, then by structural validation of the peptide-HLA complex with AlphaFold-Multimer.

#Technical Details

ProtT5-XL-U50 supplies 1024-dimensional per-residue embeddings; peptides are padded or truncated to 35 residues, giving a 35 × 1024 input tensor per candidate. The retrieval database ships with the code as two FASTA files of immunogenic and non-immunogenic epitopes. The classifier is a multi-window CNN feeding bidirectional LSTMs: three parallel 1D convolutional branches of 64 filters with kernel widths 2, 4 and 6, each followed by a 64-unit BiLSTM whose final states are concatenated into a dropout-regularised dense layer and a two-way softmax, trained with Adam under categorical cross-entropy.

The training corpus began at 6,516 neoantigen epitopes and 377,990 other peptides, reduced by sequence-similarity clustering at 0.4 and a 35-residue length cap to 558 and 2,555 sequences respectively. An 80/20 split yields 2,490 training and 623 independent-test peptides, of which 112 are neoantigens. On that independent test set the model reaches an AUC of 0.9053.

#Applications

The framework is aimed at off-the-shelf therapeutic vaccine design for breast cancer: given a recurrent driver mutation, it enumerates and ranks the MHC class II epitopes around it, then generates additional synthetic peptides matched to validated neoantigens on physicochemical and structural properties. Because CD4+ help is the axis it targets, it is most relevant to groups working on immunologically cold HR+ and triple-negative disease, including combination strategies that raise MHC-II expression pharmacologically. The scoring leg is also usable as a standalone immunogenicity filter downstream of any mutation caller.

#Impact

NeoGen-BC is a demonstration that a frozen protein language model plus retrieval can recover class II neoantigens that dedicated binding predictors miss, and that a general-purpose generative protein model can be constrained to propose plausible replacements for them. The limits are worth stating plainly. The positive training set is small — 558 epitopes after redundancy reduction — and the scope is three substitutions in two genes, so transfer to other tumour types is untested. Code, training data and the retrieval database are offered under the Academic Free License 3.0, a grant stated only in the README prose with no LICENSE file in the repository, and no trained classifier weights are distributed — the repository ships the training notebook and the splits instead. Every result, including the properties of the generated peptides, is computational; none of the de novo designs has been tested for T-cell reactivity.