epiGPTope

Epitopes are the short antigenic peptide stretches that antibodies, T-cell receptors, and MHC molecules actually recognize, and they sit at the center of vaccine design, immunodiagnostics, and antibody-based therapeutics. Building synthetic epitope libraries is limited by combinatorics: a linear epitope of n residues has 20^n possible sequences, a space far larger than phage display or droplet microfluidics screening can cover. Most computational work in this area has been predictive — scoring whether a given sequence looks epitope-like — rather than generative, leaving the problem of proposing candidates in the first place unaddressed.

epiGPTope approaches the problem from the generative side. It is an autoregressive protein language model fine-tuned from ProtGPT2 on curated linear epitopes from the Immune Epitope Database (IEDB), so that new epitope-like sequences can be sampled directly from the learned distribution rather than mined from natural proteomes. Because the resulting library is generic, the authors pair the generator with a family of binary classifiers that separate epitopes from non-epitopes within bacterial and viral subsets of IEDB, tailoring a general-purpose library toward a specific experimental context.

The model was developed by Multiverse Computing together with CIC biomaGUNE, posted to arXiv in September 2025 and published in ACS Synthetic Biology in February 2026. The models are not distributed as open weights: they are reachable through Multiverse Computing's Singularity platform, with access granted on request, and the method is the subject of a European patent application on which all authors declare a competing financial interest.

#Key Features

• De novo epitope generation: Sampling from the fine-tuned model produced 192,222 distinct synthetic epitope sequences, a library scale that is impractical to assemble by experimental screening alone.
• Sequence-only design: The approach works from primary amino acid sequence, requiring neither antigen structures nor hand-crafted physicochemical descriptors.
• Origin- and assay-specific classifiers: A family of classifiers trained on eight IEDB subsets — bacterial or viral, crossed with T-cell, B-cell, MHC, and combined assays — filters generated candidates before synthesis.
• Filtering quantified by likelihood ratio: Classifier utility is reported as the positive likelihood ratio (LR+), which is independent of the class balance of the library being filtered and so translates directly into enrichment.
• Statistical fidelity to natural epitopes: Generated sequences reproduce known IEDB properties, including a length distribution peaking at 7-9 residues, aromatic enrichment at the final position, and depletion of cysteine.

#Technical Details

The base model is ProtGPT2, a GPT-2-style causal language model with 738 million parameters and a model dimensionality of 1280, pretrained on UniRef50 (release 2021_04). Fine-tuning used 504,611 deduplicated linear epitope sequences from IEDB, filtered to a human host, T-cell, B-cell, or MHC assays, and a maximum of 11 residues, trained with the same causal language modeling objective on a single NVIDIA A10G GPU at batch size 48. Five hyperparameter configurations were compared; evaluation loss reached its minimum around epochs 2-5 before overfitting. Generation settings were selected on perplexity: a repetition penalty of 2 gave a lower mean perplexity than the default 1.2 (5207.8 versus 5414.1, Mann-Whitney p = 8.7e-52), and a temperature of 1 was used for the final library.

For classification, XGBoost ensembles built on 1280-dimensional embeddings from ProtGPT2, ProtBERT (from ProtTrans), and a natural-language BERT baseline reached LR+ values of only 1.41 for bacterial and 1.13 for viral epitopes. Fine-tuned language-model classifiers performed substantially better: ProtBERT on viral B-cell data reached ROC AUC 0.918 and ProtGPT2 0.916, MHC-assay classifiers reached the highest F1 scores (0.869 and 0.852), and LR+ values reached 16.2 on bacterial T-cell data. Neither PCA nor UMAP on the embeddings produced visible clustering.

#Applications

The intended workflow is library design for wet-lab campaigns: generate a broad set of epitope-like candidates, then filter with a classifier matched to the organism class and assay format a group intends to run, cutting the number of peptides that must be synthesized and screened by phage display, peptide microarray, or droplet microfluidics. Immunology and vaccine groups, diagnostic developers, and biosensor engineers who work from sequence rather than antigen structure are the natural users, since nothing in the pipeline requires structural input.

#Impact

epiGPTope adds a generative component to an epitope-discovery toolbox that has been dominated by predictors such as BepiPred, epiBERTope, and EpitopeVec, and it shows that transfer learning from a general protein language model carries over to peptides far shorter than typical training proteins. The authors are candid about the limits: statistical features associated with epitopes also appear in IEDB negatives and may reflect curation bias rather than immunology, many negatives may be unlabeled epitopes for antibodies absent from the assay, and the classifiers score general epitope-likeness rather than binding to a specific antibody. All results reported are computational, with no experimental validation of the generated sequences, and the absence of released weights or code limits independent reproduction.