BLMPred

Antibodies recognize antigens at epitopes, and the linear (continuous) variety — a short run of adjacent residues along the antigen's primary sequence — is the form most often exploited in serological diagnostics, peptide vaccines, and antibody development. Because experimental epitope mapping is slow and costly, computational triage has become a standard first pass. Predictors fall into two groups: tools that scan an entire antigen sequence and highlight epitopic regions, and tools trained on exact epitope peptides that answer a narrower question — is this specific peptide an antibody epitope or not? The two are complementary, but the first group is known to degrade on short peptides, precisely the regime in which candidate lists for synthesis are drawn up.

BLMPred, developed by Barnali Das and Dmitrij Frishman at the Technical University of Munich, sits in the second group. It is a binary classifier over peptide sequences: supply a peptide and it returns an epitope or non-epitope call, with no 3D structure, no multiple sequence alignment, and no taxonomic annotation required. The defining design choice is a division of labour — the classifier is a conventional support vector machine, while all representational work is delegated to a pretrained protein language model, ProtTrans, whose embeddings encode contextual, evolutionary, and structural signal that handcrafted physicochemical descriptors cannot capture.

The work was posted to bioRxiv in September 2025 and published in Computational and Structural Biotechnology Journal in December 2025. Its contribution lies less in architecture than in data curation at scale combined with modern sequence representations, and it complements structure-based conformational predictors such as the patch-centric conformational B-cell epitope predictor and taxon-adapted approaches such as EpitopeTransfer.

#Key Features

• Structure-free peptide classification: Predictions require only the peptide's amino acid sequence, avoiding the dependency on experimental or predicted 3D structures that constrains conformational epitope tools.
• Language model embeddings as the feature space: Each peptide is reduced to a single 1024-dimensional averaged ProtT5-XL-U50 embedding, replacing propensity scales and composition descriptors with a learned, context-aware representation.
• Two length-matched models: BLMPred_5–60 covers peptides of 5–60 residues, while BLMPred_8–25 targets the 8–25 residue window set by the length range of antibody complementarity-determining regions.
• Length-matched negative sampling: Negatives were drawn so their length distribution mirrors the positives, removing peptide length as a shortcut the classifier could otherwise exploit.
• Time-split benchmarking: Comparison against other peptide-level tools uses IEDB entries deposited after the training snapshot, so evaluation peptides post-date training.

#Technical Details

Training data came from the Immune Epitope Database (March 2023 release): 208,265 experimentally validated linear B-cell epitopes and 487,127 non-epitopes, filtered to remove duplicates, non-standard amino acids, peptides under five residues, and sequences appearing in both classes. Host proteins were clustered with CD-HIT at 80% identity and near-identical peptides removed, yielding 111,015 positives for the 5–60 range and 102,023 for the 8–25 range; class-balanced sets of 222,030 and 204,046 peptides were split 90/10 into training and held-out test data. Twelve scikit-learn classifiers plus XGBoost and an explainable boosting machine were compared under 10-fold cross-validation with GPU-accelerated training via RAPIDS; the support vector machine won on Matthews correlation coefficient.

On its held-out test set, BLMPred_5–60 reaches accuracy 0.846, precision 0.859, recall 0.829, F1 0.844, specificity 0.864, MCC 0.693, AUROC 0.846, and AUPRC 0.798; BLMPred_8–25 performs comparably at accuracy 0.835 and MCC 0.671. Under stricter homology reduction (CD-HIT at 70% on a November 2024 IEDB snapshot) performance drops to accuracy 0.754 and MCC 0.553, with recall falling to 0.555 against specificity 0.952. On the time-split benchmark of 2,928 epitopes and 1,000 non-epitopes, BLMPred achieves F1 0.80 versus 0.16 for epitope1D, 0.43 for LBEEP, and 0.33 for SVMTriP, though epitope1D reports a higher AUROC in its own publication (0.93 versus 0.83–0.84 here) on the strength of taxonomic and ontology features.

#Applications

BLMPred is aimed at researchers who already have candidate peptides in hand and need to prioritize which to synthesize and test. Typical uses include screening conserved blocks from multiple sequence alignments of a pathogen proteome, filtering peptide libraries for serodiagnostic assay development, and shortlisting immunogens for subunit vaccine constructs and antibody-generation campaigns. Because it accepts peptides rather than whole proteins, it pairs naturally with region-level scanners: those tools propose epitopic stretches, and BLMPred adjudicates the resulting peptides.

#Impact

BLMPred demonstrates that a large, carefully deduplicated training corpus paired with frozen protein language model embeddings can outperform purpose-built feature engineering on a long-standing immunoinformatics task, using a classifier that predates deep learning entirely. Several limitations bound it: the model returns a binary call with no residue-level localization, it cannot address conformational epitopes, its accuracy falls substantially under strict homology reduction, and the AUROC gap to epitope1D indicates weaker ranking quality than its thresholded classification suggests. Code, datasets, and training scripts are on GitHub, with the two trained SVM models and precomputed embeddings distributed via Google Drive; the repository carries no license file, so reuse terms for code and weights are unstated.