DeepEpitope

DeepEpitope predicts linear B-cell epitopes in tumour antigens — the continuous stretches of amino acids that antibodies recognize, and the building blocks of peptide-based cancer vaccines. Its premise is a data argument rather than an architectural one. Established B-cell epitope predictors such as BepiPred and LBtope are trained on the Immune Epitope Database, a heterogeneous collection dominated by microbial and viral antigens. Tumour-associated antigens are largely self-proteins with distinct immunological and biochemical character, and the authors argue that general-purpose predictors underperform on them because no cancer-specific training corpus was used.

The tool was built by Karthick Vasudevan's group at the Institute of Bioinformatics in Bangalore with collaborators at REVA University, Manipal Academy of Higher Education, Taylor's University, and Universiti Malaysia Kelantan, and posted to bioRxiv in May 2025. It follows the now-common recipe of freezing a protein language model and training a light classifier on its embeddings: sequences are encoded by ESM-2 into 1,280-dimensional vectors, and a small feedforward network learns the epitope/non-epitope decision on top.

Within this catalog it sits alongside two other ESM-based epitope predictors that occupy different niches — the Patch-Centric Conformational B-Cell Epitope Predictor from Albert Einstein, which scores discontinuous surface patches on a 3D structure, and the Structure-Fused Immunogenic B-Epitope Predictor from TCS Research, which adds contact and protrusion features. DeepEpitope is linear-only, sequence-only, and cancer-specific.

#Key Features

• Trained exclusively on cancer epitopes: The dataset is drawn from CEDAR, the Cancer Epitope Database and Analysis Resource, rather than the general IEDB corpus that competing tools use.
• Frozen protein language model embeddings: Sequences are represented as 1,280-dimensional ESM-2 vectors (facebook/esm2_t33_650M_UR50D), replacing the amino acid composition, dipeptide, and physicochemical descriptors that classical epitope predictors rely on.
• Eight classifiers benchmarked head to head: Logistic regression, random forest, XGBoost, LightGBM, and naive Bayes were compared against an MLP, a 1D CNN, and a bidirectional LSTM on identical embeddings, with the MLP selected for deployment.
• Full-length protein input: Unlike web tools that accept only single short peptides, the command-line utility takes multi-FASTA files of complete proteins and scans them with user-set minimum and maximum peptide lengths, score thresholds, and top-N ranking.
• Bundled trained classifier: The GitHub package ships the trained MLP weights and its feature scaler, so no retraining is required; ESM-2 itself is downloaded on first run.

#Technical Details

After deduplication and removal of sequences containing non-natural amino acids, the CEDAR-derived training set comprises 37,609 linear B-cell epitopes and 40,637 non-epitopes, evaluated with a stratified 80/20 split. The deployed classifier is a three-layer perceptron mapping 1,280 → 512 → 128 → 1, with batch normalization and 0.4 dropout after each hidden layer. On the internal test split it reaches ROC-AUC 0.85, accuracy 78.0%, precision 0.803, recall 0.898, and F1 0.829, ahead of the BiLSTM (AUC 0.83), the CNN (0.82), random forest (0.78), logistic regression (0.75), and naive Bayes (0.63). Calibration curves show the MLP and CNN tracking the ideal reliability line more closely than the tree-based models.

Independent validation uses a benchmark of 45 cancer-related B-cell epitopes and 31 non-epitopes curated from IEDB and excluded from training. There DeepEpitope reaches ROC-AUC 0.94 and PR-AUC 0.96, against 0.60 and 0.75 for BepiPred-2.0 and 0.54 and 0.65 for LBtope. That benchmark contains only 76 sequences, so the magnitude of the margin rests on few examples even though its direction is consistent across both metrics.

#Applications

DeepEpitope is intended for immunoinformatics pipelines that design epitope-based cancer vaccines: scanning tumour-associated antigens for antibody-accessible linear regions, shortlisting candidates for peptide synthesis and serological assays, and supplying B-cell components to multi-epitope vaccine constructs alongside separately predicted T-cell epitopes. Because it is a Linux command-line tool installable with pip install ., it drops into batch workflows more readily than the single-sequence web forms that dominate this space.

#Impact

The contribution here is the training corpus rather than the model class — the architecture is a standard probe on frozen language model embeddings, and its value rests on the claim that cancer-specific supervision matters for cancer antigens. Several limitations bound the result. The model covers only linear epitopes, leaving the conformational epitopes that make up the majority of real antibody targets to other tools. The independent benchmark is small. The GitHub repository carries no license file, which leaves reuse terms unsettled, and the preprint itself is deposited under an all-rights-reserved notice. The work has not yet been peer reviewed, and no experimental validation of the predicted epitopes is reported.