Enhancing Linear B-Cell Epitope Prediction via Evolutionary Scale Modeling (ESM-2) and Hybrid Feature Integration
Abstract
Experimental identification of B-cell epitopes is time-consuming and resource-intensive, motivating the adoption of computational prediction approaches. This study examines the progression of sequence-based strategies by contrasting traditional machine learning methods (SVM, XGBoost, etc.) and early deep learning techniques (CNN, RNN, etc.) with sophisticated protein language models such as ESM-2. By systematically evaluating multiple feature encodings, including amino acid composition, dipeptide composition, k-mer tokenization, and evolutionary embeddings, we established that capturing both global semantic context and local sequence motifs is essential for robust predictive performance. Optimal performance was achieved using a hybrid framework that combined 15B-parameter ESM-2 embeddings with k-mer features. This hybrid approach achieved an AUC of 0.743 and an MCC of 0.333 on independent validation using an Extra Trees classifier, and an AUC of 0.738 and an MCC of 0.356 using an XGBoost classifier on the same dataset. Our findings indicate that transformer-based embeddings yield enhanced protein representations; however, their predictive efficacy is optimized when integrated with local motifs and ensemble learning, presenting a scalable, high-accuracy alternative to conventional and standalone deep learning frameworks.
Related
- Explainability from Training with Applications to TCR-Epitope Prediction
- CellDiffusion: Enhancing scRNA-seq data integration with accurate inference of cell-cell association
- In Vitro Modeling of Sarcopenia-Related Pathology and Exercise-Associated Adaptations: Integration with In Silico Approaches
- From Prediction to Mechanism: Explainable AI for Plant–Microbiome Multi-Omics Integration and Biological Discovery
- An (αβ)2 Architecture of the Arabidopsis thaliana tRNA Splicing Endonuclease Complex: Structural Modeling and Evolutionary Analyses
BibTeX
@misc{pal2026enhancing,
title = {{Enhancing Linear B-Cell Epitope Prediction via Evolutionary Scale Modeling (ESM-2) and Hybrid Feature Integration}},
author = {Prasun Pal and Sai Jahnavi Pasumarthi and Sandipan Chakraborty},
howpublished = {ChemRxiv},
year = {2026},
month = sep,
doi = {10.26434/chemrxiv.15009154/v1},
url = {https://chemrxiv.org/doi/full/10.26434/chemrxiv.15009154/v1},
note = {CC BY-NC-ND},
}
RIS
TY - UNPB
TI - Enhancing Linear B-Cell Epitope Prediction via Evolutionary Scale Modeling (ESM-2) and Hybrid Feature Integration
AU - Prasun Pal
AU - Sai Jahnavi Pasumarthi
AU - Sandipan Chakraborty
PY - 2026
DA - 2026/09/21
PB - ChemRxiv
DO - 10.26434/chemrxiv.15009154/v1
UR - https://chemrxiv.org/doi/full/10.26434/chemrxiv.15009154/v1
AB - Experimental identification of B-cell epitopes is time-consuming and resource-intensive, motivating the adoption of computational prediction approaches. This study examines the progression of sequence-based strategies by contrasting traditional machine learning methods (SVM, XGBoost, etc.) and early deep learning techniques (CNN, RNN, etc.) with sophisticated protein language models such as ESM-2. By systematically evaluating multiple feature encodings, including amino acid composition, dipeptide composition, k-mer tokenization, and evolutionary embeddings, we established that capturing both global semantic context and local sequence motifs is essential for robust predictive performance. Optimal performance was achieved using a hybrid framework that combined 15B-parameter ESM-2 embeddings with k-mer features. This hybrid approach achieved an AUC of 0.743 and an MCC of 0.333 on independent validation using an Extra Trees classifier, and an AUC of 0.738 and an MCC of 0.356 using an XGBoost classifier on the same dataset. Our findings indicate that transformer-based embeddings yield enhanced protein representations; however, their predictive efficacy is optimized when integrated with local motifs and ensemble learning, presenting a scalable, high-accuracy alternative to conventional and standalone deep learning frameworks.
ER -