Fine-Tuning Protein Language Models: pMHC Binding Prediction Case Study

Machaca V.; Grados J.; Lazarte K.; Escobedo R.; Lopez C. · Smart Innovation, Systems and Technologies (2026) · DOI: 10.1007/978-981-95-1353-6_16

Resumen

The interaction between peptides and the Major Histocompatibility Complex (MHC) is a critical factor in the immune response against various threats. In this work, we fine-tuned protein language models like TAPE, ProtBert-BFD, ESM2(t6), ESM2(t12), ESM2(t30), and ESM2(t33) by adding a BiLSTM block in cascade for the task of peptide-MHC class-I binding prediction. Additionally, we addressed the vanishing gradient problem by employing LoRA, distillation, hyperparameter guidelines, and a layer freezing methodology. After experimentation, we found that TAPE and a distilled version of ESM2(t33) achieved the best results outperforming state-of-the-art tools such as NetMHCpan4.1, MHCflurry2.0, Anthem, ACME, and MixMHCpred2.2 in terms of AUC, accuracy, recall, F1 score, and MCC. © The Author(s), under exclusive license to Springer Nature Singapore Pte Ltd. 2026.

Instituciones: Universidad La Salle