MSc Thesis Defense: Şevki Aybars Türel, ACCURATE AND INTERPRETABLE PREDICTION OF SUMOYLATION SITES FROM PROTEIN SEQUENCES, Date & Time: 14 July, 2026 – 1:30 PM, Place: L067

ACCURATE AND INTERPRETABLE PREDICTION OF SUMOYLATION SITES FROM PROTEIN SEQUENCES

Şevki Aybars Türel
Computer Science & Engineering, MSc Thesis, June 2026

Thesis Jury

     Assoc. Prof. Öznur TAŞTAN (Thesis Advisor)

  Assoc. Prof. Umut ŞAHİN

  Prof. Hilal KAZAN

Date & Time: July 14th, 2026 – 1:30 PM

Place: FENS L067

Keywords : Protein Language Models (PLMs), SUMOylation, Protein

Post-translational Modifications, Interpretability, Data Augmentation for Protein Sequences

Abstract

SUMOylation is a critical post-translational modification (PTM) that regulates essential cellular processes, including transcriptional regulation, chromatin organization, cell cycle progression, and DNA damage repair. Dysregulation of SUMOylation has been implicated in numerous diseases, including Alzheimer's disease, cancer, and diabetes. Although experimental identification of SUMOylation sites is highly accurate, it remains labor-intensive, costly, and difficult to scale, motivating the development of accurate computational prediction methods.  This thesis presents an accurate and interpretable framework for SUMOylation site prediction based on protein language models (PLMs). We construct leakage-free benchmark datasets with controlled sequence similarity and investigate homology-based data augmentation strategies to improve learning from limited positive examples. Three ESM-2 650M models are fine-tuned using different augmentation strategies and evaluated on common independent benchmark datasets to enable fair comparison. Experimental results demonstrate that the best model, trained with representative homology-based data augmentation, achieves an AUC of 0.918 and an AUPR of 0.836, outperforming the strongest published method by 12.5% and 21.3%, respectively. We further evaluate the proposed model on a challenging benchmark containing SUMOylated sites without known consensus motifs and non-SUMOylated sites containing canonical motifs. On this benchmark, the proposed model improves AUC by 42.14% and AUPR by 12.43% over the strongest published baseline. Finally, we develop an interpretability framework that systematically quantifies the contribution of individual amino acid positions to the model's predictions, providing biological insight into the sequence context underlying SUMOylation site recognition.