MACHINE LEARNING - BASED PREDICTION OF INTERLEUKIN-2 INDUCING POTENTIAL OF PEPTIDES

Abstract

Interleukins are pivotal cytokines that orchestrate immune regulation, with their dysregulation contributing to a spectrum of diseases ranging from autoimmunity to cancer. Among them, Interleukin-2 (IL-2) plays a crucial role in T-cell proliferation, immune tolerance, and the efficacy of immunotherapies. The precise identification of IL-2-inducing peptides (IIPs) is fundamental for the rational design of vaccines and immunotherapeutic agents. However, experimental discovery of IIPs is inherently laborious, costly, and limited in throughput, underscoring the urgent need for robust computational approaches. Recent advances in machine learning (ML) have revolutionized peptide immunoinformatics, enabling the high-throughput, accurate, and reproducible prediction of cytokine-inducing peptides directly from sequence data. This thesis provides a comprehensive review and critical assessment of ML methodologies developed for predicting the interleukin-inducing potential of peptides, with a particular emphasis on IL 2. The work begins by contextualizing the immunological significance of interleukins and the central role of IL-2 in immune modulation and therapy. It then systematically addresses the challenges and limitations associated with experimental identification of IIPs, motivating the transition to computational strategies. A detailed methodology is presented for the extraction, preprocessing, and curation of high-quality peptide datasets from the Immune Epitope Database (IEDB), focusing on experimentally validated IL-2 inducers and non-inducers. Rigorous data cleaning, feature engineering, and exploratory data analysis are performed to uncover sequence-level and physicochemical patterns distinguishing IL-2 inducers. The thesis explores a broad spectrum of ML algorithms, feature selection techniques, and validation strategies, providing insights into model interpretability and performance. Comparative analyses highlight the strengths and limitations of current computational tools for IL-2 prediction, identifying key gaps and opportunities for further innovation. The review concludes with a forward-looking discussion on the integration of novel ML architectures, multi-omics data, and explainable AI to enhance the predictive power and biological interpretability of IIP models. The findings underscore the transformative potential of ML-driven approaches in immunological research, facilitating the rapid discovery of therapeutic peptides and advancing the frontiers of translational medicine. This thesis not only serves as an authoritative resource on the state-of-the-art in IL-2 peptide prediction but also provides a reproducible framework and actionable recommendations for future computational immunology studies.