Antibody Deimmunization via Targeted, Machine Learning Guided Sequence Design Using Fine Tuned Protein Language Models for Computationally Efficient Design
ISEF · 2026 Computational Biology and Bioinformatics
Overview
Rheumatoid arthritis (RA) is an autoimmune disease affecting approximately 18 million people worldwide. Many treatments rely on monoclonal antibody drugs that block inflammatory signals such as TNF-a and IL-6. However, up to 40% of patients develop anti-drug antibodies (ADAs), which reduce treatment effectiveness by targeting these therapies as foreign proteins. Reducing immunogenicity while preserving drug function remains a major challenge in biologic design. This project presents a computational pipeline for identifying antibody sequence modifications that reduce predicted immunogenicity while maintaining structural stability. Antibody sequences were analyzed using NetMHCIIpan to identify immune-reactive regions, and alternative amino acids were generated using the protein language model ESM-2. A key innovation of this work is a targeted masking strategy that biases model training toward immunogenic "hotspot" residues, ensuring precise edits in high-risk regions. Through multiple iterations of optimization and hybridization with biological antibody design principles, the final approach achieved a 440% improvement over random masking - a 66% success rate in reducing predicted immunogenicity (meaning the model was able to successfully produce antibodies with reduced GIS scores for 66% of the antibodies in the test set) out of 106 test sequences, with an average sequence similarity of 88.3% and 13.5 targeted mutations per sequence. These results demonstrate that protein language models can be guided with biologically-informed strategies to perform targeted, clinically relevant antibody deimmunization, providing a robust framework for ML-assisted therapeutic design.
Competition history
- ISEF 2026
Resources
Related projects
ISEF · 2026
De novo Design of Broadly Neutralizing Antibodies for Rapidly Evolving Pathogens
ISEF · 2023
IntelliBind: A Novel Deep-Learning Framework for High-Affinity Antibody Design
ISEF · 2018
Designing a Modified HAC-PD-1 to Antagonize PD-L2 for the Improvement of Cancer Immunotherapy
ISEF · 2025
Machine learning optimized in silico design of a de novo RAGE inhibitor: a small molecule approach to slowing the progression of inflammatory disease
Closest projects by meaning, across every fair and year in the corpus.
Source: Regeneron International Science and Engineering Fair