Пошук академічних текстів

Інформація × Реєстраційний номер 2124U004705, Матеріали видань та локальних репозитаріїв Категорія Препринт Назва роботи Advancing Prediction of MHC-Peptide Complex Structures through Enhanced AlphaFold Fine-Tuning Автор Kalitin Dmytro Дата публікації 01-01-2024 Постачальник інформації Український католицький університет Першоджерело https://hdl.handle.net/20.500.14570/4835 Видання Опис Protein structure prediction plays a pivotal role in understanding biological pro- cesses, with a particular emphasis on Major Histocompatibility Complex (MHC)- peptide complexes due to their significance in immune responses. This thesis fo- cuses on my contributions to the research that delves into the realm of enhanced Al- phaFold fine-tuning, specifically tailored for the precise prediction of high-resolution MHC-peptide (pMHC) complex structures. As a result, we introduce the MHC-Fine model, an augmentation of the original AlphaFold framework. The proposed AlphaFold multimer training pipeline incorporates key elements to refine MHC-peptide complex predictions. This includes leveraging pre-trained weights obtained from DeepMind, allowing for fine-tuning to address the intricate task of predicting peptide structures in detail. The utilization of templates, reference protein structures that enrich prediction accuracy, was integrated into the pipeline. Further enhancements involve meticulous selection of hyperparameters such as loss weight factors, learning rates, and schedulers. Notably, the exploration of impact of additional data provided to the model constitutes a significant advancement in training pipeline optimization. The results of this research demonstrate the superiority of the MHC-Fine model in comparison to other methods. Our model outperforms both the original Al- phaFold and homology-modeling-based method PANDORA, achieving a remark- able median RMSD of 0.65 Å. This groundbreaking precision in pMHC structure prediction emphasizes the efficacy of the proposed AlphaFold fine-tuning pipeline. Додано в НРАТ 2025-05-09 Закрити