Interpretable and extendable deep learning model for biological sequence analysis and prediction

点击翻译按钮获取中文摘要

Project Abstract Bioinformatics and computational biology have become the core of biomedical research. The PI Dr. Dong Xu's work in this area focuses on development of novel computational algorithms, software and information systems, as well as on broad applications of these tools and other informatics resources for diverse biological and medical problems. He works on many research problems in protein structure prediction, post-translational modification prediction, high-throughput biological data analyses, in silico studies of plants, microbes and cancers, biological information systems, and mobile App development for healthcare. He has published more than 300 papers, with about 12,000 citations and H-index of 55. In this project, the PI proposes to develop deep-learning algorithms, tools, web resources for analyses and predictions of biological sequences, including DNA, RNA, and protein sequences. The availability of these data provides emerging opportunities for precision medicine and other areas, while deep learning as a cutting-edge technology in machine learning, presents a new powerful method for analyses and predictions of biological sequences. With rapidly accumulating sequence data and fast development of deep-learning methods, there is an urgent need to systematically investigate how to best apply deep learning in sequence analyses and predictions. For this purpose, the PI will develop cutting-edge deep-learning methods with the following goals for the next five years: (1) Develop a series of novel deep-learning methods and models to specifically target biological sequence analyses and predictions in: (a) general unsupervised representations of DNA/RNA, protein and SNP/mutation sequences that capture both local and global features for various applications; (b) methods to make deep-learning models interpretable for understanding biological mechanisms and generating hypotheses; (c) “rule learning”, which abstracts the underlying “rules” by combining unsupervised learning of large unlabeled data and supervised learning of small labeled data so that it can classify new unlabeled data. (2) Apply the proposed deep-learning model to DNA/RNA sequence annotation, genotype-phenotype analyses, cancer mutation analyses, protein function/structure prediction, protein localization prediction, and protein post-translational modification prediction. The PI will exploit particular properties associated with each of these problems to improve the deep-learning models. He will develop a set of related prediction and analysis tools, which will improve the state-of-art performance and shed some light on related biological mechanisms. (3) Make the data, models, and tools freely accessible to the research community. The system will be designed modular and open-source, available through GitHub. They will be available like integrated circuit modules, which are universal and ready to plug in for different applications. The PI will develop a web resource for biological sequence representations, analyses, and predictions, as well as tutorials to help biologists with no computational knowledge to apply deep learning to their specific research problems.