Neural Network Models for Modelling, Design and Optimization of Structurally Complex Entities in Biomedical Data Science

点击翻译按钮获取中文摘要

The rise of deep learning approaches has reshaped the research and development landscape in machine learning and data science. Deep learning modelling has been rapidly evolving with superior performances in many challenging tasks ranging from feature selection, classification, data understanding, decision making, to creative designs. The success of deep learning is largely due to (1) the acceptance of distributed representation and symbolic embedding theories, (2) the wide use of convolutional operations for automatic feature learning, (3) the modelling of memory and attention mechanisms, (4) the marriage of deep neural networks with reinforcement learning, and (5) the born of neural or deep generative models (DGMs). As a fast-growing family of deep learning techniques, DGMs use deterministic neural networks to model dependencies among visible and latent variables. Majorly applied in computer vision and natural language processing, DGMs enable versatile functionalities such as sampling, seasoning, simulation, design, optimization, and domain transformation. The adoption of deep learning in biomedical data science exhibits huge potentials but meanwhile faces grand challenges to represent and model structurally complex objects, learn on small data, and design from exponential number of possibilities. As a continuation of our research excellence in bioinformatics, the long-term goal of this research program is to investigate and implement novel deep-learning-based approaches (particularly DGMs) for the effective representation and modelling of extremely long sequences (ELSs) and structurally complex entities (SCEs), and apply them to solving challenging real-world biomedical data science tasks that are currently largely limited due to structural complexity and data availability. It is anticipated that our research and development will be disruptive by providing fresh ideas, inspirations, and findings. Three short-term objectives will be pursued within a five-year period. First, novel multi-module, multi-modal, and few-shot deep learning methods will be devised to model extremely long sequences and structurally complex entities which are often encountered in biomedical data analytics. Second, data-model co-evolutionary learning processes will be fully investigated to better solve black-box global multi-objective optimization tasks. We seek novel learning paradigms that can integratively improve both model learning and data quality through the data representation space. Third, our devised and newly emerged deep learning methods will be applied to address three challenging biomedical data science problems: genome annotation, aptamer drug design, and interactome predictions. This research will deliver a collection of next-generation machine learning approaches and data science tools that are beneficial to the research communities and the public good.