NEW METHODS FOR OMICS AND PHARMACEUTICAL DEEP LEARNING

Abstract

Artificial intelligence (AI) has emerged as a solution recently and is transforming healthcare and therapeutics. However, there are still challenges for deep learning (DL)-based AI in non-image data learning of biomedical area. Biomedical investigations such as omics-based disease diagnosis and pharmaceutical drug discovery frequently rely on non-image data of high-dimensions, arbitrary-order but intercorrelated features, relatively low-sample sizes, diverse samples, and sparse features (Bio-HARDS), which pose challenges in exploring highly efficient DL algorithms for these tasks. There is thus a pressing need to develop more robust, high-performing, but explainable DL approaches for Bio-HARDS tasks. This thesis describes a novel methodology for omics and pharmaceutical DL using unsupervised and supervised learning to learn the omics and pharmaceutical data. The unsupervised data representation methods of AggMap/MolMap combined with supervised CNN-based AggMapNet/MolMapNet architectures establish the new paradigms for enhanced learning and interpretability of omics and pharmaceutical molecular data.