AI for single-cell multi-modality biology

Abstract

Cells are the fundamental building blocks of all life on Earth. All cell types of one organism, despite sharing the same genetic blueprint, undertake distinct cellular fates guided by underlying molecular networks. These networks, encompassing gene transcription, translation, and their intricate intra- and extra-cellular interactions, regulate and delineate cellular function, identity, and fate. Understanding the spatial-temporal dynamics of these molecular regulatory networks and their impact on cellular functions is pivotal for understanding development, disease onset, and aging. With the advancement of single-cell multi-modality measurement technologies (e.g., scRNA-seq, patch-seq), we are now able to capture an unprecedented level of detail from individual cells from molecular to functional level. However, this influx of data, laden with intricate information about diverse molecular interactions, presents an analytical challenge that traditional methods struggle to address comprehensively. AI, with its capability to discern complex non-linear patterns in high-dimensional datasets, holds great promise to illuminate these complexities. Yet, its prevalent black-box nature obscures confidence in AI-derived insights. Moreover, current analytical methods are designed to perform a single task, only providing a partial picture of the multi-modality data. In this thesis, I focus on introducing single-cell multi-modality biotechnologies for gene-to-function mapping and developing explainable and predictive AI models to analyze different modalities of single cells. First, I review recent progress in flexible electronics capable of tracking the electrical activity in the neural and cardiac systems (Chapter 1). Then, by fusing the flexible electronics with spatially resolved sequencing, I introduce the in situ electro-seq, which allows stable mapping of electrical activity at the millisecond time scale and profiling of gene expression from the same cells across intact biological networks. In addition, I develop machine-learning-based cross-modal analysis to identify gene-to-electrophysiology relationships throughout cardiomyocyte development (Chapter 2). Finally, I focus on predictive and explainable AI models capable of generic analysis for single-cell multi-modality measurements (Chapter 3).