Deep Learning for Causal Inference and Latent Dynamical Modeling in Biomedical Research

Abstract

Biological systems are ubiquitous, encompassing complex molecular networks governing single-cell organisms to expansive ecosystems profoundly impacting our planet's environment. In biology, the adoption of a systems approach seeks to achieve a comprehensive, quantitative understanding of living organisms comparable in some ways to the kind of understanding we have of systems in engineering and physics. In this context, a major challenge in scientific AI is causal learning. To address emerging biomedical questions, this work proposes a deep neural architecture that learns causal relationships between variables by combining high-dimensional data with prior causal knowledge. In particular a combination of convolutional and graph neural networks is utilized within a causal risk framework, specifically designed to handle the high dimensionality and typical sources of noise frequently occurring in large-scale biological data. In experimental evaluations, the proposed learner demonstrate its effectiveness in identifying novel causal relationships among thousands of variables. The results are based on extensive gold-standard simulations with known ground-truth. Additionally, real biological examples are considered, where the models are applied to high-dimensional molecular data and their output compared against entirely unseen validation experiments. These findings showcase the feasibility of using deep neural approaches to learn causal networks at a large scale. <br /> Additionally, this work presents a novel method for learning dynamical systems from high-dimensional empirical data combining variational autoencoders and spatio-temporal attention within a framework that enforces scientifically-motivated invariances. The focus is set to scenarios in which data are available from multiple different instances of a system whose underlying dynamical model is entirely unknown at the outset. The presented approach builds upon a separation, dividing the encoding into instance-specific information and a universal latent dynamics model shared across all instances. This separation is achieved automatically and driven solely by empirical data. The results offer a promising new framework for efficiently learning dynamical models from heterogeneous data. This framework has the potential for applications in various fields, including physics, medicine, biology, and engineering. <br /> In a different approach, this work explores interventional experiments to shed light on the causal structure within a system. Under the framework of instrumental variables, a new and mathematically sound cause-effect estimator is proposed to uncover sparse causal relations based on unpaired data regimes. The primary focus lies in predicting the outcomes of interventions that have not been performed before, based on data gathered from observed interventions with unknown characteristics. To illustrate, this framework addresses inquiries such as how hypothetical alterations through gene-level interventions could impact the growth rate of a cell. The efficacy of this method is studied on simulated benchmarks and semi-simulated test cases incorporating human single cell measurements. <br /> Last, this work intends to advance the prediction and comprehension of individual treatment effects in a longitudinal setting. Specifically, this work is investigating clinical records of patients afflicted with wet age-related macular degeneration which if untreated can lead to severe vision loss and legal blindness. To gain a comprehensive understanding of this disease progression, supervised end-to-end models are devised and evaluated to estimate drug responses based on highly irregular time-series data and forecast future treatment effects at individual patient level.