Causal Discovery at Scale for Biological Mechanism Inference

Mechanism discovery is a central objective in science. In biology, mechanisms (cascades of molecular events that lead to observable phenotypes) provide insight into cellular programming, explain disease processes, and identify targets for therapeutics. Despite advances in high-throughput experimentation that enable genome-scale measurement and the success of deep learning models in prediction, data-driven discovery of biological mechanisms remains elusive. Recently, causal discovery has emerged as a method for learning cause and effect relationships, represented by directed acyclic graphs (DAGs), from data. Given the success of networks for representing mechanistic knowledge in biology (e.g., gene regulatory networks), causal discovery algorithms are promising approaches for mechanism discovery in biology. In this thesis, I examine the current state of causal discovery and its application to biological mechanism inference. I address challenges including (1) scale, enabling inference over genome-wide systems with limited samples; (2) structural priors, incorporating biological knowledge or correlative structure to guide inference; and (3) robustness to real-world biological data characterized by heterogeneous environments, perturbations, and limited ground truth.

In our first contribution, we introduce a hybrid constraint- and score-based causal discovery algorithm, named SP-GIES, that first estimates a skeleton from observational data and then learns the DAG from interventional data over a constrained search space. We see an improvement in convergence of greedy algorithms, in both accuracy and runtime, by first estimating a structural prior. Moreover, we analyze SP-GIES for downstream optimal experimental design for choosing maximally informative perturbational experiments. 

In the second contribution, we propose a distributed framework for causal discovery using our novel causal graph partition on a structural prior we call a superstructure. We prove under certain assumptions that causal discovery can be decomposed into independent subproblems by leveraging a closely related maximal ancestral graph class, which is defined by projections of DAGs onto subsets of nodes. We scale causal discovery to 10,000 variables with our distributed framework, achieving genome-scale recovery of synthetic biological networks. 

Finally, these contributions are applied to a low-sample, high-dimensional perturbational dataset to uncover cellular processes involved in response to radiation from human transcriptomics data. We compare a VAE-based causal discovery algorithm DAG-GNN, a gene regulatory network inference algorithm GENIE3, and the standard bioinformatics approach of differential expression analysis to identify important genes in response to radiation. We use bioinformatics techniques to map these genes to known pathways, and show that graph algorithms achieve higher enrichment of radiation pathways. Additionally, we find that the presence of a strong upstream perturbation induces different environments from which we can learn invariant mechanisms; in the causal graphs, these invariant edges corroborate evidence of a bifurcation of reactive oxidative stress and cell death pathways in response to radiation.

These contributions demonstrate the impact of scaling causal discovery in genome-wide studies to infer regulatory response networks and identify important genes for downstream pathway analysis. More broadly, this work aims to enable causal, testable insights from high-dimensional biological data.