Causality

Gene regulatory networks (GRNs) are intricate causal networks where the biomolecule abundance of each gene causally influences others. The reverse-engineer of causal relations from observational and perturbational data is widely studied in the field of causal inference, which was recognized with the Nobel Prize in 2021. In fact, biomedicine is one of the earliest adopters of causal inference through the design of randomized controlled trials. Despite its Nobel recognition, causal inference remains a youthful discipline, rife with unresolved challenges.

One key challenge is causal cycles, which mainstream causal inference methods tend to circumvent by assuming their absence. This assumption, often encapsulated as Directed Acyclic Graph (DAG), greatly simplifies causal inference into a series of independent regression problems. Nonetheless, real-world systems are often dynamical systems consisting of stablizing feedback loops. Existing approaches struggle to recover the true causal structure or analyze causal network rewiring [1]. Furthermore, many causal inference approaches are NP-hard, without any practical exact solution in high dimensions. Several other profound challenges are often left undiscussed in causal GRN inference studies, such as measurement noise [2].

We initially worked within the mainstream DAG assumption for causal GRN inference [2-4], and only recently came to the idea to address these challenges [1]. Rather than modeling causality as how one variable influences another, we regard causality as the direct effect of one variable on the time derivative of another variable. Our model draws inspiration from physical laws and accommodates causal cycles with the same steady-state data used by existing methods. We implemented model fitting and causal GRN inference through probabilistic programming alongside acceleration from Graphical Processing Units (GPUs). We plan to investigate its unique statistical properties, compare its performance with existing approaches, and broaden its applications in single-cell GRN inference.

[1] Wang, L. et al. Dictys: dynamic gene regulatory network dissects developmental continuum with single-cell multiomics. Nat Methods 1–11 (2023).

[2] Wang, L. & Michoel, T. Efficient and accurate causal inference with hidden confounders from genome-transcriptome variation data. PLOS Computational Biology 13, e1005703 (2017).

[3] Wang, L. & Michoel, T. Controlling false discoveries in Bayesian gene networks with lasso regression p-values. arXiv:1701.07011 (2017).

[4] Wang, L., Audenaert, P. & Michoel, T. High-Dimensional Bayesian Network Inference From Systems Genetics Data Using Genetic Node Ordering. Front. Genet. 10, (2019).