A central problem in biology and for biomedical discovery is the inference of gene regulatory networks, causal networks that allow predicting the effect of any intervention in the system. Key outcomes of this project were the development of theory, algorithms, and computational methods for: 1) causal structure discovery, i.e., learning the underlying cause-and-effect relationships such as which gene up- or down-regulates which other gene, from a mix of observational and interventional data; 2) optimal experimental design of interventions, a key problem given that the space of possible perturbations that can be performed in biology (e.g. all combinations of genetic perturbations, any combination of drugs) is huge and cannot be fully explored experimentally; 3) predicting the effect of untested intervention-context pairs, such as a drug in a new disease context.

With respect to causal structure discovery, we developed methods that could deal with key issues that make this problem challenging in practice, including: the presence of latent (unmeasured) confounders, off-target intervention effects (knock-outs may target also other genes with similar sequences), measurement error in the data collection process (single-cell RNA-seq data is highly zero-inflated), as well as the data coming from unknown disease subtypes and hence data coming from a mixture of causal models. In particular, we developed methods for causal structure discovery that are provably consistent under strictly weaker assumptions than previous algorithms and scale to the large graph sizes needed for applications to gene regulation.

With respect to experimental design, we developed methods for identifying interventions that are optimal in different ways, including: with respect to the amount of information they carry about the underlying causal graph (e.g. the gene regulatory network), as well as with respect to moving the distribution from any given state to a desired state via interventions.

With respect to predicting the effect of untested interventions from a set of tested interventions, we viewed this causal transportability problem as a tensor completion problem and developed novel algorithms based on infinitely wide neural networks that are fast, flexible and effective for the problem of causal transportability. We also benchmarked these algorithms on the problem of virtual drug screening.

Throughout the project, the principal investigator (PI) has actively engaged in activities related to education and research by building a diverse research group attracting talents from underrepresented groups and women, and training these graduate students, undergraduate students, and postdoctoral fellows at the intersection of statistics, machine learning, and the biomedical sciences. The training spanned theory, method development, and applications to important biological and medical problems. To help build the research area and community at large, the PI initiated and organized various conferences, including introducing a new machine learning conference focused solely on causal inference "Causal Learning and Reasoning (CLeaR)" as well as co-organizing the semester-long program on causality at the Simons Institute at UC Berkeley.

The research resulting from this project has been widely disseminated: Several open-source software packages have been developed; in particular, all causal inference algorithms resulting from this project are implemented and freely available in the group's causaldag python library which can be found in the group's github repository (https://github.com/uhlerlab). In addition, the PI disseminated the research results through keynote presentations at various conferences, conference presentations by the involved PhD students, as well as many publications, which were made freely available already pre-publication on the arXiv or bioRxiv preprint repositories to enable accelerated scientific discovery. Finally, material developed as part of this project has been integrated into two courses that the PI teaches at MIT.

Last Modified: 09/28/2023
Modified by: Caroline Uhler