New causal discovery algorithm over censored variables identifies subtype-specific drivers of breast cancer progression.

Abstract

Background: Many research domains are producing large, multi-scale, multi-modal datasets at growing rates with mixed variable types (continuous, discrete, censored). Identifying possible cause-effect associations in such datasets is essential for predicting outcomes and proposing possible interventions. Probabilistic graphical models (PGMs) have emerged as a robust, interpretable way to analyze such datasets, but current graph learning algorithms cannot incorporate time-to-event (censored) variables, which are important in many systems (e.g., patient survival). Instead, regression models are typically used for survival analysis of single censored variables, but these cannot assess cause-effect interactions.

Results: Here, we present a new mathematical framework to incorporate multiple censored variables into mixed graphical models. A novel efficient algorithm, CausalCoxMGM, is implemented, which is extensively evaluated on synthetic and real-life high-dimensional biomedical datasets (cardiovascular disease, breast cancer). CausalCoxMGM was able to recover effectors of censored variables, supported by literature, and provided new mechanistic insights on the differences between ER+ and ER- breast cancers.

Conclusions: CausalCoxMGM is a flexible computational framework for learning potential cause-effect relations from observational data of mixed data types, including multiple censored variables. The resulting graphs are interpretable and can be used to generate testable hypotheses or build efficient predictors of any outcome.