Quantifying interventional causality by knockoff operation

Causal inference between measured variables is crucial to understand the underlying mechanism of complex biological processes at a network level but remains challenging in computational biology. We propose an innovative causal criterion, knockoff conditional mutual information (KOCMI), to accurately infer interventional direct causality without prior knowledge of the network structure using either time-independent or time-series data. KOCMI performs knockoff operation on a variable as its virtual intervention, which preserves the original network structure, and then identifies the causality between two variables by estimating the distributional invariance before and after such a virtual intervention. We show that, algorithmically, KOCMI enables quantification of causal relationship, even for networks with loops, and, theoretically, is also consistent with the do-calculus causal analyses but without their prerequisite of the network structure. KOCMI shows superior performance on benchmark and real datasets, comparing with existing methods. Overall, KOCMI provides a powerful tool in inferring interventional causality, which is theoretically ensured and experimentally validated by real intervention data.