Causal discovery for the microbiome

Abstract

Measurement and manipulation of the microbiome is generally considered to have great potential for understanding the causes of complex diseases in humans, developing new therapies, and finding preventive measures. Many studies have found significant associations between the microbiome and various diseases; however, Koch's classical postulates remind us about the importance of causative reasoning when considering the relationship between microbes and a disease manifestation. Although causal discovery in observational microbiome data faces many challenges, methodological advances in causal structure learning have improved the potential of data-driven prediction of causal effects in large-scale biological systems. In this Personal View, we show the capability of existing methods for inferring causal effects from metagenomic data, and we highlight ways in which the introduction of causal structures that are more flexible than existing structures offers new opportunities for causal reasoning. Our observations suggest that microbiome research can further benefit from tools developed in the past 5 years in causal discovery and learn from their applications elsewhere.

Figure 1:. Two DAGs describing different causal structures

The DAGs describe different causal structures (A, B) for a system involving a microbiome community (C), an outcome node of interest (O), and an environmental or confounding factor (E). The directed edges represent causal relationships between the variables. DAGs=directed acyclic graph.

Figure 2:. Lower bound of the causal effect of on

Causal effect estimated by intervention calculus when the directed acyclic graph is absent under three different scenarios (A–C) illustrated by the graph structures. The red arrows represent the causal mechanism in which the microbiome affects the outcome status (true effect=0·75) and the blue arrows represent the reversed causal mechanism in which the outcome status affects the microbiome (true effect=0). The dashed arrows represent potential interactions between C, C*, and the rest of the operational taxonomic units. The box–bar plots summarise the results of the simulations obtained for different sample sizes shown on the horizontal axis. The bars show the proportion of non-zero estimates (right vertical axis) and the boxes show the distributions of the of the non-zero estimates. C=microbiome community.

Figure 3:. A labelled directed acyclic graph describing the causal structure of Clostridioides difficile

C represents a person’s microbiome community, which is either dominated (C=1) or not dominated (C=0) by C difficile; E represents whether or not the person receives antibiotic treatment. O represents whether or not the person has persistent diarrhoea. The directed edges represent causal relationships between the variables; and the label implies that the causal direct effect (E to O) vanishes when the microbiome community is dominated by C difficile (ie, when C=1).

Figure 4:. Causal structures and results of the LDAG simulation

The discovery rate (vertical axis) is the proportion of cases in which either method successfully discovered the causal mechanism: LDAG-based method (blue), DAG-based method (red). The different plots correspond to different a-values and the different curves within each plot correspond to different RRs. C=microbiome community. DAG=directed acyclic graph. LDAG=labelled DAG. O=outcome node of interest. RR=risk ratio.