Finding a link between the microbiome and disease

In 1884, Robert Koch identified a causal relationship between a certain Mycobacterium species and tuberculosis. As the backbone of this discovery, he postulated a four-step protocol for inferring simple causal relationships between bacteria and diseases, resulting in the additional discovery of pathogens inducing cholera and anthrax poisoning. Although his findings were considered a breakthrough in the understanding of infectious diseases, Koch’s postulates are largely obsolete. We now know that viral particles, as well as symbiotic, commensal, and predatory microbial interactions, are integral to our health and well-being as part of complex microbial ecosystems. Our bodies contain hundreds of trillions of bacterial, fungal, archaeal, and viral microbes. Although their impact on our health is yet to be fully understood, the abundance and diversity of these microbes has led to a recent explosion of microbiome research.

The current gold standard of causal inference is through randomised trials. However, pathogen and disease related trials raise a number of ethical conundrums, therefore microbiome research to date has mostly only produced associations and correlations to specific diseases (1). While informative in their own right, associations can lead to skewed interpretations of the causal mechanisms in a system. For instance, a study showed that babies delivered by cesarean sections have been associated with a higher rate of obesity. The hypothesis in this study was that if babies were delivered by caesarean section they were not exposed to maternal microbiota at birth, therefore heightening their chance of microbiome-related disorders later in life, such as obesity (2). However, the prevalence of caesarean sections has also been shown to be higher for obese mothers (3). Therefore, maternal obesity may be a common cause for caesarean sections and child obesity, with no real cause-effect relation between the latter two.

In ‘The Book of Why,’ Judea Pearl argues that there are three mathematically distinct levels of causal inference. The first is association which is described as being able to detect if some events are related, e.g., bacterium X is associated with an increased risk of gastric cancer. The second is intervention which is being able to predict possible effects of a given action, e.g., stimulating bacteriophage Y will reduce the risk of obesity. The third is counterfactual reasoning which is being able to reevaluate situations given that some circumstances are different, e.g., would a patient still have Covid-19 if she had a higher abundance of fungus Z in her gut? The vast majority of current machine learning methods are concerned only with the first level. This is also the case in microbiome research, where studies often show correlations and associations rather than causal mechanisms. Pearl argues that causal inference can only be derived from the causal graph, i.e., by making and testing assumptions on the mechanism of the systems of interest (2).

While these proposed treatments have shown promise in both mouse and human models, they are far from perfect. This is where the third level becomes important. When a drug or treatment shows high promise, yet is still ineffective, it is important to understand why this is the case and which conditions could be altered to improve the result. This kind of counterfactual reasoning is currently making its way into the state-of-the-art of machine learning, as it stands to reason that predictions can be made more accurately when the learning algorithm is fed a causal mechanism, rather than a series of associated factors (4). The collaborative work done between the Ghent University and our company BioLizard, approaches Pearl’s levels systematically. Starting from proven causal links, such as those observed between gastric cancer and Helicobacter pylori in the stomach microbiome or between dental plaque and Fusobacterium nucleatum in the oral microbiome, state-of-the-art causal inference models are being fit to the ecological systems of the human body. In silico frameworks developed through this research will either substantiate or refute the associative links made between the microbiome and disease, and could dramatically accelerate the development of drugs, biomarkers and personalised nutrition.

This research is supported by Flanders Innovation & Entrepreneurship (VLAIO) through the Baekeland grant HBC.2020.2292.

References:

1.Wang, B., Yao, M., Lv, L., Ling, Z., & Li, L. (2017). The Human Microbiota in Health and Disease. In Engineering (Vol. 3, Issue 1, pp. 71–82). Elsevier Ltd.

2. Pearl, J., & Mackenzie, D. (2018). The Book of Why: the new science of cause and effect. In Notices of the American Mathematical Society. Basic Books.

3. Marotz, C. A., & Zarrinpar, A. (2016). Treating obesity and metabolic syndrome with fecal microbiota transplantation. In Yale Journal of Biology and Medicine (Vol. 89, Issue 3, pp. 383–388). Yale Journal of Biology and Medicine Inc.

4. Richens, J. G., Lee, C. M., & Johri, S. (2020). Improving the accuracy of medical diagnosis with causal machine learning. Nature Communications, 11(1), 1–9.