Inferring the role of the microbiome on survival in patients treated with immune checkpoint inhibitors: causal modeling, timing, and classes of concomitant medications

Treatment with Immune Checkpoint Inhibitors (ICIs) has improved patient outcomes across a wide variety of cancers. Not all patients respond to these drugs and there is a need to identify biomarkers of response. Three recent studies have shown that microbes are associated with response and overall survival (OS) in renal cell carcinoma (RCC), non-small cell lung cancer (NSCLC) and melanoma [1‐3]. The microbiome may be a key player in response to ICI therapy and a potential biomarker of treatment response.

The microbiome is known to interact with the immune system, but how it affects response to ICIs is not fully understood. The effectiveness of ICI treatment relies on active T-cell infiltration of a tumor; microbes have been associated with increased Tumor Infiltrating Lymphocytes in an IL12-depended manner [2]. However, other immune cells dampen response to ICIs such as myeloid-derived suppressor cells and FOXP3 & CD4 + CD25+ T-regulatory cells, the levels of which have also been associated with the microbiome [4]. Moreover, the microbiome has been associated with another, systemic form of immune repression characterized by the production of prostaglandins [5‐8].

Several medications commonly used during routine oncologic care and ICI treatment can influence inflammation pathways and/or the microbiome. Corticosteroids (CS) affect both of the aforementioned T-cell subtypes and the prostaglandin-related inflammatory pathways [9]. Additionally, antibiotics (ABx) have a direct effect on the microbiome by killing or halting the growth of bacteria. Proton pump inhibitors (PPIs), histamine 2 blockers (H2Bs), non-steroid anti-inflammatory drugs (NSAIDs), and CS have also been associated with changes in the microbiome but, in contrast to antibiotics, this mechanism is indirect [10]. PPIs, by inhibiting gastric acid secretion, alter the pH of the gut and change the number and types of bacteria that pass through the stomach [11]. Notably, if the taxa enriched by the PPI-induced pH change are also important for response to ICIs, then PPI treatment during ICI may influence clinical outcomes. The effect of other medications on clinical response may be challenging to interpret given that the effects may influence both the microbiome and ICI response.

In order to disentangle these complex interactions, we created a model of the relationship between patient characteristics, medications that affect the microbiome, inflammation, and survival. Second, we performed a retrospective analysis of patients who received ICI therapy for advanced cancer between 2011 and 2017 including medications with known effects on either the microbiome or its pathway toward affecting ICI response. Third, we estimated the association for each medication with OS. Fourth, we analyzed the effects of medications longitudinally, in order to decouple confounding variables at different time points. Fifth, we controlled for variables that broadly describe differences in baseline statuses (e.g. Eastern Cooperative Oncology Group performance status (PS)) of individuals who received concomitant medications and those who did not. Sixth, we compared the associations across several cancers, for which the medications are prescribed in subtly different ways that can be leveraged to gain further insight into the causal effects. Finally, we related these results to the microbes shown to be enriched or depleted in individuals who respond to ICIs. The combination of these strategies gives layers of support to defining the role of the microbiome in the context ICI treatment of cancer.

The effects of medications or other variables are difficult to parse in a dynamic setting such as during treatment for cancer. We used a variety of methods to show that ABx and CS are significantly associated with decreased OS in several cancer types.

The association of CS with ICI response and OS remains controversial. Our observed association is consistent with other observations of decreased OS in NSCLC [9]. However, Ricciuti et al. showed no effect of CS on OS in NSCLC when given on the same day as ICI start, when the CS was prescribed for reasons other than “cancer-related palliative indications” [48]. Our records lack some variables needed to replicate those results, however our results are consistent with aspects those findings. For example, dexamethasone treatment showed a strong negative association with OS across several cancer types, consistent with its use for brain metastases and anorexia, which are all indicators of poor clinical outcome. On the other hand, several of our analyses demonstrated associations between CS and OS that may not be due to selecting a sub-cohort with a poor prognosis. Our first causal strategy, the time analysis, showed similar results when restricting CS medications to a single day, but a larger effect when a wider time window was used (Table 2). Similar effects have been observed previously, but with little consistency in the time window tested [2, 3, 9, 48‐52]. Our second causal strategy, controlling for covariates, cannot be directly compared because our dataset did not include central nervous system metastases. However, when we control for metastatic stage and PS, the CS association remains. Our third causal strategy, comparisons between cancers, shows that the CS association with OS is observed in cancers for which brain metastases are not common, such as RCC, and for specific CS not typically used for brain metastases, such as methylprednisolone in HNSC. This suggests that understanding the association between CS and the response to ICIs may require more granular assessment of CS types (i.e. rather than collapsing to 10 mg prednisone equivalent) and cancers.

We applied the same logical framework to ABx treatment to demonstrate an effect on OS. Unlike CS, the majority of studies have found an association between ABx use and ICI response, independent of the time window (Table 2). Our longitudinal analysis showed a global maximum HR well before the start of ICIs, consistent with the ABx effects persisting for long periods. Given this result, it is unlikely that acute illnesses drive the association between ABx and OS. However, a recent prospective study found that ABx given currently with ICI treatment did not significantly affect OS for a group of patients with lung, skin, or several other cancers [49] (Table 2). We observe lower HRs for the effect of ABx after ICI start, however it remains significant until approximately 120 days post ICI start. We note that within cancers the effect of ABx is highly variable (Fig. 4c); the difference may be due to the composition of the cohorts (e.g. more patients with bladder cancer, where ABx has a strong effect, and fewer with NSCLC, where the effect is less). Our results are consistent with a recent meta-analysis across several cancers, in which the greatest HR was observed in the 42 days before the start of ICIs [50].

When controlling for illness-related covariates that report on the overall health status of the individual (e.g. CCI, PS) the effect of ABx remained significant. Third, the associations of ABx and OS were observed across cancer types (e.g. patients with bladder cancer versus melanoma). A larger fraction of bladder cancer patients were treated with ABx than any other cancer (56%), consistent with their use for urinary tract or as prophylaxis for invasive urologic procedures. On the other hand, melanoma patients treated with ABx were the smallest fraction of any cancer (25%), consistent with this population being less likely to undergo procedures in which prophylactic ABx are used. It is reasonable to suspect that melanoma patients treated with ABx are therefore more compromised than those not treated with ABx. However, an effect of ABx remains, even for bladder cancer. Although it remains probable that the cohorts who receive ABx are different from those who did not in ways that have not been controlled for in analyses, these three analyses add confidence to the association of ABx with OS in the context of ICIs.

We next related the strength of the association of ABx classes with OS and the microbes that those ABx classes affect. The β-lactam ABx were shown to have the strongest association with OS across cancer types. The literature review of antibiotic susceptibilities showed that this diverse class is effective against the Gram-positive phylum Firmicutes. The literature review of the bacterial taxa associated with response to ICIs, showed that the Firmicutes are enriched in responders to ICIs. Moreover, β-lactams are not consistently effective against members of the phylum Bacteroidetes, which was found to be enriched in non-responders. This suggests that the β-lactams may show the strongest signal across all cancers in our dataset because they disrupt the microbiome in such a way that they reduce response to ICIs by depleting the Firmicutes more so than the Bacteroidetes.

The association between ABx prescriptions and OS that we observe is consistent with direct measurements of the microbiome and response to ICIs [1‐3]. However, there is no consensus for which taxa are enriched in the responders to ICIs (Fig. 5). For example, there is causal evidence for Akkermansia muciniphila increasing response to ICIs, however, it was not among the most enriched in the other datasets [1‐3]. Nonetheless, some agreement can be observed between the effects of ABx on isolated taxa and OS. Narrow spectrum β-lactams (e.g. cephalosporins), which show the strongest association with OS, are not effective against Bacteroidetes (enriched in non-responders (1)) but are against A. muciniphila (enriched in responders (2)). However, we note that the effects of ABx can be difficult to predict over long time scales; some broad spectrum β-lactams have resulted in increased Firmicutes post-recovery, despite being effective against them [51].

The results presented here contrast with several assumptions gathered from the literature and described by the causal model (Fig. 1). First, we found that ABx and CS are the only medications significantly associated with OS, despite the inclusion of several medications associated with changes to the microbiome (Fig. 2). This may be due to the types of changes incurred (e.g. PPIs may not significantly change the abundances of those taxa linked to ICI response) or the strength of the effect amid the noise in the data. However, the other two hypotheses were borne out by the analyses.

The CS and ABx medications showed an additive effect on OS, consistent with a collider interaction in the model (Fig. 4a). Also, there was an effect of ABx after controlling for many covariates, consistent with its direct effect on the microbiome and the microbiome playing a role in ICIs (Fig. 4b). This result was consistent with the relationship between the strength of the ABx signal and the bacterial taxa susceptible to that ABx (Fig. 5).

ABx and CS, but not other medications known to affect the microbiome, are associated with reduced OS when administered near the start of ICI treatment. Our results show this finding several cancer types, and for several subclasses of these drugs. These results are consistent with a role of the microbiome in response to ICIs and identify clinical settings where the microbiome is likely to play the largest role, namely NSCLC, melanoma, RCC, HNSC, and bladder cancer. A clear understanding of which microbes are important for ICI responses and in what cancers will require the collection of microbiome samples across a wide variety of clinical settings. However, some information can be gathered by indirect means, which identifies the settings where the microbiome is likely to have the greatest effects. Medications that affect the microbiome given concomitantly with ICIs provide evidence for where microbes play a role. Further work is needed to identify which microbes are important and identify solutions to mitigate these effects and perhaps promote greater response to ICIs.