Discovering Causal Mechanistic Pathways in Sepsis-associated Acute Respiratory Distress Syndrome

In this issue of the Journal, Jones and colleagues (pp. 47–56) find that sRAGE (soluble receptor for advanced glycation end products) blood levels are related to sepsis-associated acute respiratory distress syndrome (ARDS) (1). Moreover, they make the bold claim that sRAGE causally contributes to sepsis-associated ARDS. This is an important step to take, for multiple reasons. First, the sepsis and ARDS fields are littered with hundreds of reports of an association between inflammatory pathway molecule measurements and clinical outcomes (2). Although they are interesting, these associations have not led to the introduction of new therapies to prevent or ameliorate sepsis-induced lung injury/ARDS or any other clinically important outcome (3). Second, identification of a true causal pathway points the field toward a plausible intervention strategy and relevant biomarkers, as modulation of a causal pathway would result in altered clinically relevant outcomes, not just a change in a blood cytokine level. Third, claiming causality is bold because it invites us to examine the methodology that supports the inference of causality (4). Understanding when the data support causal inference and when they do not is a key issue that many previous association studies have not adequately addressed. This may have contributed to the critical care medicine field embarking on inadequately supported interventional approaches, as indicated by the many failed sepsis-related clinical trials (5). Examining this methodology also makes us ask, what additional information should we seek, and what are the best next research steps? Fourth, using the current study by Jones and colleagues as a “proof of principle,” it is clear that we may be able to extract more knowledge from a thoughtful combination of clinical and basic biological measures from our current and growing patient datasets.

Is sRAGE, among Many Inflammatory Biomarkers, a Good Starting Point?

Many blood biomarkers have been reported to be associated with the development of ARDS (6). Among these, the investigators chose to study sRAGE (1). The investigators support this as a reasonable choice by referring to the substantial body of evidence already in place that demonstrates that RAGE is found in the right place (lung [7]), at the right time (when an inflammatory response is induced [8]), doing the right thing (potentiates the inflammatory response [9]), and with the right effect on perturbation (RAGE knockout mice have better clinically related outcomes [10]) to allow the hypothesis of causality to be considered. But does this model-derived knowledge translate to the clinically relevant outcome of sepsis-associated ARDS in humans? Evidence of causality in humans is therefore critically important.

Approaches to Causal Inference

The most common approach to test for causality is to intentionally modify the putative causal agent and then measure the effect. This works in animal models (e.g., RAGE knockout mice [10]) but typically is very challenging and/or very expensive to do in humans (e.g., clinical trials of experimental interventions in humans). A more recently applied approach to infer causality from observational data is to use instrumental variables (4). For example, if a new tobacco tax (the statistical instrument that reduces smoking) reduces the incidence of lung cancer, we can infer that smoking causally contributes to lung cancer because there is no other plausible way that a tobacco tax could have this effect (11). When genetic variation is used as the statistical instrument, the instrumental variable approach is often called Mendelian randomization, because the genotype is “randomly” assigned at conception and is not influenced by a host of confounding issues (measurement and environmental factors, considered broadly) (4).

Rather than using a single genetic variant, these investigators used multiple genetic variants, associated to varying degrees with sRAGE levels, as statistical instruments. They then tested for the associations that allow for causal inference. First, the investigators confirmed that sRAGE is associated with sepsis-associated ARDS. Then, they found that sRAGE genetic instruments that were associated with greater differences in sRAGE levels were also associated with a greater effect on sepsis-associated ARDS, and genetic instruments that were associated with smaller differences in sRAGE levels were proportionally associated with a smaller effect on sepsis-associated ARDS (Figure 2 in Reference 1). Based on these findings, the investigators reasoned (by the Mendelian randomization flow of logic) that the sRAGE genetic instruments could only be plausibly associated with sRAGE levels and sepsis-associated ARDS in this way if sRAGE causally contributed to sepsis-associated ARDS.

Examining Limitations of the Methodology and Next Steps

From previous work, it is clear that the RAGE pathway modulates the septic inflammatory response (7–10). The current data build on that knowledge and support the hypothesis that the RAGE pathway causally contributes to the development of sepsis-associated ARDS in humans. But this is just a start. Was it really sRAGE? It seems that sRAGE genetic instruments would also be effective in reflecting a molecule directly upstream within the RAGE pathway that may be the true causal contributor, with sRAGE being a proportional but noncausal byproduct. In addition, as these investigators point out, this analysis supports the hypothesis of causality but does not define the underlying mechanism of causality, indicating a need for future investigations. Furthermore, it would be helpful to validate the genetic scores independently to reduce the potential for type 1 error when nearly a million variants were considered. Would the same genetic instruments predict a somewhat similar effect on sRAGE levels in separate populations, and would the same genetic instruments yield similar associations with the development of ARDS in an analysis of a completely separate population? The limited overlap of genetic variants (and even genes) identified in European ancestry and African ancestry populations highlight this concern. Effect size is an even more important issue that this statistical methodology starts to address; that is, would an intervention impacting the RAGE pathway have a sufficiently large effect to meaningfully alter clinically relevant outcomes?

A very positive outcome of this study is that sRAGE, as an intermediate phenotype, becomes a powerful tool in the route toward developing novel therapeutic interventions. A phase 2 trial of a novel intervention likely would not be powered to detect a statistically significant difference in a relevant clinical outcome. However, a relatively small phase 2 trial may be sufficiently powerful to detect a significant difference in sRAGE levels that, based on the current work, may predict clinical benefit. The simple measurement of sRAGE levels opens up the therapeutic development field to investigators who do not have the same financial resources that big pharma may have. Simple but creative ideas can also be tested; for example, do different ventilation or resuscitation strategies alter sRAGE?

Proof of Principle

Stepping outside of the RAGE pathway, this study demonstrates that the large datasets that are being developed can be powerful discovery tools pointing toward potential new therapeutic strategies (12). These valuable datasets optimally would contain clinically relevant physiological and outcome data, intermediate phenotype data (e.g., protein, gene expression, and other omic measurements), and genotype measurements (13). There is no fundamental reason why even complex multipathway approaches and precision medicine approaches (i.e., biomarker-defined subpopulations) could not be examined, so long as the study populations are very large.