Misinterpretation and misapplication of biomarkers in inflammatory bowel disease: how do we avoid this?

ABSTRACT

Introduction

The management of inflammatory bowel disease (IBD) has evolved substantially over the past decade, with the emergence of new advanced therapies presenting unprecedented challenges in clinical decision-making. While these therapies provide patients with more opportunities to get better, biomarkers to guide their use remain elusive.

Areas covered

This article highlights the challenges associated with biomarker discovery, interpretation, and application in IBD – based on literature review, first-hand experience of biomarker discovery, and personal opinion. We highlight problems including the misinterpretation of predictive capabilities, lack of independent validation, and reverse causation in retrospective studies, and explain why associations with clinical parameters or seropositivity to microbial antigens often fail to meet the rigorous performance metrics required for clinical utility. The relative need for different biomarkers is also discussed – particularly in light of recent evidence from the PROFILE trial, which emphasizes the considerably greater risk posed by uncontrolled disease than by the potential side-effects of medications.

Expert opinion

Despite multiple challenges, the potential of biomarkers for precision medicine in IBD remains promising, particularly in combination with other clinical and biochemical parameters. Further research into combinatorial biomarker approaches is needed, but must be combined with learning how to communicate results that are inherently uncertain.

1. Introduction

Less than 10 years ago, the management of IBD was relatively straightforward. Only one class of advanced therapy was available and these drugs were simply too expensive to use in anyone who had a chance of responding to cheaper, less effective treatments. Now, however, things are much more complicated – a situation that has parallels in rheumatoid arthritis (RA), systemic lupus erythematosus (SLE) and multiple sclerosis (MS), to name just a few. Having many more advanced therapies means many more decisions have to be made, and all with little advance knowledge of which drug would be best for individual patients. Indeed, recent improvements in IBD care have been driven more by closer monitoring than by individualized care [1]. While research has therefore increasingly focused on finding more sensitive measures of subclinical disease activity – to provide better ‘targets’ for a ‘Treat-to-Target’ approach – such biomarkers, including promising candidates like serum leucine-rich α2-glycoprotein [2], are simply measures of current disease activity. For this reason, efforts to develop prognostic and predictive biomarkers continue, both to predict the future disease course and the likely outcome of potential therapy (including efficacy and side effects). In this article, we discuss the errors that are commonly made in the application and interpretation of biomarkers in IBD. For a more granular overview of available biomarkers and their strengths and limitations, please see our previous publications [3,4].

2. What is prediction?

The commonest error with predictive biomarkers is perhaps also the most surprising: thinking you have one when you don’t. Predicting anything implies that the outcome has not already happened. Thus, previously requiring bowel surgery and/or advanced therapies or having developed disease complications are not ‘predictors’ of aggressive IBD [5–7] – just as having joint deformities or renal failure are not predictors of aggressive RA or SLE. Rather, these are clear indicators that aggressive disease is already present. Association is also commonly confused with prediction, with many statistically robust associations being cited as potential predictors when it is clear that their predictive performance would actually be poor [3]. Indeed, very few reported predictors have sufficient sensitivity, specificity, positive and negative predictive values to be useful predictors. A good example of this are clinical parameters, including younger age at diagnosis, early need for steroids and perianal disease in Crohn’s Disease (CD) or extensive disease in ulcerative colitis (UC), which are still commonly used in clinical practice – perhaps because of a lack of alternatives – but which have long been known to be poor predictors [8–16].

3. Interpreting predictions

A fundamental aspect of prediction is that the projected outcome is not certain. This uncertainty – and risk in general – is often difficult for patients (and some clinicians) to understand. For biomarkers to be successfully adopted in clinical practice, the ability to understand and explain their results – and more importantly, their implications – will be essential. Even good biomarkers will appear to get it wrong in some cases (Figure 1). In these situations, the risk of misclassification should be weighed against the risk of not using the biomarker. For example, a companion diagnostic was recently reported to identify patients with a higher chance of responding to Tulisokibart, an anti-TL1A monoclonal antibody [17]. However, 76% of patients were negative for the biomarker, and the difference between the remission rate in a biomarker positive group (32%) and the unstratified cohort (without considering the biomarker; 26%) was small. As such, use of such biomarkers – to select patients for anti-TL1A therapy – would risk denying some patients the chance of receiving a therapy that could have been effective.

Figure 1.

Understanding biomarker performance.

A hypothetical prognostic biomarker can identify a cohort of UC patients who have either an 86% chance or a 15% chance of developing aggressive disease. The biomarker’s performance characteristics are therefore excellent, but 15% of patients are ‘misclassified.’ Blood vial image: NIAID Visual & Medical Arts. (10/7/2024). BDBlood Vial. NIAID NIH BIOART Source. bioart.niaid.nih.gov/bioart/52

4. Challenges in identifying reliable biomarkers

Experience in oncology has highlighted the many challenges associated with identifying reliable biomarkers [18]. Indeed, while the potential of biomarkers to deliver precision medicine is undisputed – promising prediction of diagnosis, prognosis, treatment response and overall outcomes – the reality is that most proposed biomarkers have failed to fulfill that potential. There are many reasons for this, although some common themes have emerged from previous experience.

4.1. Lack of independent validation

A leading cause of biomarker failure is the lack of adequate independent replication. This is especially important for putative biomarkers identified in ‘−omic’ studies, since the number of parameters measured in these studies (p) vastly exceeds the number of samples (n), resulting in a high likelihood of overfitting (the phenomenon by which spurious associations are detected by chance due to the sheer number of data-points considered) [19]. The situation where p >> n is a recognized challenge in classifier development and can only be confidently overcome by confirming predictive performance in independent cohorts [20]. This is not a trivial task, since recruiting new patient cohorts is both time-consuming and costly, and inherently risks that a biomarker may not validate. Indeed, even when associations do replicate, the effect size observed in replication studies is usually smaller than in the index studies in which they were first reported (‘winner’s curse’). Nonetheless, since the performance, and thus value, of unvalidated biomarkers is essentially unknown, independent validation must be a pre-requisite for all future biomarkers.

4.2. Reverse causation

To rapidly obtain large study cohorts with extended follow-up data, many investigators have opted for a retrospective approach to biomarker discovery. While this is undoubtedly easier than recruiting prospective cohorts and waiting for follow-up time to elapse, there are several risks inherent with this strategy. A major concern is reverse causation, in which an apparent predictor of an outcome is in fact the result of that outcome. For example, seropositivity to various microbial antigens has been reported to predict more complicated CD and higher need for surgery [20–23]. However, most of these studies were based on serologic testing of samples from CD patients after many years of disease (and correlation with prior disease course). Subsequent work has highlighted that seropositivity increases over time, and while there are undoubtedly some CD patients who are seropositive at diagnosis, our experience is that this is a minority – consistent with the subsequent poor predictive performance reported in attempted replication studies [24,25]. Large prospective studies of newly diagnosed CD patients will therefore be needed to determine the true utility of anti-microbial antibodies as prognostic biomarkers. In the meantime, other serological markers – including anti-αvβ6 antibodies which can precede the diagnosis of ulcerative colitis by up to 10 years – may prove more useful by facilitating prevention studies [26].

5. Using biomarkers appropriately

When interpreting biomarkers, it is important to recognize the contribution of other factors that may influence either the result itself or at least the interpretation. An obvious example is the influence of disease location on fecal calprotectin results. It is now well-recognized that fecal calprotectin levels are generally lower in patients with isolated small bowel disease compared to those with colonic involvement, leading to a reduction in sensitivity [27–29]. It is also important to consider the pretest probability prior to conducting a test, as this may alter the interpretation of the result. For example, in UC patients in symptomatic remission, it has been estimated that the overall rate of moderate or severe inflammation is approximately 15% [30]. In this setting, a fecal calprotectin result of < 250mcg/g carries a 92% negative predictive value. On the other hand, a symptomatic UC patient with rectal bleeding and an elevated stool frequency has an 85% chance of moderate or severe inflammation. Using the same calprotectin cutoff in this setting results in a value < 250mcg/g only having a 26% negative predictive value [30].

6. What are the key performance parameters?

The perfect biomarker is unlikely to ever exist. Nevertheless, there are some aspects of biomarker performance that are more important than others. Exactly which aspects are most important will vary depending on the clinical context, but this should be carefully considered when developing, optimizing and validating biomarkers. This typically requires an assessment of risk – for example, which is more dangerous for a patient: the risk of inadequately treated disease, or the potential risks of the proposed therapy? If uncontrolled disease carries a greater risk, then negative prediction is more important (i.e. reliably identifying some patients who do not need the treatment, even if this means that a few patients could be over-treated) [31]. Alternatively, if the treatment carries a greater risk than the disease, then the reverse will be true and it will be more important to identify patients who would definitely benefit, even if this does not pick up every patient in this group. In IBD, it is increasingly recognized that risks of uncontrolled inflammation outweigh those of most therapies, and so negative prediction is likely to be more important in most scenario [31,32].

7. What biomarkers do we actually need?

In recent years, we and others have highlighted the need for both prognostic biomarkers (to predict disease course) and predictive biomarkers (to identify the most suitable treatment for an individual) [4]. However, the results of the PROFILE trial have provided a new perspective on the relative importance of these aims [32]. In PROFILE – a biomarker-stratified trial designed to test a prognostic biomarker – newly diagnosed CD patients were randomized to accelerated step-up or top-down therapy (infliximab and azathioprine) after biomarker testing. While the biomarker did not show prognostic utility, there was an overwhelming benefit of top-down therapy, with higher endoscopic remission rates than have ever been reported in any Phase 3 trial. Moreover, approximately 1 in 20 patients treated with accelerated step-up required surgery for disease complications within the first year, while the only surgery required in the top-down arm was for a gallstone ileus. Notably there was also no excess of serious infections in newly diagnosed patients treated with top-down therapy – consistent with other reports that infection risk in IBD patients may be driven as much by physiological deconditioning due to longstanding uncontrolled disease than by treatments per se [33,34]. While PROFILE only enrolled patients who had both symptoms and some degree of detectable inflammation (and so cannot be generalized to patients diagnosed incidentally) the message is clear: top-down therapy should be considered the standard of care for most newly diagnosed CD patients, especially with the advent of biosimilar drugs, which mean this is no longer unaffordable in many healthcare settings. Importantly, while some clinicians will be more comfortable with adverse outcomes arising from the underlying disease, rather than from the drugs they have prescribed, it is undoubtedly true that the greatest harm to patients comes from IBD itself [35,36].

8. Expert opinion

Despite considerable ongoing research, the exact role that predictive and prognostic biomarkers will play in the future of IBD management remains unclear. Some investigators have even questioned whether predictive or prognostic biomarkers will ever be realistically implemented into clinical practice – especially following the observed lack of biomarker utility in PROFILE. While this result clearly highlights the challenges associated with developing biomarkers – and emphasizes the importance of formally testing biomarkers in stratified trials (of which PROFILE remains the only one to date in IBD) – it is worth remembering that this situation is not dissimilar to the experience in oncology. Indeed, despite oncology being frequently cited as an exemplar of biomarker success, many more cancer biomarkers have failed than have succeeded. Similar results have also been seen in other immune-mediated diseases [37]. The result of PROFILE should therefore not deter future biomarker discovery efforts, and provides a useful framework for how putative biomarkers can and should be tested. The other important lesson from PROFILE was that there was clear benefit in using top-down therapy from diagnosis in CD patients, even without a biomarker. This benefit included both disease control and avoidance of complications, and was not associated with a clear downside with respect to side effects or infection risk. This result therefore mandates a reevaluation of the relative need for different types of biomarker and – ironically – suggests that a prognostic biomarker may not be so important after all. Selecting which therapy to use first, however, remains a major unmet need, but we suspect this will not be solved with a single biomarker. Indeed, even determining whether a specific drug would be suitable for an individual patient will require both prediction of efficacy and the risk of side-effects. It is also conceivable that some biomarkers will provide more information when combined with clinical features than they provide alone. These possibilities can, and should, be formally tested to provide patients with the best chance of receiving the right treatment at the right time. Aside from challenges associated with the development of biomarkers, clearly communicating their results to patients will bring additional challenges. Like any forecast, biomarker-based predictions will not be definitive and being able to communicate the associated uncertainty will be important for maintaining doctor-patient relationships, and managing expectations.

So where will the biomarker field be in 10 years? Based on the time it takes to develop biomarkers, validate their performance, and conduct stratified trials to show whether their use would improve care, it seems possible that we may still be having the same discussions regarding the unfulfilled potential of IBD biomarkers. However, there are several promising studies that are currently underway, and – providing we can learn the lessons that have been painfully learnt in oncology – it is equally possible that we may finally be in a position to match patients to the treatment(s) that are most suitable for them.

Funding Statement

JC Lee is funded by the Francis Crick Institute, which receives its core funding from Cancer Research UK [CC2219], the UK Medical Research Council [CC2219], and the Wellcome Trust [CC2219]. JC Lee is a Lister Institute Prize Fellow.

Article highlights

• Prediction is often confused with association, but requires distinct performance characteristics.

• Prediction is inherently uncertain – recognizing and effectively communicating will be critical.

• Biomarker discovery must involve independent validation.

• Caution is needed when trying to identify biomarkers retrospectively in case of reverse causation.

• Predicting drug response and side-effects should be a greater priority than predicting prognosis in light of the results of PROFILE.

Declaration of interest

JC Lee reports consultancy and/or speaker fees from AbbVie, AgPlus Diagnostics, PredictImmune, Pfizer, Johnson & Johnson, and C4X Discovery outside the submitted work; and grants from GSK outside the submitted work. The authors have no other relevant affiliations or financial involvement with any organization or entity with a financial interest in or financial conflict with the subject matter or materials discussed in the manuscript apart from those disclosed.

Reviewer disclosures

Peer reviewers on this manuscript have no relevant financial or other relationships to disclose.