Intelligify (United Kingdom)

Abstract Background Biologic drugs such as anti-TNFα antibodies are predominantly used in a step-up setting and prescribed for Crohn’s disease (CD) patients who have failed, or are contraindicated for, first-line therapies. However, patient response to biologics is highly variable, with only a minority benefitting from a strong and sustained clinical response. Presently, there is no operationally effective means in clinical care to stratify patients on individual biologic drugs. Consequently, standard of care for those with CD remains inadequate and inefficient. To identify protein biomarkers associating with treatment outcomes, we performed an untargeted proteomics analysis of baseline serum samples from individuals with moderate to severe CD starting anti-TNFα treatment. Methods Biobanked serum samples from 47 (42 from Amsterdam, 5 from Oxford) moderate to severe CD patients starting anti-TNFα treatment (adalimumab, infliximab or their biosimilars) with documented 12 and 26-week treatment response were used for untargeted serum proteomic analysis using SomaLogic’s SomaScan® technology platform. Data was checked for outliers using robust principal component analysis. Supervised machine learning (ML) was used to predict treatment outcome. Results There were 29 responders and 18 non-responders to anti-TNFα treatment (week 26). Untargeted proteomics profiles passing SomaLogic’s quality controls were generated for all samples with 7,596 proteins detected in each sample. Robust principal component analysis for anomaly detection did not reveal any substantial outliers (see Figure 1). Five different ML algorithms independently generated mean AUCs and balanced accuracies from 5x 5-fold cross-validation ranging from 0.47 to 0.53 (see Table 1) indicating no reliable signal in the data. Notably, applying the ML algorithms to detect sample origin, samples from the 2 different clinical sites (Oxford or Amsterdam) could be confidently distinguished suggesting careful protocol alignment for serum collection is essential. Of 26 proteins previously reported to be associated with biologic treatment outcome, only PF4 was found to be significantly different between anti-TNFα responder and non-responders (P=0.0043). Overall, these metrics all fall well below what would be needed for a predictor to have value in the clinic. Conclusion ML analysis of serum proteomics profiles for 47 CD patients did not identify any strong and reliable biomarkers of biologic treatment response suitable for clinical applications. This small study suggests serum proteins are not suitable for anti-TNFα response prediction in this population. In addition, the serum proteomic profile is highly susceptible to differences in serum collection and storage methods.