Establishing a many-cytokine signature via multivariate anomaly detection.

Establishing a cytokine signature associated to some medical condition is an important task in immunology. Increasingly, large numbers of cytokines are used for signatures, via lists of reference ranges for each individual cytokine or ratios of cytokines. Here we argue that this common approach has weaknesses, especially when many different cytokines are analysed. Instead, we propose that establishing signatures can be framed as a multivariate anomaly detection problem, and hence exploit the many statistical methods available for this. In this framework, whether or not a given subject's profile matches the cytokine signature of some condition is determined by whether or not the profile is typical of reference samples of that condition, as judged by an anomaly detection algorithm. We examine previously published cytokine data sets associated to pregnancy complications, brain tumours, and rheumatoid arthritis, as well as normal healthy control samples, and test the performance of a range of anomaly detection algorithms on these data, identifying the best performing methods. Finally, we suggest that this anomaly detection approach could be adopted more widely for general multi-biomarker signatures.