An algorithm for deriving new combinatorial biomarkers based on ridge regression

Motivation: Combinatorial biomarkers are considered more specific and sensitive than single markers in medical diagnostics and prediction, yet even detection of such these combinatorial biomarkers requires deep computational analysis. The principles of analytic combinatorics, linear and kernel ridge regression, and machine learning were applied to derive new combinatorial biomarkers of muscle damage. Results: Lactate, phosphate, and middle-chain fatty acids were most often included into biochemical combinatorial markers, while the following physiological parameters were found to be prevalent: muscle isometric strength, H-reflex length, and contraction tone. Several strongly correlated combinatorial biomarkers of muscle damage with high prediction accuracy scores were identified. The approach — based on computational methods, regression algorithms and machine learning — provides a flexible, platform independent and highly extendable means of discovery and evaluation of combinatorial biomarkers alongside current diagnostic tools. Availability: The developed algorithm was implemented in Python programming language on a quantitative dataset comprising 23 biochemical parameters, 37 physiological parameters and 3,903 observations. The algorithm and our dataset are available free of charge on GitHub. Supplementary information: Supplementary data are available at Journal of Bioinformatics and Genomics online.