A data-driven, knowledge-based approach to biomarker discovery: application to circulating microRNA markers of colorectal cancer prognosis

Outline of the method. a The construction steps of the miRNA-mediated regulatory network: (1) miRNA target genes (TGs) that are either validated experimentally or predicted by two different data sets were retrieved using multiMiR which is an R package providing access to 11 publicly available data sets. Transcription factor (TF) targets were retrieved from ORTI database which compiles validated mammalian TF-TG interactions from six public data sets as well as the literature. The miRNA-mediated regulatory network was constructed using a recursive algorithm described in Supplementary Figure S3. (2) The network was then annotated using 339 CRC-associated genes identified by MalaCards; 35 ‘elite’ genes with strong causal associations with CRC progression were ranked ‘1’ and the rest of CRC genes were ranked ‘2’. (3) Using the annotated network, a functional relevance (FR) score was calculated for each miRNA (using Eq. (1)) and a look up table was returned to be used in the subsequent biomarker discovery. b FR calculation on an example network. c Schematic view of the proposed multi-objective optimisation-based biomarker discovery workflow: The pre-processed samples were partitioned to validation and discovery sets using fivefold cross-validation. The multi-objective optimiser was run on discovery set where objectives are prediction errors and averaged FR scores of the population of putative signatures. Optimal miRNA signatures (i.e., Pareto front solutions) and their corresponding predictive models were then used to classify test samples and the performance measures were reported. The whole process repeated for 50 times to account for random partitioning of samples and the average performance measures were reported (Fig. (Fig.33