Development of the miRNA signature

miRNAs induced (i.e. upregulated) with fold change >1.0 under 0.2% hypoxia in ≥2 bladder cancer cell lines were used as seed genes to derive a signature. The Cancer Genome Atlas (TCGA) bladder cancer (BLCA) mature miRNA-seq counts (n = 409) were downloaded from GDCRNATools²² and normalised to reads per million mapped (RPM) counts. TCGA BLCA mRNA data—RNA-Seq by Expectation-Maximisation (RSEM) normalised counts (n = 408)—were downloaded from Firebrowse (http://firebrowse.org). RPM and RSEM values with a pseudo-count of 1 added were log₂ transformed. There were 405 TCGA BLCA patients who had both miRNA-seq and mRNA-seq data. The 405 TCGA BLCA patients were split 70:30 into the training and test cohorts balanced for the proportion of hypoxia classifications (Supplementary Table 1). Overall survival (OS) data were available for 403 of the 405 patients and progression-free survival (PFS) data were available for 404 of the 405 patients.

The Boruta algorithm²³ was used in the training data set to identify miRNAs important in predicting hypoxic and normoxic tumours defined using the Winter 99-gene hypoxia signature²⁴ (Supplementary Methods). Spearman correlations were calculated between the expression levels of the miRNAs identified and Winter signature scores and those significant were selected for the miRNA signature. miRNA signature scores were calculated as mean (miRNAs positively correlated with the Winter hypoxia scores) − mean (miRNAs negatively correlated with the Winter hypoxia scores).

The online tool miRWalk base 2.0 was used to select miRNA targets predicted by three target prediction databases: miRWalk, TargetScan, and miRDB.²⁵ Gene enrichment analysis was carried out on the predicted targets in the MiRWalk database.