Fig. 1

a Workflow to obtain discriminant transcript isoforms and predictive models. Given two patient groups, we subsampled two equal sized subsets, one from each group (e.g. metastatic and non-metastatic), which were compared using information-based measures, denoted as I _iso . At each iteration step, the group labels were randomized to obtain an expected measure, denoted as I _rand . After 100 iterations, two distributions were produced for each isoform corresponding to observed (I _iso ) and expected (I _rand ) values. Transcript isoforms with a difference of mean PSI values >0.1 in absolute value between the two patient groups and with a positive difference of the means of the observed and expected distributions for all information-based measures used were then considered as discriminant, which were then used to evaluate enriched cancer hallmarks. Discriminant isoforms were further filtered for redundancy with a Correlation Feature Selection strategy to build a predictive model, which was evaluated using cross-fold validation (see “Methods”). b Enriched hallmarks in the set of discriminant isoforms for each stage class, metastasis (M), tumor size (T), lymph-node involvement (N), and overall staging (S), using all isoforms selected across all tumor types. c Enriched hallmarks for each tumor type using all discriminant isoforms selected across all stage classes in each tumor type independently. d Accuracies of the classifiers for each tumor type for the T, N, M, and S annotation, given as the distributions of the areas under the receiving operating characteristic (ROC) curves (AUC). The variation on each bar indicates the minimum and maximum AUC values. Some models are absent due to lack of sufficient samples (Table 1). e PSI distributions for the transcript isoforms of IDO1 in PRAD, SYK in SKCM, and GAS7 in OV, for the N-, M-, and S-models, respectively (Wilcoxon test p values <0.001)