Concordance and differential pathway analysis

Concordance analysis was performed by first binarizing the mean expression values of either the patient tissue or EV data based on a 1.5 log expression cutoff using non-platform-corrected microarray data. If a gene has a higher than 1.5 log expression cutoff in both the patient EV  and tissue data, it was labeled as concordant. Furthermore, if a gene had a lower than 1.5 log expression cutoff in both the patient EV and tissue data, it was also labeled as concordant. The remainder of the genes were labeled as either unique to tissue or unique to EV based on which compartment had greater than 1.5 log expression cutoff. To account for replicates in this analysis, pre-treatment samples that arose from the same biological replicates were averaged. Post-treatment samples arising from the same patient were not averaged, as they arose different biological samples from differing time points. To find differential pathways that are different between patient tumors and patient EV, we used gene-set enrichment between tumors and EV samples with default parameters and GO biological processes database with the R ‘GAGE’ program¹¹. To find the canonical (C2) MSigDB¹⁵ pathways that are significantly different between responders and non-responders in both the discovery and validation cohorts, we utilized the Gene Set Variation Analysis (GSVA)¹⁶ program with default settings to generate per-patient GSVA scores (a normalized statistic summarizing enrichment relative to the entire cohort analogous to ssGSEA scores) across our platform-corrected discovery and validation cohorts datasets. We then used a Mann-Whitney U-Test to test for differential GSVA scores between responders and non-responders. Similar to the rationale used for DEG analysis, we utilized a nominal p-value cutoff of 0.1 to flag differential pathways (Fig. 2a and Fig. 3a). A pathway was considered validated if it achieved significance in both the discovery and validation cohorts.