Expanding the coverage of regulons from high-confidence prior knowledge for accurate estimation of transcription factor activities

Key results

1. ShownIn TF knockdown and knockout data, activities inferred from CollecTRI-derived regulons separated perturbed from unperturbed TFs better than every other regulon collection tested. Most ChEA3 libraries did no better than a random baseline.

   Median AUROC 0.7 and AUPRC 0.75 using filtered KnockTF data (388 experiments, 234 TFs), higher than DoRothEA, Pathway Commons, RegNetwork and the ChEA3 libraries (adjusted p-value < 2.2×10^−16; mean t 252 and 275.1). ChEA3 libraries other than ARCHS4 did not beat a permuted CollecTRI baseline (adjusted p > 0.05). · Benchmark data; Benchmark procedure; Systematic comparison of TF activity inference…

2. ShownSigns were assigned in tiers: PubMed (PMID) evidence for each link first, then TF-level annotations, then the majority sign in the TF's regulon, then activation by default. The signed regulons inferred TF activity better than a version where every link was set to activating.

   Signs came from PMIDs for 18,345 links, TF prior knowledge for 9,226, regulon majority for 949 and default activation for 17,364, giving 83% activating vs 17% repressing links. Signed regulons beat the all-activating version (adjusted p-value < 2.2×10^−16; mean t 64.9 for AUROC, 51.6 for AUPRC). · Assigning the TF-gene mode of regulation; Construction of TF regulons from CollecTRI; Systematic comparison of TF activity inference…

3. ShownAmong collections built from curated knowledge rather than co-expression, CollecTRI-derived regulons cover the most TFs and include TFs found nowhere else. Agreement on interactions across resources is generally low.

   1,183 TFs and 45,856 signed links. Only ChEA3 ARCHS4 (1,612), GTEx (1,578) and Enrichr (1,393), which are built from co-expression or co-occurrence, cover more TFs. 47 TFs are unique to CollecTRI (vs 80 for RegNetwork, 42 for Pathway Commons). On average, 72.4% of interactions are unique to each collection. · Construction of TF regulons from CollecTRI

4. ShownThe better performance does not seem to come from a bias toward TFs with many targets. When each TF was scored only against experiments for that TF, CollecTRI performed about the same as the other networks overall and better for a few TFs.

   TFs in the benchmark had more targets (t = 4.68, 3.27, 3.84 for CollecTRI, DoRothEA, RegNetwork). Mean Pearson r between target count and absolute activity was ≤ 0.3 for all networks and lowest for CollecTRI (0.14). Of 12 TFs, REST, TP53, FLI1, NRF2F2 and SOX2 reached average median AUROC 0.84 and AUPRC 0.88; REST was classified perfectly. · Evaluation of size bias; Benchmark procedure; Systematic comparison of TF activity inference…

5. SuggestedComparing tumor with normal tissue in three CPTAC cancer types recovered known oncogenic and tumor-suppressor TF activity changes. Several TFs covered only by CollecTRI had literature support, though one contradicted its expected role.

   limma t-values for UCEC, LUAD and CCRCC gave 62 significantly deregulated TFs: MYC, the Jun/FOS families and E2F TFs went up, FOXO TFs went down. Of TFs absent from DoRothEA ABC, four of six in LUAD and four of eight in UCEC had a previously reported role. ZBTB4 appeared overactive despite its proposed tumor-suppressor role. · Case study; Case study