Mapping the Human E3 Ubiquitin Ligase Landscape with Metric Learning

The ubiquitin‑proteasome system (UPS) controls protein turnover, and E3 ubiquitin ligases are the key enzymes that confer substrate specificity, making them attractive drug targets despite their diversity.

To create a comprehensive view of the human E3 ligome, the authors aggregated data from eight independent sources (including E3Net, UbiHub, UbiNet 2.0, UniProt and BioGRID), initially collecting 1,448 protein entries. After cross‑checking and scoring consistency across sources, they filtered out duplicates and likely false positives, then used InterPro domain annotations (RING, HECT, RBR) to retain 462 high‑confidence catalytic E3s.

For multi‑subunit complexes such as Cullin‑RING ligases, the team manually annotated 151 adaptor proteins, 106 receptor proteins and 8 scaffold proteins, and mapped their protein‑protein interactions to define substrate‑binding modules.

Using a weakly supervised hierarchical metric‑learning framework, they computed twelve pairwise distances that capture sequence similarity (LMS, γ), domain‑architecture similarity (Jaccard, Goodman‑Kruskal γ, domain‑repeat), 3D structural similarity (AlphaFold2 TM‑score), functional similarity (GO semantic similarity across MF, BP, CC), subcellular localization and tissue‑specific expression. All distances were scaled to [0,1] and combined via weighted summation, with weights optimized by element‑centered similarity (SEC) and weak supervision.

The combined distance matrix was clustered with Ward’s minimum‑variance method and bootstrapped support, yielding an emergent hierarchy that, at a cut‑height of 0.25, partitions the 462 E3s into 13 families (10 RING, 2 HECT, 1 RBR). Each family was examined for characteristic sequence and domain features, and sub‑families and outliers were identified.

To assess functional importance, the authors performed a CRISPR‑Cas9 loss‑of‑function screen of UPS genes in cultured cells, using cell viability as the phenotype. The screen identified 53 catalytic and 32 non‑catalytic E3 components essential for survival. GO enrichment of the essential catalytic E3s highlighted nuclear functions, DNA damage response, replication and repair, underscoring their role in genome integrity.

Family‑level GO enrichment showed distinct substrate preferences, cellular locations and catalytic activities. For example, RBR members RNF14, RNF144A and PRKN preferentially bind K6‑linked ubiquitin chains, while TRIM (RING) proteins are enriched in antiviral innate‑immune pathways.

Drug‑targetability was explored by mapping known PROTACs and E3 binders onto the classified ligases. Only 16 proteins (9 catalytic E3s and 7 adaptors) have existing ligandable sites, mainly on adaptor proteins such as VHL and CRBN. Nearest‑neighbour analysis uncovered five highly similar proteins (BIRC8, RN166/181/141, UBR2) that could be repurposed for existing ligands, and small‑molecule clustering identified 20 representative clusters whose binding propensities suggest 25 additional catalytic E3s and 15 non‑catalytic components as tractable targets.

The authors argue that the multi‑scale metric‑learning framework is a transferable methodology for integrating heterogeneous omics data, and can be extended to other biological systems, spatial proteomics, chemical libraries and disease‑mechanism studies.