Identifying protein targets for phenotype-first or mechanism-uncertain compounds is challenging because proteome-scale docking can produce thousands of structurally plausible protein–compound interactions per compound. The study develops and benchmarks a workflow that transforms ranked proteome-scale docking outputs into compound-level Gene Ontology (GO) Biological Process enrichment signatures. The core goal is to evaluate whether GO-derived signatures can reduce the candidate search space while preferentially retaining known drug–target relationships.
Docking results were retained at the PDB-chain level and then mapped to unique human gene identifiers. This mapped gene set for each compound was used as the input universe to assess overrepresentation of GO Biological Process terms using the PANTHER overrepresentation analysis against the structural gene universe that had been screened. The approach thus converts a ranked list of structural docking hits into a set of enriched biological processes associated with the top candidate proteins.
The authors evaluated GO enrichment across progressively larger top-ranked protein lists for each compound, specifically from the top 25 through the top 505 proteins in increments of 10. They then defined a stability criterion for a compound-level GO profile: a selected profile must show a Jaccard similarity of at least 0.80 across three consecutive transitions in the evaluated ranking window. This procedure yields a single, stable GO-enrichment signature per compound for downstream comparisons and prioritization.
Benchmarking used the Yamanishi drug–target interaction collection. From that resource, 682 compounds were mapped and assigned to a prespecified development/held-out split of 552 compounds for development and 130 held out for validation. The analyses evaluated biological organization and recovery at three hierarchical label resolutions: canonical (broad class), mechanistic (intermediate), and fine-mechanistic (detailed) labels. The abstract reports results for corrected full benchmarks and held-out validation; exact mapping procedures and additional dataset preprocessing details are not reported in the abstract.
In the corrected full benchmark cohort, 642 compounds produced usable canonical GO profiles. For those compounds, within-class GO profile similarity exceeded between-class similarity (within = 0.1019; between = 0.0931; delta = 0.0088), with a permutation test (100,000 permutations) yielding p = 0.00222. By multivariate analysis of variance using PERMANOVA, canonical class membership explained 1.34% of the variation in GO profiles (p < 1e-4).
At the finer mechanistic resolution, organization was stronger across the full dataset: the within-versus-between-class similarity difference increased (delta = 0.0245) and PERMANOVA R-squared rose to 0.0976, with both tests significant at p < 1e-4. These findings indicate that GO-derived profiles reflect biological organization that aligns with known mechanistic labels, particularly at more granular mechanistic definitions.
Performance on the held-out split was more modest and depended on the metric and label resolution. For mechanistic labels, PERMANOVA indicated a significant effect in the held-out set (R-squared = 0.0626, p = 0.0437). A frozen analysis using fine-mechanistic labels produced greater within-class similarity (0.1350) than between-class similarity (0.1110) with p = 0.038, and nearest-neighbor recovery was significant (p = 0.0495). However, the same frozen fine-mechanistic analysis did not reach significance by PERMANOVA (p = 0.119). These mixed outcomes highlight that held-out generalization is present but limited and sensitive to the evaluation choice.
The principal held-out search-space experiment quantified how GO-based gating affects candidate retention and recovery of gold-standard targets. This experiment evaluated 107 held-out compounds, covering 753,492 candidate protein rows, and included 424 represented gold-standard targets. When applying a direct GO gate (using direct GO associations without ontology propagation), 1.26% of candidate proteins were retained, and these retained candidates included 11.32% of known targets, corresponding to an enrichment of 8.96-fold over random expectation.
When GO associations were propagated through the ontology (i.e., using hierarchical propagation of GO terms), retention changed: 5.10% of candidates were retained while including 21.46% of known targets, corresponding to a 4.21-fold enrichment. At matched candidate-space sizes, the GO-associated prioritization approach retained a higher fraction of known targets compared with the direct GO gate: for roughly 5% of candidates, GO-associated retained 27.59% of known targets versus 22.41% for the direct gate; at 10% of candidates retention was 39.39% versus 36.08%; at 20% it was 58.73% versus 56.13%. These comparisons indicate ontology-aware propagation can broaden candidate retention while maintaining or improving retrieval of known targets at comparable candidate-space fractions.
An alternative tested was additive protein-level GO reranking, which applies GO information back onto individual protein ranks. Among 605 evaluable compounds, 21.7% showed improvement in their best known-target rank after this additive reranking. However, the overall mean reciprocal rank for known targets decreased from 0.0276 to 0.0189, indicating heterogeneous and often adverse effects on average ranking performance. Thus, while some compounds benefit from protein-level GO reranking, it is not uniformly advantageous across the benchmark.
The results indicate that GO-term enrichment derived from proteome-scale docking profiles can serve effectively as an intermediate biological search-space reduction and prioritization layer. GO gating and ontology-propagated associations substantially shrink the raw candidate set while enriching for known targets, yielding multiple-fold increases in target density within the prioritized subset. However, GO-derived scores do not consistently improve protein-level ranking across all compounds; additive reranking produced mixed outcomes and reduced mean reciprocal rank. Therefore, the authors position GO enrichment as a filter and prioritization step rather than a universal direct replacement for protein-level target scoring.
Note: the abstract reports methods, benchmark splits, and summarized results; additional methodological and implementation details beyond the abstract text were not reported in the source provided here.