PGRanker: Enhancing Phenotype Knowledge with Graph Learning for Mendelian Gene Prioritization

Abstract

Phenotype-driven causal-gene prioritization depends on both the coverage of curated annotations and the quality of patient phenotype descriptions. Direct phenotype matching provides evidence for annotated genes, while gene networks can extend this evidence to functionally related candidates. Integrating these sources effectively, while maintaining performance under incomplete or imprecise phenotyping, remains an important challenge in Mendelian disease diagnosis. We present PGRanker, a phenotype-driven method that uses direct HPO gene support to seed patient-specific graph learning. A residual graph attention network refines this support through gene interactions, and a learned scorer combines the resulting gene representations with an attention-pooled phenotype profile to rank 21,592 candidate genes without variant data. We evaluated PGRanker against Phen2Gene, CADA and GADO on 277 diagnosed cases and a separate clinical cohort of 21 cases with sparse phenotype descriptions. On the 277-case benchmark, PGRanker achieved the highest recall at every evaluated cutoff, increasing mean reciprocal rank from 0.142 for the strongest baseline, CADA, to 0.192, a 35% relative improvement. Median causal gene rank improved from 99 to 73, and recall@1 from 9.0% to 14.0%. PGRanker also achieved the highest recall@10 area under the degradation trajectory across four controlled perturbations: phenotype removal, addition of unrelated terms, substitution with unrelated terms, and replacement by less-specific ancestors. These results support patient-specific integration of curated phenotype evidence and gene network learning for more accurate causal-gene prioritization across the tested phenotype-input conditions.