Phenotype-Centric Variant Classifier (PCVC): A Machine-Learning Framework for Binary Triage of Pathogenic Variants in Rare Disease Genomics

Background: Rare disease molecular diagnosis remains bottlenecked by the need to manually review numerous candidate variants per case—up to millions per case in whole-genome sequencing (WGS) despite increasingly rich annotation resources and phenotype-aware prioritization tools. To address this operational hurdle, we developed the Phenotype-Centric Variant Classifier (PCVC), a machine-learning framework designed to convert candidate variant prioritization into a binary triage task for streamlined rapid identification of reportable pathogenic variants. Methods: The training dataset comprised 108,614 independent variant-by-case observations from 915 pediatric cases evaluated for suspected rare genetic disorders who underwent clinical whole-exome or whole-genome sequencing (WES/WGS). The dataset contained 610 clinically reported pathogenic or likely pathogenic (P/LP) variants (positive class), while variants of uncertain significance (VUSs), benign/likely benign (B/LB) variants, and non-diagnostic background variants were coded as the negative class. VUSs were grouped with B/LB variants for operational purposes, but this is not an assertion that these variants are biologically benign. Each observation was annotated across 254 features, integrating gene-level metrics, functional consequences, in silico effect predictions, population frequencies, ClinVar evidence, and Exomiser-derived pathogenicity and phenotype-matching scores. Models were trained using H2O.ai AutoML across multiple positive-class weighting strategies. External benchmarking was conducted on a GA4GH Phenopacket-derived simulated spike-in dataset (N = 60,000 observations, 279 positive variants) processed through an identical annotation pipeline. Results: Systematic evaluation of AutoML configurations established that tree-based algorithms consistently outperformed deep learning and linear models, with Extreme Gradient Boosting (XGBoost) dominating the top ranks. A leading single model (t2_weight_1_v2_no_clinvar) achieved an area under the precision–recall curve (AUCPR) of 0.872 on an internal holdout dataset (N = 21,785 observations) with a sensitivity of 0.83 (recall confidence interval/CI 0.758–0.897), 0.825 positive predictive value (PPV with 95% CI 0.745–0.895), and an F1-score of 0.828. On the external GA4GH dataset, it achieved 0.8817 sensitivity (95% CI: 0.8387–0.9145) and 0.6685 PPV. A multi-model decision-tree meta-learner identifies a high-confidence tier-1 node whose variants reached 0.9855 and 0.955 node-level PPV on the internal holdout and external GA4GH datasets respectively. Based on a previous estimation on negative cases’ variants and assuming approximately 20% positive diagnosis rate for rare disease cohort, approximately 0.5 predicted candidate variants, including 0.42 false-positive calls, would be called per case. So, integration of the PCVC into the clinical interpretation workflow may reduce candidate review burden to approximately 0.5 predicted candidate variants and 0.42 false-positive calls per case, while capturing approximately 84–96% of clinically reportable pathogenic variants in the first pass. SHapley Additive exPlanations (SHAP) analysis demonstrated that phenotype-matching metrics (EXOMISER_VARIANT_SCORE and HUMAN_PHENO_SCORE) and Maverick deleteriousness scores were primary drivers of model predictions. Conclusions: The PCVC provides a robust, phenotype-centric pathogenicity triage framework trained on expert-curated clinical labels. These retrospective results support projected workflow acceleration by the PCVC in generating actionable binary triage tiers instead of ranked candidate lists. The PCVC is thus projected to reduce manual interpretation burden while maintaining high sensitivity. Prospective validation and workflow time studies are required to confirm turnaround-time impact on rare disease clinical sequencing workflows.