P-KNN: Maximizing variant classification evidence through joint calibration of multiple pathogenicity prediction tools

Clinical guidelines for Mendelian disease diagnosis require that outputs from variant pathogenicity prediction tools be converted into well-calibrated probabilities. However, the existing calibration process is only valid when pre-committing to a specific tool, preventing clinicians from using multiple tools with complementary strengths. To lift this restriction, we introduce Pathogenicity K-Nearest Neighbors (P-KNN), a simple, flexible method that jointly calibrates any set of tools. P-KNN scores each variant by the fraction of pathogenic variants among those with the most similar scores across all underlying tools. We compared P-KNN to single-tool calibration over 13 real tools, including two meta-predictors. Compared to BayesDel, the best-performing tool, P-KNN produced better-calibrated probabilities and stronger evidence strength, with mean log likelihood ratios of 2.56 vs. 2.24 for pathogenic variants, 2.29 vs. 1.56 for benign variants, and 2.28 vs. 2.07 for variants of uncertain significance. We also evaluated P-KNN at four historical time points to assess scalability and robustness, finding that it improved steadily as new tools became available, while remaining well-calibrated. It also correctly integrated computational predictions with experimental measurements, whereas guidelines-based evidence summation systematically overestimated pathogenicity. In summary, P-KNN allows full flexibility to work with any set of tools in a reliable and robust manner. It is consistent with clinical variant classification guidelines, making it well positioned to improve diagnostic yield for rare genetic diseases. P-KNN is available via command line (https://github.com/Brandes-Lab/P-KNN) and precomputed scores (Dataset: https://huggingface.co/datasets/brandeslab/P-KNN, User Interface: https://huggingface.co/spaces/brandeslab/P-KNN-Viewer).