P-KNN: joint calibration of multiple pathogenicity prediction tools streamlines variant classification.

Joint calibration of multiple pathogenicity predictors: each variant is placed in a multidimensional space of tool scores and its probability of pathogenicity is estimated from the proportion of pathogenic neighbours.

Clinical classification guidelines require converting pathogenicity predictor outputs into well-calibrated probabilities, but the existing calibration method is only valid when pre-committing to a single tool, preventing laboratories from using tools with complementary strengths. P-KNN lifts this restriction by jointly calibrating any set of tools: each variant is represented in a multidimensional space defined by tool scores, and the probability of pathogenicity is estimated from the proportion of pathogenic neighbours. Compared with standard single-tool calibration of multiple predictors and meta-predictors at four historical time points, P-KNN performs better on overall evidence strength and on the alignment of calibrated probabilities with true pathogenicity frequencies. Evidence strength keeps improving as newer tools are added, and the method correctly integrates correlated computational and experimental evidence that existing protocols overestimate. Command-line code and precomputed scores are publicly available.

Synthesis written by Geno'X. For the full original abstract, please refer to the source publication.

This is precisely the bottleneck laboratories face: having to choose a predictor before seeing the variant, when tools diverge by gene and substitution type. The handling of correlation between computational and functional evidence is the most practically useful contribution, since this is where classification frameworks double-count the same information and artificially inflate evidence strength. The limitation is structural: the whole calibration depends on reference variant sets, and a neighbourhood-based method is only as unbiased as the density of classified variants around the variant being assessed.

Analysis by Dr Thibaut Benquey