predNMD: prediction of nonsense-mediated mRNA decay for improved clinical variant pathogenicity classification

The clinical consequence of a stop-gain variant depends on whether it ablates protein production by triggering nonsense-mediated mRNA decay (NMD) or yields truncated protein with residual, dominant-negative, or gain-of-function activity. This distinction is the key branch-point in applying PVS1, the strongest ACMG/AMP pathogenic evidence criterion. Current PVS1 guidelines and proposed SVC v4.0 successors risk unwarranted evidence assignment by employing only the 50nt rule for NMD prediction, which is central but incompletely captures NMD biology. We developed predNMD, a random forest classifier trained on 5,304 nonsense variants from GTEx, TCGA, GEUVADIS, and GREGoR. Feature selection reduced 166 candidates to 20 final features, 7 new to NMD prediction, including m6A density and TranslationAI. For variants predicted to not trigger NMD, predNMD infers the likely protein truncation. predNMD predictions aligned with deliberate clinical decisions where ClinGen Variant Curation Expert Panels (VCEPs) drew on evidence leading them to diverge from standard decision tree. predNMD agreed with VCEP on 6 of the 8 stop-gain variants for which they declined full PVS1, though the variants were predicted to trigger NMD by the 50nt rule and in loss-of-function disease genes. In leave-one-chromosome-out cross-validation predNMD reached AUC=0.79, and on independent test set outperformed 50nt rule (AUC 0.78 vs 0.65), nearly doubling discriminative signal above random (0.28 vs 0.15). predNMD likewise effectively discriminated NMD targets on BRCA1 and BARD1 saturation genome-editing data, where RNA abundance reflects NMD. These results support replacing the 50nt rule in clinical variant classification with predNMD, available as precomputed predictions covering all 13,968,776 possible stop-gain SNVs in GRCh38, installable codes, Docker image, and at predNMD.org.