Intrinsic sequence evidence cannot reach a pathogenic classification for missense variants under a partitioned ACMG framework

Background. Missense variants in rare-disease genes are predominantly returned as variants of uncertain significance. Under the American College of Medical Genetics and Genomics / Association for Molecular Pathology framework, the criteria a laboratory can apply to a missense variant from its sequence and annotations alone are few and individually weak, and every criterion capable of closing the remaining distance to a pathogenic classification requires functional, segregational, case-control or phenotypic evidence. The criterion encoding phenotypic specificity presupposes a model of disease that a classifier reasoning solely over sequence lacks. How large that shortfall is has not been measured within a framework that refuses to count any item of evidence more than once. Results. DISCERN couples a calibrated classification engine to a likelihood-ratio model that discriminates ten clusters of confusable diseases, within a joint posterior in which each criterion is assigned to exactly one factor. Evaluated in inherited bleeding and platelet disorders on the expert-panel surface of the ClinGen Evidence Repository, blinded to ClinVar, it reached an area under the receiver operating characteristic curve of 0.939 (95% confidence interval: 0.906–0.967) on missense variants and 0.927 under a time-split, statistically indistinguishable from REVEL, the predictor supplying its own computational criterion. Out-of-sample expected calibration error was 0.017. The evidence partition achieved complete criterion coverage across 38,802 expert-applied codes and prevented the reuse of evidence that would inflate 33.2% of variants in a coupled model. A treatment-divergence interlock flagged every contraindicated management plan and none of the harmless ones. On intrinsic sequence evidence alone, the framework assigned a pathogenic classification to 0 of 425 missense variants, reaching at most 3.0 of the 6 points required; restoring every criterion that a partitioned pipeline withholds or cannot annotate recovers only 89 variants. Conclusions. The ceiling follows from the point arithmetic of the guidelines, not from any implementation, and holds wherever they are applied. Reducing the uncertain fraction requires evidence of a different kind, not a better predictor.