Callability agrees with an independent catalog of short-read problem genes
Our GIAB-derived callability classification, tested against the independent Mandelker 2016 homology catalog — its high-stringency problem list, computed from reference sequence alone and checked against 30 real exome datasets, from a different decade and sharing no inputs with ours. Educational reference — not diagnostic.
Discrimination (ROC)
How well the score ranks cataloged short-read problem genes ahead of the rest.
Low-mappability of CDS · AUC 0.971Callability class · AUC 0.924
Empirical failure rate by class
Share of each class on the catalog’s high-stringency list. Bars: all 19,250 genes. Dark ticks: the 5,694-gene clinical panel.
Hover a bar for gene counts in both populations.
Read: over all 19,250 genes the classification separates the genes an independent catalog flags as short-read problems with AUC 0.924 for the five classes and 0.971 for the low-mappability fraction (0.921 and 0.973 on the clinical panel). The failure rate climbs steadily from Reliable to Refractory on the panel, but not across all genes: there, Reference-sensitive genes (40.6%) fail more often than Technology-sensitive ones (35.7%). The classes are a classification by mechanism — which fix applies — that happens to be ordered by severity inside the panel, not a severity scale. The data-driven optimal low-mappability cutoff is 7.4% across all genes.
n = 19,250 genes (1,019 on the high-stringency list), of which 5,694 are the clinical panel (195). Release v1.12, measured 2 October 2026. Callability from GIAB HG002 v5.0q / v4.2.1 & T2T-CHM13; failure labels from Mandelker 2016 (PMID 27228465), high-stringency list, all-gene scope. Rank-based AUC with mid-ranks. Absence from the catalog is treated as a negative, so AUC is a conservative lower bound.