ABSTRACT Introduction Current CALR mutation screening in myeloproliferative neoplasms (MPNs) faces a trade‐off between coverage and cost: allele‐specific qPCR and fragment analysis are inexpensive but cover only part of the variant landscape, while sequencing costs 300–500 per sample with 3–5 day turnaround. High‐resolution melting (HRM) offers both but suffers from subjective interpretation. We developed HRM–CALR–AI, an automated CALR screening platform. Methods HRM–CALR–AI was developed on 2667 samples and combines YOLOv8s‐cls deep learning with CatBoost meta‐classifier for open‐set recognition of untrained mutations. Validation included three independent laboratories, 188 consecutive JAK2 ‐negative MPN patients, and a head‐to‐head comparison with capillary electrophoresis fragment analysis. Results HRM–CALR–AI processed 96 samples in 90‐min at < 10 per test, with limits of detection of 6. 25% (CALR‐1) and 12. 5% (CALR‐2) variant allele frequency. On a 480‐sample held‐out set, it achieved 93. 9% accuracy (95% CI: 91. 7–95. 8), AUC of 0. 996 (95% CI: 0. 993–0. 998), and 89. 1% (57/64) recognition of untrained rare variants. Clinical screening identified 160 wild‐type (85. 1%), 15 CALR‐1 (8. 0%), nine CALR‐2 (4. 8%), and four rare variants (2. 1%), including a previously unreported variant (c. 1145₁158delinsTCCT) classified as CALR‐2 ‐like. Multicenter accuracy was 98. 1% (ICC 0. 868). Against fragment analysis, HRM–CALR–AI showed equivalent detection of length‐altering variants (kappa = 0. 933) and additionally covered nonlength‐altering variants that are invisible to size‐based separation. Conclusions HRM–CALR–AI combines open‐set recognition with low‐cost HRM to detect rare and unreported variants at ~50‐fold lower cost and same‐day turnaround. Multicenter reproducibility and benchmarking against fragment analysis support implementation as a frontline screening tool where NGS access is limited.
Zhang et al. (Mon,) studied this question.