Artificial Intelligence Based High Resolution Melting Analysis for Comprehensive Calreticulin Mutation Screening and...

Area of Science:

• Genetics
• Oncology
• Bioinformatics

Background:

• Current CALR mutation screening for myeloproliferative neoplasms (MPNs) presents a cost-coverage dilemma.
• Traditional methods like allele-specific qPCR and fragment analysis are inexpensive but limited in variant detection.
• Next-generation sequencing (NGS) offers broader coverage but incurs high costs and extended turnaround times.
• High-resolution melting (HRM) analysis provides comprehensive coverage but often involves subjective interpretation.

Purpose of the Study:

• To develop and validate HRM-CALR-AI, an automated platform for cost-effective and accurate CALR mutation screening in MPNs.
• To overcome the limitations of existing screening methods by combining the speed and low cost of HRM with advanced AI for variant detection.
• To enable same-day turnaround for CALR mutation analysis, improving clinical workflow efficiency.

Main Methods:

• Development of HRM-CALR-AI using a dataset of 2667 samples.
• Integration of YOLOv8s-cls deep learning for image classification and a CatBoost meta-classifier for open-set recognition of novel mutations.
• Validation across three independent laboratories, including 188 JAK2-negative MPN patients.
• Head-to-head comparison with capillary electrophoresis fragment analysis for performance benchmarking.

Main Results:

• HRM-CALR-AI processed 96 samples in 90 minutes at a cost under $10 per test.
• Achieved limits of detection of 6.25% for CALR-1 and 12.5% for CALR-2 variant allele frequency.
• Demonstrated high accuracy (93.9%) and AUC (0.996) on a held-out set, with successful recognition of 89.1% of untrained rare variants.
• Clinical screening identified known CALR mutations and a novel unreported variant (c.1145_1158delinsTCCT).
• Multicenter accuracy reached 98.1% (ICC 0.868), showing equivalent detection of length-altering variants compared to fragment analysis and identifying additional non-length-altering variants.

Conclusions:

• HRM-CALR-AI effectively combines open-set recognition with low-cost HRM for detecting rare and unreported CALR variants.
• The platform offers a ~50-fold cost reduction and same-day turnaround compared to sequencing.
• Demonstrated multicenter reproducibility and superior variant coverage compared to fragment analysis.
• Supports implementation as a frontline screening tool for MPNs, particularly where NGS access is limited.