Cardiogenetics and Artificial Intelligence: the Mutscore Algorithm

Abstract Funding Acknowledgements Type of funding sources: None. Backgroud The analysis of whole genome, exomes or deputed genes in the context of multigene panels, while being extensively available in routine practice, has raised new challenges in result interpretation. Novel sequence variants have become an increasingly frequent finding which, differing from the sequence of the referenced general population, remain of uncertain significance. Purpose We report our experience on a cohort of consecutive patients (pts) referred for suspicion of cardiac hereditary disease to the cardiogenetic outpatient clinic of our center. After the identification of likely pathogenic/pathogenic variants (LP/PV) and of variants of uncertain significance (VUS), we applied the new computational predictive algorithm Mutscore, based on a machine-learning approach, to test its performance and ability in reclassifying VUS. Methods We retrospectively reviewed DNA sequencing results of pts tested for suspected cardiac hereditary disease and addressed at our center from July 1st 2007 through December 31st 2021. For each LP/PV and for VUS we applied the Mutscore algorithm using a random forest approach. The Mutscore, which integrates already existing predictive algorithms to data concerning variant topographic localization, has been already validated. Of note, VUS Mutscore distribution has been analyzed according to two cut-off values (i.e. 0.14 and 0.73) identified in our previous study. In our precedent validation dataset, a Mutscore of 0.14 represented the threshold below which the 95% of variants were likely benign/benign and 0.73 the threshold above which the 95% of variants were likely pathogenic/pathogenic. VUS were then reclassified as likely benign variants when their Mutscore was <0.14 and as likely pathogenic when their Mutscore was >0.73. Results Among the 488 tested pts, a missense LP/PV was found in 126 (25.8%) pts, while 57 missense VUS were identified in 41 (8%) pts. Panel A represents Mutscore distribution among the 115 pts with missense LP/PV and a computable Mutscore. We observed an excellent predictive performance of the Mutscore in our dataset, as attested by an AUC (Area Under the Curve) of 0.895 (Panel B) and by the statistically significant positive correlation (r = 0.65, p-value = 0.000) between the Mutscore and the variant interpretation reported in ClinVar. Among the 55 pts with missense VUS and a computable Mutscore, Panel C represents Mutscore distribution according to two cut-off values (i.e. 0.14 and 0.73) identified in our previous study. The Mutscore supported the reclassification for the 47.2% of VUS into likely pathogenic/benign variants. Conclusions We showed the excellent predictive performance of the Mutscore algorithm and its accuracy in reclassifying VUS. Although further validation studies are needed, the Mutscore appears as a valuable algorithm contributing to VUS disambiguation. Artificial intelligence-based algorithms are promising tools towards precision medicine.