DAVE: how to use explainable AI to interpret missense variants for genome diagnostics based on functional protein modeling

Abstract

<ns5:p>Background Diagnostic yield in NGS genome diagnostics is constraint by the high fraction of variants of uncertain significance (VUS), largely due to poor interpretability of missense variation. Current pathogenicity predictors often provide strong performance but lack mechanistic insight. This study introduces MOLGENIS Digital Approximation of Variant Effects (DAVE), a supervised learning model designed to predict and explain missense variant pathogenicity using twelve biophysically grounded features. These features capture changes in protein folding energy, hydrophobicity, electrostatics, and interactions with ligands, nucleic acids, and other proteins. Results Trained on curated Dutch diagnostic data, DAVE delivers robust accuracy while decomposing predictions into interpretable feature contributions. Performance was benchmarked against updated clinical classifications and external databases, with selected variants analyzed through structural modeling. Applied to over eleven thousand VUS, DAVE showed high concordance with subsequent reclassifications and highlighted variant specific molecular mechanisms, such as altered folding stability, binding pocket properties, and interaction surfaces. Structural visualizations demonstrated how mechanistic insights can aid interpretation of VUS. Conclusions By integrating predictive accuracy with explainable outputs, DAVE offers a practical approach to prioritize VUS, generate testable hypotheses, and support informed clinical decision-making. All source code and data required to reproduce annotations and analyses are available. The DAVE results for the selected VKGL missense VUS are also available through an interactive dashboard at https://dave.molgeniscloud.org/.</ns5:p>