AI‐Based Mitochondrial Network Characterization as a Biomarker for DNM1L ‐Related Disease Diagnosis

Pathogenic variants in DNM1L impair Drp1 function, a key regulator of mitochondrial and peroxisomal fission, leading to a neurodevelopmental disorder with a broad clinical spectrum. The clinical and functional heterogeneity of Drp1 deficiency complicates the interpretation of new variants of unknown significance, hindering the disease diagnosis and proper management. This proof‐of‐concept study evaluates whether AI‐based mitochondrial network phenotyping could serve as a characteristic feature of DNM1L mutations, supporting the disease diagnosis. Using fluorescent confocal microscopy images, supervised machine‐learning classification of the mitochondrial network features of DNM1L ‐ patients’ primary fibroblasts enables discrimination from controls and other mitochondrial disorders. To avoid manual data extraction, the deep‐learning Convolutional Neural Network (CNN) EfficientNetB0 is used, following two approaches: direct Fine‐Tuning (FT) and Curriculum‐learning Fine‐Tuning (CFT), in which the model is trained from easy‐to‐complex microscopy images to extract hierarchical visual features. Among these, CFT achieves the best performance, with 77% accuracy and 0.92  AUC (Area Under the ROC Curve), thereby optimally discerning DNM1L mitochondrial networks from other mitochondrial disorders. Overall, these findings provide proof‐of‐concept that mitochondrial network morphology in patient‐derived fibroblasts may constitute a disease‐associated phenotypic signature of DNM1L deficiency and a potentially useful complementary functional readout for the interpretation of DNM1L variants. Further validation in larger cohorts and across independent experimental settings will be required before considering its clinical diagnostic application.