DNA2Graph enables automated identification of non-linear DNA molecules in electron microscopy
Electron microscopy of spread and rotary-shadowed DNA molecules provides a direct readout of DNA structure and remains uniquely informative for studying replication, recombination and repair intermediates. However, quantitative EM analysis is limited by the rarity of biologically informative structures and by the time required for expert operators to inspect very large numbers of molecules. Automated image acquisition and stitching have increased the scale of EM datasets but have shifted the main bottleneck from image collection to image analysis. Standard image segmentation tools and machine-learning approaches fail to faithfully preserve molecular continuity in EM images of DNA molecules contrasted by rotary shadowing. Here we present DNA2Graph, an open-source software for segmentation of DNA molecules from EM images. Rather than treating segmentation as a purely pixel-level task, DNA2Graph represents each molecule as a spatial graph of nodes and edges. It applies dedicated graph-based error-correction algorithms that repair signal interruptions and spurious connections introduced during segmentation, while enforcing the biological priors of continuity and thinness. DNA2Graph classifies molecules as linear or non-linear, thereby converting large imaging datasets into focused lists of candidate structures for operator review. We validated DNA2Graph on two genomic DNA datasets: a structure-poor, non-enriched human sample and a structure-rich yeast sample. DNA2Graph reduced the number of molecules requiring operator review by 45-fold and 7-fold, respectively. In addition, DNA2Graph-assisted review slightly improved the operators structure-detection sensitivity compared to unaided image review. The software also enables automated length measurement of individual DNA molecules and generates machine-readable outputs for downstream quantitative and computational analyses. DNA2Graph does not require training on manually annotated data, can run on a personal computer and can be adapted across experimental preparations through interpretable parameters.