DeepTRACE brings flexible machine learning to single-molecule track analysis.

Single-molecule imaging was developed to resolve behaviours obscured by ensemble averaging, but early tracking experiments typically captured only brief temporal windows, restricting analysis to individual states rather than the progression between them. Observation times now extend to minutes, revealing complete multi-stage biological processes that require new analytical approaches to capture sequences of events. Here we present DeepTRACE, a flexible tool for analysing single-molecule tracks in living cells that learns sequences of molecular events using past and future context from subcellular location, mobility, and photometric properties. It learns any molecular behaviour that can be annotated with natural-language labels, enabling users to tailor models themselves to specific biological questions without ML expertise. DeepTRACE generalises rapidly from very small datasets, training in minutes on a few hundred tracks, and supports extensive downstream analysis, including discovery of relationships absent from the training data. As DeepTRACE natively handles any numerical feature outside of its standard feature set, it incorporates photometric readouts, including measurements of internal conformation that reflect molecular action, alongside motion, temporal context, and subcellular location. We anticipate that researchers will use DeepTRACE to define biological states by molecular behaviour rather than mobility alone in complex multi-stage processes.