DiverScan: attention-guided exploratory analysis of animal behaviour across diverse comparative experiments

High-throughput tracking yields rich time-series data of animal behaviour, yet extracting biological insight remains a bottleneck, even for simple control--treatment comparisons. Existing analyses struggle to pinpoint the multiple, nonlinearly intertwined differences hidden in the data and fail to generalise across diverse recording set-ups. Here we present DiverScan, a deep-learning framework for attention-guided exploratory analysis of behavioural time series. DiverScan disentangles condition-dependent behavioural differences by pairing a loss function tailored to data exploration with an attention-branch architecture, and highlights condition-specific behavioural features visually and with explanatory text generated by a large language model. DiverScan is validated on a synthetic benchmark and on datasets from seven laboratories spanning multiple species and both single- and multi-animal behaviour. Beyond recovering known phenotypes, we demonstrate that DiverScan helps researchers generate new hypotheses across these diverse datasets. Available as Jupyter notebooks with a graphical user interface and flexible input formats, DiverScan can integrate into existing behavioural-analysis pipelines.