TIMELY: Time-series Integration across Modalities for Evaluation of Latent DYnamics

What is the project about?

TIMELY is a standardized, multimodal benchmark for biological time-series data to advance modeling of complex dynamical systems. By aligning datasets, metadata, and evaluation protocols, it fosters development of statistical methods and foundation models tailored to the analytical needs of research in biomedicine and neuroscience.

What gap in the scientific community led to the creation or expansion of this benchmarking dataset?

Life science and biomedical research generates increasingly large and complex time-series datasets. Yet, limited high quality benchmarks for making these datasets accessible to the broader machine learning community exist. Existing time-series benchmarks rarely capture the complexity, noise, and multimodality of biological and health data, limiting the development and evaluation of models suited for real biological systems and scientific discovery.

What is the project’s structure — from data curation to expected outputs such as publications or competitions?

TIMELY follows an agile, seven-cycle development structure. In each cycle the project team curates, standardizes, and integrates one dataset from a new problem domain into the benchmarking package with consistent metadata and evaluation tasks. The final release will include open-access datasets, reference code, and a leaderboard. To broaden impact, we will incorporate community contributions through workshops at upcoming machine learning conferences.

In what ways does the project foster cross-domain, cross-center, or interdisciplinary collaboration?

The initial benchmark integrates datasets from seven partner groups across domains and centers. It fosters collaboration through shared data standards and co-design of the benchmark API. Community contributions will be invited via dedicated workshops at Helmholtz and NeurIPS 2026.

Image: https://scidraw.io/