Comprehensive protein datasets and benchmarking for liquid-liquid phase separation studies

Proteins self-organize in dynamic cellular environments by assembling into reversible biomolecular condensates through liquid-liquid phase separation (LLPS). These condensates can comprise single or multiple proteins, with different roles in the ensemble's structural and functional integrity. Driver proteins form condensates autonomously, while client proteins just localize within them. Although several databases exist to catalog proteins undergoing LLPS, they often contain divergent data that impedes interoperability between these resources. Additionally, there is a lack of consensus on selecting proteins without explicit experimental association with condensates under physiological conditions (non-LLPS proteins or negative proteins). These two aspects have prevented the generation of reliable predictive models and fair benchmarks.

In this work, we use an integrated biocuration protocol to analyze information from all relevant LLPS databases and generate confident datasets of client and driver proteins. We introduce standardized negative datasets, encompassing both globular and disordered proteins. To validate our datasets, we investigate specific physicochemical traits related to LLPS across different subsets of protein sequences and benchmark them against 16 predictive algorithms. We observe significant differences not only between positive and negative instances but also among LLPS proteins themselves. The datasets from this study are available as a website at https://llpsdatasets.ppmclab.com and as a data repository at https://doi.org/10.5281/zenodo.15118996 .

Our datasets offer a reliable means for confidently assessing the specific roles of proteins in LLPS and identifying key differences in physicochemical properties underlying this process. Moreover, we describe limitations in classical and state-of-the-art predictive algorithms by providing the most comprehensive benchmark to date.