A comprehensive benchmark of sequence-based subcellular localization predictors for human proteins

Knowledge of the spatial distribution of proteins at a subcellular level is essential for understanding protein functions, interactions, and cellular mechanisms. Dozens of methods have been developed to computationally predict protein subcellular locations from various kinds of information about a protein including amino acid sequence, but it is far from a solved problem. Most methods predict only one location per protein (i.e. single-label prediction) and neglect the fact that about half of all human proteins localize to multiple cellular compartments, where they potentially perform multiple functions. Recently, there have been a few methods proposed to perform multi-label protein localization prediction. While the performance seems promising, the evaluation is limited to around 5% of the human proteome and a small number of coarse-grained major cellular compartments. The existing multi-label protein localization prediction methods only utilize protein features like amino acid sequence and protein structure, and do not take advantage of other available data modalities like RNA expression levels and protein-protein interactions (PPI), thereby potentially falling short of the best performance we can achieve. We will perform a comprehensive analysis on existing computational models on the task of multi-label protein subcellular localization on a new 4-fold larger (4k proteins) highly validated benchmark dataset. In an explorative analysis, we will augment existing state-of-the-art sequence-based predictors with protein-protein interaction data to assess how the performance changes. We believe that being able to accurately predict cellular localization from sequence is a key development necessary to underpin the use of generative AI for sequence-to-protein-image cell modeling and this comparative analysis will help provide a clear picture of where we are at on the important task of multi-label protein location prediction and how we can best improve from here.