Image, sequence and interactome based prediction of subcellular localization of proteins

Knowledge of subcellular localization (SL) of proteins is essential for drug development, systems biology, proteomics, and functional genomics. Due to the high costs associated with experimental studies, it has become crucial to develop computational systems to accurately predict proteins' SLs. With different modes of biological data (e.g., biomolecular sequences, biomedical images, unstructured text, etc.) becoming readily available to ordinary scientists, it is possible to leverage complementary types of data to increase both the performance and coverage of predictions. In this study, we propose HoliLoc, a new method for predicting protein SLs via multi-modal deep learning. Our approach makes use of three different types of data (i.e., 2D confocal microscopy images, amino acid sequences, and protein protein interactions – PPIs) to predict SLs of proteins in a multi-label manner for 22 different cell compartments using protein language models, graph embeddings and convolutional and feed forward neural networks. The system was trained in an end to-end manner, and the performances were calculated on the unseen hold-out test dataset. The average test performance of individual models (each using a single data type) was 0.18 (macro F1-score) and 0.55 (accuracy), whereas for HoliLoc (the fusion of 3 modalities) it was observed to be 0.26 (F1-score) and 0.60 (accuracy), indicating the effectiveness of the multi-modal learning approach proposed. According to our comparison against state-of-the-art SL predictors, HoliLoc displays highly competitive performance. HoliLoc is distributed as an open-access programmatic tool, which is anticipated to benefit life science researchers by reducing the cost and time required for wet-lab experiments by accurately predicting the SLs of the protein of interest in advance.