Prot-SCL: State of the art Prediction of Protein Subcellular Localization from Primary Sequence Using Contrastive Learning

Protein subcellular localization is a critically important parameter to consider when designing expression constructs and production strategies for industry scale protein production. In this study, we present Prot-SCL an innovative self-supervised machine learning approach to predict protein subcellular localization exclusively from primary sequence. The models herein were learned from a dataset of subcellular localizations derived by exhaustively analyzing the Uniprot database. The set of localization data was rigorously curated for machine learning by employing group sampling following clustering of the protein sequences. The novel component of this approach lies in the development of a triplet neural network architecture capable of generating meaningful embeddings for classification of protein subcellular localization. We observed a robust predictive power for our classical gradient boosted machine learning models trained on these triplet embeddings in both cross validation and in generalization to the testing set. Importantly, we have made this extensive dataset of protein subcellular localizations publicly accessible, facilitating future, need-based, localization studies. Finally, we provide the relevant codebase to encourage a wider adoption and expansion of this methodology.