Jakub Mikolajczak
Pronouns: he/him
Research Mentor(s): Yongqun He
Research Mentor School/College/Department: / 0
Program:
Authors: Jakub Mikolajczak, Yongqun He
Session: Session 4: 1:40 pm – 2:30 pm
Poster: 78
Abstract
Fast automated protein localization is crucial to medicinal and ontological research. This project aims to develop a deep learning method for predicting subcellular locations of microbial proteins in particular, including bacterial proteins and unicellular parasite proteins. Our dataset is sourced from UniProt, which includes many reviewed and unreviewed microbial protein sequences in its TrEMBLE dataset. Our downstream supervised learning task uses only reviewed protein sequences, but the upstream self-supervised transformer model used both reviewed and unreviewed protein sequences. For the unicellular parasite protein subcellular localization prediction, we focused on Plasmodiidae and Trypanosomatidae. Note that Plasmodiidae have only a few hundred manually annotated proteins despite having hundreds of thousands of proteins on file. This lack of manual annotations demonstrates a need for a specialized tool developed for unicellular parasites specifically. The model uses features extracted by ProtBert (URL: https://github.com/sacdallago/bio_embeddings), an adapted natural language processing (NLP) tool, which are fed into a downstream multilayer perceptron model (URL: https://github.com/jeighkoub/protein_loc) for final classification. The model relies on sequence information only, and does not consider homologues in existing knowledge bases. Our bacterial and parasite protein subcellular localization predictions achieved the performance of 72% and 42%, respectively. We are continuously improving the performance. Our unicellular parasite protein prediction model is the first of its kind, making it uniquely suited for the study of these parasites and the development of treatments for the diseases they cause.
