Mit Protein Transformer: Identification Mitochondrial Proteins with Transformer Model

Abstract

Mitochondrial proteins carry out unique physiological functions within subcellular localization regions, making accurate subcellular localization prediction essential in uncovering pathological disease mechanisms and guiding drug development. Mitochondria serve as the energy production centers of cells, comprising four major subcellular localization regions, namely matrix, outer membrane, inner membrane, and intermembrane space. While traditional research methods analyze physical and chemical protein sequence properties, coupled with machine learning or deep learning algorithms, such methods necessitate considerable time and effort in data preprocessing. To overcome such challenges, this study proposes a novel approach to efficiently predicting mitochondrial protein subcellular localization by perceiving semantic information of protein sequences directly, using an ESM2-35B pre-trained model based on transformer architecture. The study utilized four datasets, comparing two models - the Transformer-Encoder Only model trained from scratch and the classification predictor centered on ESM2-35B pre-trained model fine-tuning. Results show that fine-tuning the large pre-trained model presents a superior performance in subsequent mitochondrial protein subcellular localization tasks in comparison to the Transformer-Encoder Only model. In conclusion, ESM2-35B pre-trained model based on transformer architecture offers vast application prospects in addressing mitochondrial protein subcellular localization prediction issues.