Deep ensemble and uncertainty estimation for more accurate protein localization in deeploc2.0

Abstract

Accurately predicting protein subcellular locations is critical in bioinformatics, as it plays a key role in understanding protein functions and their implications in various biological processes. Proteins operate in specific cellular compartments, and their localization is essential for proper function. However, mislocalization can disrupt cellular activities and contribute to the onset of diseases such as metabolic disorders, cardiovascular diseases, neurodegenerative conditions, and cancer. [1] Existing computational models, including deep learning-based approaches like DeepLoc 2.0, have significantly improved localization prediction by leveraging advanced feature extraction techniques. However, a major challenge remains the lack of uncertainty estimation in these models. Given the complexity of protein sequences and the potential for multi-compartment localization, there is an inherent uncertainty in predictions that current models fail to quantify. To address this limitation, we propose a method utilizing deep ensemble learning combined with entropy calculation to mitigate uncertainty in protein localization predictions. By leveraging multiple models and aggregating their predictions, we can achieve higher overall confidence in localization assignments, ensuring more robust and reliable outcomes for protein sequence classification. [2], [3]