Towards thermophilic protein stability prediction: A comprehensive study of machine learning approaches

Abstract. Thermophilic proteins are critical in biology due to their enhanced thermal stability. Different machine learning approaches have been applied to estimate protein thermal stability. Current study categorizes the previous research into classification and regression tasks and explores the impact of different data representations, including tabular, sequence, and graph data on evaluation performance. Pipelines for prediction using different representations are thoroughly described. Current challenges, such as insufficient and imbalanced datasets, are addressed with potential solutions, such as transfer learning and re-sampling methods. Additionally, model interpretability, discussing various approaches to obtain model explanations and highlighting that some explanations are inconsistent have also been included in current study. Such comprehensive overview provides insights into existing methodologies and suggests potential research directions and improvements.