Advancing the accuracy of clathrin protein prediction through multi-source protein language models

Abstract

Clathrin is a key cytoplasmic protein that serves as the predominant structural element in the formation of coated vesicles. Specifically, clarithin enables the scission of newly formed vesicles from the plasma membrane’s cytoplasmic face. Efficient and accurate identification of clathrins is essential for understanding human diseases and aiding drug target development. Recent advancements in computational methods for identifying clathrins using sequence data have greatly improved large-scale clathrin screening. Here, we propose a high-accuracy computational approach, termed PLM-CLA, to achieve more accurate identification of clathrins. In PLM-CLA, we leveraged multi-source pre-trained protein language models (PLMs), which were trained on large-scale protein sequences from multiple database sources, including ProtT5-BFD, ProtT5-UR50, ProstT5, and ESM-2. These models were used to encode complementary feature embeddings, capturing diverse and valuable information. To the best of our knowledge, PLM-CLA is the first attempt designed using various PLM-based embeddings to identify clathrins. To enhance prediction performance, we utilized a feature selection method to optimize these fused feature embeddings. Finally, we employed a long short-term memory (LSTM) neural network model coupled with the optimal feature subset to identify clathrins. Benchmarking experiments, including independent tests, showed that PLM-CLA significantly outperformed state-of-the-art methods, achieving an accuracy of 0.961, MCC of 0.917, and AUC of 0.997. Furthermore, PLM-CLA secured outstanding performance in terms of MCC, with values of 0.971 and 0.904 on two existing independent test datasets. We anticipate that the proposed PLM-CLA model will serve as a promising tool for large-scale identification of clathrins in resource-limited settings.