DeepSaltPro: Enhancing halophilic protein prediction accuracy and efficiency via multi-protein language model...

Halophilic proteins refer to a class of biomolecules capable of maintaining function under high-salt conditions, playing significant roles in industrial catalysis, environmental adaptation, and various biotechnological applications. Accurate identification of halophilic proteins remains a challenge but is crucial for understanding their functional mechanisms and harnessing their potential applications. However, experimental methods are often time-consuming, labor-intensive, and impractical for large-scale analysis. This highlights the urgent need for efficient computational approaches. In this study, we propose a deep learning framework named DeepSaltPro for predicting halophilic proteins. Firstly, we extract feature embeddings from two pre-trained protein language models (PLMs), i.e., Ankh and ESM-2, to capture both local and global sequence information simultaneously. Then, these features are processed through separate convolutional neural networks (CNNs) and integrated via a bidirectional gated recurrent unit (BiGRU) to model sequential dependencies. To further capture complex nonlinear interactions, we employ a Kolmogorov-Arnold Network (KAN), which utilizes learnable spline-based transformations instead of fixed linear weights, thereby enhancing both the representational capacity and interpretability of the model. Finally, we evaluate the performance of DeepSaltPro on benchmark datasets by means of five-fold cross-validation and independent test. On the independent test set, DeepSaltPro achieved the overall accuracy of 97 %, representing a 12 % improvement over the existing state-of-the-art method HPClas. These results demonstrate that DeepSaltPro could serve as an effective computational tool for identifying halophilic proteins, offering valuable insights into their functional mechanisms and facilitating future industrial and biotechnological applications.