AHAPC: Multi-source feature fusion and ensemble learning for multiclass extremophilic protein prediction

Acidophilic, alkaliphilic, and halophilic proteins function under extreme conditions and hold great industrial value. However, their experimental identification is time-consuming and costly. Here, we introduce AHAPC, a unified computational framework for multiclass classification of extremophilic proteins. First, we construct a new benchmark dataset, TriExtrem, by combining three rigorously curated datasets. Then, we extracted two types of protein features, i.e., embeddings from pretrained protein language models (PLMs) and hand-crafted embeddings which include evolutionary descriptors derived from position-specific scoring matrix (PSSM), and sequence features, followed by the feature fusion and selection. Finally, convolutional neural network (CNN), gated recurrent unit (GRU), and bidirectional long short-term memory (BiLSTM) were used for three binary classification tasks respectively, while a multi-branch BiLSTM was adopted for the multiclass setting. Comprehensive evaluation and visualized analysis demonstrate that AHAPC achieves strong performance and provides interpretable predictions, facilitating reliable discovery of extremophilic proteins.