PTAC: An Interpretable Dual-Path Framework for Temperature-Adaptive Protein Discovery
Abstract Reliable prediction of protein temperature adaptation can facilitate enzyme discovery and engineering. However, most existing AI-based methods rely on a single protein language model and inadequately capture sequence patterns across temperature-adaptation classes. Here, we developed PTAC, a dual-path feature-fusion framework that formulates temperature adaptation as a three-class classification task through integrating a general protein language model ProtT5 and a temperature-related encoder ProPrime. On the benchmark datasets, PTAC achieved an overall accuracy of 94.9% and substantially improved recognition of the under-represented psychrophilic class, boosting F1 score to 93.3%, significantly outperforming the reported methods (70.7%). PTAC exhibits strong generalization across remotely homologous and previously unseen sequences, as well as across distinct protein property prediction tasks, maintaining an accuracy of over 80% for halophilic and acidophilic proteins. Interpretability analyses identified sequence features associated with temperature adaptation and deepened our understanding of environmental adaptation mechanisms. Moreover, PTAC achieved 92.0% average accuracy in a practical secreted-protein screening scenario. These results establish PTAC as an effective computational framework for screening temperature-adaptive proteins and mining target enzymes.
Authors
- Guimin Zhang (ORCID: https://orcid.org/0000-0001-6674-5820)
- Jiayu Wang
- Ella A. Yi
- Wenya Wang
- Junchao Shen
Institutions
- Beijing University of Chemical Technology (CN)
Publication Details
- Journal
- Research Square
- Published
- 2026-11-13
- DOI
- https://doi.org/10.21203/rs.3.rs-11572915/v1
- Primary Topic
- Machine Learning in Bioinformatics
- Type
- preprint