SuSha: A multi-model ensemble learning framework for predicting microbial salinity adaptation

Current research on microbial salinity adaptation faces substantial challenges, including the limited predictive accuracy of traditional single-gene models and difficulty in dissecting systemic biological responses to salinity stress in complex natural habitats. To overcome these bottlenecks, the multi-model ensemble learning tool SuSha, which leverages genome-wide amino acid composition features, was developed. By extracting features from the whole-genome data of 123 bacterial and archaeal species with well-defined salinity adaptations, a 24-dimensional feature vector was constructed, comprising the frequencies of 20 standard amino acids and four aggregated functional categories. Based on this, an ensemble model was developed by integrating algorithms such as random forest, bagging, and extra trees. Five-fold cross-validation demonstrated that this 24-dimensional feature-based ensemble model achieved a global accuracy of 0.765 and an area under the curve of 0.941, significantly outperforming individual baseline models. Furthermore, the model was externally validated using 2678 metagenomic samples from six global regions, encompassing freshwater, marine, and hypersaline habitats. SuSha exhibited high robustness, ecological consistency across diverse salinity gradients, and a classification accuracy of over 90% for extreme halophiles, particularly within the extreme halophilic range. By enabling high-precision genotype-to-phenotype predictions using a habitat-adaptive algorithm-switching strategy, SuSha provides a robust computational framework for inferring the physiological potential of uncultivated microorganisms and mining microbial resources in extreme environments.