Environmental adaptations in metagenomes revealed by deep learning

Background: Deep learning has emerged as a powerful tool in the analysis of biological data, including the analysis of large metagenome data. However, its application remains limited due to high computational costs, model complexity, and difficulty extracting biological insights from these artificial neural networks (ANNs). In this study, we applied a transfer learning approach using the ESM-2 protein structure prediction model and our own smaller ANN to classify proteins containing the domain of unknown function 3494 (DUF3494) by their source environments. DUF3494 is found in a diverse group of putative ice-binding and substrate-binding proteins across a range of environments in prokaryotic and eukaryotic microorganisms. They present a compelling test case for exploring the balance between prediction accuracy and interpretability in sequence classification.

Results: Our ANN analysed 50,669 DUF3494 sequences from publicly available metagenomes, and successfully classified a large proportion of sequences by source environment (polar marine, glacier ice, frozen sediment, rock, subsurface). We identified environment-specific features that appear to drive classification. Our best-performing ANN was able to classify between 75.9 and 97.8% of sequences correctly. To enhance biological interpretability of these predictions, we compared this model with a genetic algorithm (GA), which, although it had lower predictive ability, provided transparent classification rules and predictors. Further in silico mutagenesis of key residues uncovered a vertically aligned column of amino acids on the b-face of the protein which was important for environmental differentiation, suggesting that both methods captured distinct evolutionary and ecological aspects of the sequences. Feature importance analysis identified that steric and electronic properties of the protein were associated with predictive ability.

Conclusions: Our findings highlight the utility of deep learning for classification of diverse biological sequences and provide a framework for combining methods to improve model interpretability and ecological insights.