Environmental adaptations in metagenomes revealed by deep learning.

Area of Science:

• Bioinformatics
• Computational Biology
• Genomics

Background:

• Deep learning (DL) offers powerful biological data analysis but faces challenges in cost, complexity, and interpretability.
• Artificial neural networks (ANNs) are complex, limiting biological insight extraction.
• Domains of unknown function 3494 (DUF3494) are widespread in microorganisms, suggesting diverse ecological roles.

Purpose of the Study:

• To classify DUF3494 proteins by their source environments using a transfer learning approach.
• To explore the balance between prediction accuracy and biological interpretability in sequence classification.
• To develop a framework for combining DL with other methods for enhanced ecological insights.

Main Methods:

• Applied transfer learning with ESM-2 protein structure prediction and a smaller ANN.
• Analyzed 50,669 DUF3494 sequences from public metagenomes.
• Compared ANN performance with a genetic algorithm (GA) for interpretability.

Main Results:

• Successfully classified DUF3494 sequences by source environment (e.g., polar marine, glacier ice, rock).
• Achieved classification accuracy between 75.9% and 97.8% with the best ANN.
• Identified environment-specific features and key amino acid residues driving classification.
• GA provided transparent rules, complementing DL's predictive power.

Conclusions:

• Deep learning is effective for classifying diverse biological sequences.
• Combining DL with methods like GA enhances model interpretability and ecological understanding.
• The study provides a framework for integrating predictive and interpretable models in bioinformatics.