DeepSEA: an alignment-free explainable approach to annotate antimicrobial resistance proteins.

Area of Science:

• Bioinformatics
• Computational Biology
• Genomics

Background:

• Antimicrobial resistance (AMR) is a critical global health challenge, exceeding the impact of diseases like HIV and malaria.
• Current AMR surveillance relies on genomic comparisons with high similarity thresholds (>95%), leading to significant false-negative rates due to incomplete reference databases.

Purpose of the Study:

• To develop a deep learning approach for accurate identification and classification of antimicrobial resistance (AMR) proteins.
• To overcome the limitations of traditional sequence alignment methods in AMR surveillance.

Main Methods:

• A convolutional neural network (CNN) was trained to distinguish AMR proteins from non-resistance proteins.
• The CNN was designed to classify proteins into nine distinct resistance classes.
• The model's internal states were analyzed to understand feature importance in protein classification.

Main Results:

• The CNN model achieved high recall values (> 0.9) across all tested protein classes, surpassing alignment-based methods.
• The model successfully identified key protein domains involved in antimicrobial inactivation.
• An open-source bioinformatics tool, DeepSEA, was developed for annotating AMR proteins.

Conclusions:

• Deep learning, specifically CNNs, offers a superior alternative to sequence alignment for AMR protein surveillance.
• The developed tool provides valuable insights into AMR mechanisms and aids in tracking resistance.
• Accurate annotation of AMR proteins is crucial for effective global health strategies.