DeePaC: predicting pathogenic potential of novel DNA with reverse-complement neural networks.

Area of Science:

• Genomics
• Bioinformatics
• Computational Biology

Background:

• Emerging infectious diseases and the potential for engineered pathogens necessitate advanced detection methods.
• Traditional pathogen detection relies on known organism databases, failing to identify novel or unmapped sequences.
• Machine learning offers a promising avenue for inferring pathogenicity from limited sequence data.

Purpose of the Study:

• To develop and evaluate DeePaC, a novel deep learning framework for classifying bacterial pathogenicity.
• To assess the performance of different neural network architectures, including convolutional neural networks (CNNs) and LSTMs, for pathogen detection.
• To improve the accuracy of pathogen identification, particularly for unknown or novel sequences.

Main Methods:

• Implementation of a flexible deep learning framework (DeePaC) utilizing convolutional neural networks and LSTMs.
• Incorporation of reverse-complement parameter sharing within neural network architectures.
• Integration of predictions from paired-end sequencing reads to enhance classification accuracy.

Main Results:

• DeePaC, employing CNNs and LSTMs, surpasses existing state-of-the-art methods in pathogen classification.
• The deep learning approach significantly outperforms sequence homology and traditional machine learning methods.
• Integrating predictions from read pairs nearly halves the error rate compared to previous methods.

Conclusions:

• Deep learning, specifically DeePaC, offers a powerful and accurate solution for open-view pathogen detection.
• The framework demonstrates superior performance in identifying pathogenic bacteria, even from unmapped sequences.
• DeePaC represents a significant advancement in computational methods for microbial threat assessment.