Regular emergence of novel pathogens is one of the greatest threats to global health. DNA sequencing enables ‘reading’ the genetic information of viruses and microbes. However, standard approaches for pathogen detection from sequencing data can only recognize already known pathogens, as they rely on comparing the obtained DNA sequences to databases of reference genomes or genetic markers. For the same reason, detecting a hypothetical engineered pathogen and screening synthetic DNA against potential threats remain difficult.
We solve this problem by training deep neural networks that predict if a DNA read originates from potential human pathogens (bacteria, viruses or fungi) or viruses and microbes that do not infect humans. Such networks generalize to novel strains and species, allowing fast detection of previously unknown threats. This is possible also in real time, as the DNA sequencer is running. Further, we develop new methods of interpreting neural networks, visualize the learned patterns and find regions of novel genomes that are associated with elevated pathogenic potential.
The associated python package, DeePaC, support easy training and evaluation of custom models for arbitrary DNA sequence data.
Code: [DeePaC] [DeePaC-Live]
Database of fungal pathogens: [DeePaC-F]