DeePaC models for novel fungal pathogens and real-time detection of multiple pathogen classes

A collection of DeePaC ResNet models for

1) pathogenic potential prediction for novel fungal species (input Illumina read length: 250bp)

2) real-time detection of novel bacterial, viral and fungal pathogens (input Illumina read length: 25-250bp). Those models assume four classes: non-pathogens (i.e. commensal bacteria or non-human viruses), pathogenic bacteria, human-infecting viruses, and human-infecting fungi. Two alternative models are provided: we recommend either using the 'log' model (for faster inference), or an ensemble averaging predictions of both models (for better results).

See the code and manual at https://gitlab.com/dacs-hpi/deepac. Model weights for the fungal (-fun-) and multi-class (-multi4-) models in .h5 files and config .ini files.

The models were trained on read sets hosted here: https://zenodo.org/record/5713153 based on a curated database of pathogenic fungi (https://zenodo.org/record/5711852).

See also the preprint: https://www.biorxiv.org/content/10.1101/2021.11.30.470625v1