AMAnD: an automated metagenome anomaly detection methodology utilizing DeepSVDD neural networks

The composition of metagenomic communities within the human body often reflects localized medical conditions such as upper respiratory diseases and gastrointestinal diseases. Fast and accurate computational tools to flag anomalous metagenomic samples from typical samples are desirable to understand different phenotypes, especially in contexts where repeated, long-duration temporal sampling is done. Here, we present Automated Metagenome Anomaly Detection (AMAnD), which utilizes two types of Deep Support Vector Data Description (DeepSVDD) models; one trained on taxonomic feature space output by the Pan-Genomics for Infectious Agents (PanGIA) taxonomy classifier and one trained on kmer frequency counts. AMAnD's semi-supervised one-class approach makes no assumptions about what an anomaly may look like, allowing the flagging of potentially novel anomaly types. Three diverse datasets are profiled. The first dataset is hosted on the National Center for Biotechnology Information's (NCBI) Sequence Read Archive (SRA) and contains nasopharyngeal swabs from healthy and COVID-19-positive patients. The second dataset is also hosted on SRA and contains gut microbiome samples from normal controls and from patients with slow transit constipation (STC). AMAnD can learn a typical healthy nasopharyngeal or gut microbiome profile and reliably flag the anomalous COVID+ or STC samples in both feature spaces. The final dataset is a synthetic metagenome created by the Critical Assessment of Metagenome Annotation Simulator (CAMISIM). A control dataset of 50 well-characterized organisms was submitted to CAMISIM to generate 100 synthetic control class samples. The experimental conditions included 12 different spiked-in contaminants that are taxonomically similar to organisms present in the laboratory blank sample ranging from one strain tree branch taxonomic distance away to one family tree branch taxonomic distance away. This experiment was repeated in triplicate at three different coverage levels to probe the dependence on sample coverage. AMAnD was again able to flag the contaminant inserts as anomalous. AMAnD's assumption-free flagging of metagenomic anomalies, the real-time model training update potential of the deep learning approach, and the strong performance even with lightweight models of low sample cardinality would make AMAnD well-suited to a wide array of applied metagenomics biosurveillance use-cases, from environmental to clinical utility.