DETECTING GENETIC ENGINEERING WITH A KNOWLEDGE-RICH DNA SEQUENCE CLASSIFIER

Detecting evidence of genetic engineering in the wild is a problem of growing importance for biosecurity, provenance, and intellectual property rights. This thesis describes a computational system designed to detect engineering from DNA sequencing of biological samples and presents its performance on fully blinded test data. The pipeline builds on existing computational resources for metagenomics, including methods that use the full set of reference genomes deposited in GenBank. Starting from raw reads generated from short-read sequencers, the dominant host species are identified by k-mer analysis. Next, all the sequencing reads are mapped to the imputed host strain; those reads that do not map are retained as suspicious. Suspicious reads are de novo assembled to suspicious contigs, followed by sequence alignment against the NCBI non-redundant nucleotide database to annotate the engineered sequence and to identify whether the engineering is in a plasmid or is integrated into the host genome. Our initial system applied to blinded samples provides excellent identification of foreign gene content, the changes most likely to be functional. We have less ability to detect functional structural variants and small indels and SNPs produced by genetic engineering but which are more difficult to distinguish from natural variation. Future work will focus on improved methods for detecting synonymous recoding, used to introduce watermarks and for compatibility with synthesis and assembly methods, for using long read sequence data, and for distinguishing engineered sequence from natural variation