GARFIELD-NGS: Genomic vARiants FIltering by dEep Learning moDels in NGS

Exome sequencing approach is extensively used in research and diagnostic laboratories to discover pathological variants and study genetic architecture of human diseases. Even if present platforms produce high quality sequencing data, false positives variants remain an issue and can confound subsequent analysis and result interpretation. Here, we propose a new tool named GARFIELD-NGS (Genomic vARiants FIltering by dEep Learning moDels in NGS), which uses deep learning algorithm to dissect false and true variants in exome sequencing experiments performed with Illumina or ION platforms. GARFIELD-NGS consists of 4 distinct models tested on NA12878 gold-standard exome variants dataset (NIST v.3.3.2): Illumina INDELs, Illumina SNPs, ION INDELs, and ION SNPs. AUC values for each variant category are 0.9267, 0.7998, 0.9464, and 0.9757, respectively. GARFIELD-NGS is robust on low coverage data down to 30X and on Illumina two-colour data, as well. Our tool outperformed previous hard-filters, and calculates for each variant a score from 0.0 to 1.0, allowing application of different thresholds based on desired level of sensitivity and specificity. GARFIELD-NGS processes standard VCF file input using Perl and Java scripts and produces a regular VCF output. Thus, it can be easily integrated in existing analysis pipeline. GARFIELDNGS is freely available at https://github.com/gedoardo83/GARFIELD-NGS. 2 . CC-BY-NC 4.0 International license not peer-reviewed) is the author/funder. It is made available under a The copyright holder for this preprint (which was . http://dx.doi.org/10.1101/149146 doi: bioRxiv preprint first posted online Jun. 14, 2017;