Smart variant filtering 

Variant filtering consists of preserving highly confident variants and removing falsely called variants. Secondary genomic DNA analysis is mainly oriented toward alignment and variant calling because these two processes strongly influence overall quality. Previously, the variant filtering step was mostly overlooked or analyzed only in deeper testing. However, variant filtering can boost precision of variant calls substantially. Here, we created a Smart Variant Filtering (SVF) framework. Conceptually, the SVF framework has three phases: (i) selecting a locally-optimal machine learning algorithm configuration for the Genome In A Bottle variant-called samples (HG001-HG005); (ii) learning parameters for that configuration with a training set; (iii) using learned parameters to perform variant filtering on novel datasets. SVF is available on Github (https://github.com/sbg/smart-variant-filtering) and also as a Public project (https://igor.sbgenomics.com/u/sevenbridges/smart-variant-filtering) on the Seven Bridges Platform. It is open-sourced and free to use by any party (BSD-3 license). Phase (i) included brute-force testing across 372 different algorithm and parameter configurations. It included 10-fold, automatitized cross-validation using 123,000 variants. Based on these results we selected a locally optimal classifier and configuration (Multi Layer Perceptron with 250 nodes in the hidden layer). Phase (ii) trained the network selected in the prior phase with 25 million variants to learn the network weights (model) to be applied in Phase (iii). We will show results from deep, 3-stage testing to demonstrate that SVF  outperforms standard variant filtering solutions currently used within most secondary DNA analyses. Smart Variant Filtering increases the precision of called SNVs (removes false positives) for up to 0.2% while keeping the overall f-score higher by 0.12-0.27% than in existing solutions. Indel precision is increased by up to 7.8%, while the f-score increase is in range of 0.1 to 3.2%.