Customized de novo mutation detection for any variant calling pipeline: SynthDNM

Abstract

Motivation

As sequencing technologies and analysis pipelines evolve, de novo mutation (DNM) calling tools must be adapted. Therefore, a flexible approach is needed that can accurately identify DNMs from genome or exome sequences from a variety of datasets and variant calling pipelines.

Results

Here, we describe SynthDNM, a random-forest based classifier that can be readily adapted to new sequencing or variant-calling pipelines by applying a flexible approach to constructing simulated training examples from real data. The optimized SynthDNM classifiers predict de novo SNPs and indels with robust accuracy across multiple methods of variant calling.

Availabilityand implementation

SynthDNM is freely available on Github (https://github.com/james-guevara/synthdnm).

Supplementary information

Supplementary data are available at Bioinformatics online.

1. Introduction

Genome wide analysis of de novo mutation by whole genome sequencing (WGS) is being applied in a variety of settings, from small studies of individual families (Kong et al., 2012) to cohorts of several thousand genomes (An et al., 2018; Zhou et al., 2019). Thus, there is a need for variant classification methods that can detect DNMs in a variety of datasets with high accuracy. When new sequencing protocols, datasets or data processing pipelines are introduced, the process of assembling new training data is labor intensive.

2 Materials and methods

We have developed SynthDNM, a software package written in python that enables optimal DNM detection in any new dataset or variant calling pipeline by rapid optimization of the classifier using real data. SynthDNM requires a Fam and VCF file as input.

3. Classifier

The default SynthDNM classifier is designed for GATK variant calls on Illumina sequence data. In addition, SynthDNM can be re-trained for any new dataset or data type by re-implementing the training procedure with a new VCF file containing jointly-called variants on trio families. SynthDNM creates a large set of simulated DNMs from real variants by swapping parents and offspring between families (Fig. 1A). When private inherited variants in offspring are matched to the surrogate (unrelated) parents that have homozygous-reference genotypes, they are classified as ‘synthetic DNMs’. This ensures that simulated DNMs have the same properties as other ultra-rare germline variants. Lastly, since the majority of putative DNMs in the original trio represent errors (either a false-positive variant in the offspring or a false-negative genotype in a parent), all putative DNMs detected prior to the swap are assigned to the negative training set. The training workflow is described in detail in Supplementary Materials. The default classifier of SynthDNM, was trained on a jointly-called VCF from >30× Illumina sequence from the Simplex Collection (SSC) WGS dataset (SSC-GATK classifier).

Fig. 1.

(A) The methods of generating ‘synthetic DNMs’. (B) ROC curves of the SSC-GATK classifier, Triodenovo and DenovoGear. All classifiers were predicted on the same 3787 individuals. (C) A comparison of DNM prediction performance of the different calling methods

4. Performance results

SSC-GATK DNM calls intersecting with exons were validated using a set of DNMs detected by Whole Exome Sequencing (WES) of the same samples (Iossifov et al., 2014). We then compared the accuracy of the SSC-GATK classifier to DNM calls made with DeNovoGear and TrioDeNovo on the same SSC data. The accuracy of SSC-GATK was improved (AUC = 0.997) compared to DeNovoGear (AUC = 0.978) and TrioDeNovo (AUC = 0.971) (Fig. 1B).

In addition, we compared the quality of SSC-GATK calls to DNMs reported on the SSC WGS dataset in two previous studies (Turner et al., 2017; Werling et al., 2018). In this case, ‘false positive’ DNMs were defined as exonic DNM calls in the WGS dataset that were not detected in a previous exome sequencing study. This definition of false discovery rate (FDR) is an overestimate, since a subset of non-validated DNMs may be true positives, but this metric allows us to assess accuracy of the call sets relative to each other. The true positive rate (TPR) and F1 of SSC-GATK (0.977 and 0.802) is higher than for the other two call sets (Werling: 0.828,0.771; Turner: 0.308,0.402), which suggests that a single optimized method can provide greater sensitivity and overall accuracy compared to the intersection of multiple methods or set of stringent filtering heuristics.

In recent years, a number of new variant-calling platforms have been introduced (Poplin et al., 2018). Each uses a unique set of algorithms for evaluating variant and genotype quality, and for these our default GATK-based classifier is not compatible. We therefore evaluated the reproducibility of DNM calls for multiple SynthDNM classifiers trained on variants from different pipelines applied to a single dataset, including GATK (SPARK-GATK), DeepVariant (SPARK-DV) and weCall (SPARK-WC) all of which were applied to the SPARK exome sequencing study (Feliciano et al., 2019). Customized synthDNM classifiers for the three platforms made consistent DNM predictions (Fig. 1C). For SNPs, 82.6% (N = 8277) of variants overlapped across all three platforms; For indels, 94.5% (N = 432) of variants overlapped across all three platforms. To further confirm the accuracy of our synthDNM calls, we compared the de novo SNP and indel calls on the SPARK dataset to a set of validated DNMs that were confirmed by Sanger sequencing in the same cohort in a previous pilot study (Feliciano et al., 2019) to test our classifier. For SNPs, Spark-GATK could recall 112/114 (98.2%) de novo snps; Spark-DV could recall 109/117 (93.2%) de novo snps; Spark-WC could recall 100/108 (92.6%) de novo snps. For Indels, Spark-GATK could recall 112/113 (99.1%) de novo indels; Spark-DV could recall 84/84 (100%) de novo indels; Spark-WC could recall 68/69 (98.6%) de novo indels.

5. Conclusion

Optimized SynthDNM classifiers detect de novo SNVs and indels with high accuracy across multiple datasets and variant calling pipelines. In addition to the classifers trained for this publication, the synthDNM method for creating custom classifiers is readily available to users.

Data Availability

WGS data from the SSC are available from the Simons Foundation Autism Research Initiative (SFARI) (https://www.sfari.org/resource/autism-cohorts).