In order to evaluate the performance of DeteX, we used a range of real and standard datasets as shown below:

1 The public whole genome sequencing datasets from TCGA: 11 pairs of glioma data in sam format, with average sequencing depth of 50X-100X. The variant set of Mutect2 software for these 11 samples (Supplementary Table S2) was obtained from the TCGA (https://portal.gdc.cancer.gov/) website was used as the standard variant set for subsequent analysis.

The data were obtained by the process of filtering, Bwa (.7.12-r1039) (McKenna et al., 2010) comparison, marker duplication by Picard MarkDuplicates.jar (1.119) package, re-matching by Gatk (Lai et al., 2016) and correction of base mass values to obtain the bam data of the samples. The average sequencing depth of samples ranged from 500X to 2500X.

3 The NCCL-EQA data: Illumina Hiseq platform data of NCCL-EQA in 2017 and 2019 were analyzed to obtain bam data through the same process of reads filtering, alignment, marking duplication, realignment and correction of base quality value.

DeteX starts with Binary Alignment/Map (BAM) file, which is generated from NGS reads alignment or procedures, such as BWA (Li and Durbin, 2009), TopHat (Trapnell et al., 2009; Eenst et al., 2017), Bowtie (Langmead et al., 2009), and Bowtie2 (Langmead and Salzberg, 2012; Langmead et al., 2019). It is developed in Perl language. The source code can be downloaded at https://github.com/mvlzwtd/DeteX. This website is maintained by Marvelbio O&M team.

DeteX can detect somatic and germline mutations in tumor-only or somatic mutations in tumor-normal paired samples. The detection strategy includes the following steps:

(i) Filtering reads with low base quality, low mapping quality, duplicates, multiple mapped and outside of the detection interval.

In order to reduce the number of false positive variants caused by non-independence errors and sequencing errors, we have set various filtering conditions in our software, which are derived from the experience of developers in data analysis and suggestions of other software. The filtering conditions applied in our software, including reads and variants filtering, are shown in Supplementary Table S1. These conditional values can be set by parameters except for mapping type.

Meeting all the reads filtering and the first 4 variants filtering conditions in Supplementary Table S1 and then these reads can be used to detect mutations by the following BC and FET model.

We calculate the ratio of the maximum likelihood values (LOD) of the two models for each variant locus by using the model introduced in the MuTect supplemental method (Cibulskis et al., 2013). The larger LOD value of a variant, the more reliable it is. We calculate the LOD value for each variant in the tumor sample, and filter out the variants with LOD value less than 3.9.

If there is no control sample, perform variations filtering according to the conditions 5, 6, and 7 in Supplementary Table S1, and output the final variations in VCF format. The results contain germline and somatic variants. If there is no control samples, all we need is the results of somatic mutation. We can screen out somatic mutations by dbSNP (Smigielski et al., 2000; Sherry et al., 2001; Chiang et al., 2017; Arifuzzaman et al., 2020) (https://www.ncbi.nlm.nih.gov/snp/), Cosmic (Forbes et al., 2017) (https://cancer.sanger.ac.uk/cosmic), 1000G (https://www.ncbi.nlm.nih.gov/variation/tools/1000genomes/), gnomAD (http://hgdownload.cse.ucsc.edu/gbdb/hg19/gnomAD/vcf/), EXAC (Song et al., 2016) (http://exac. broadinstitute.org) or other databases. The dbSNP, 1000G, gnomAD and EXAC databases are healthy human variant databases. If this variant exists in these databases, and the population variant frequency is greater than or equal to 1%, then this variant is considered as a germline variant, otherwise it is considered as somatic mutation. Cosmic database is database of tumor variants. If there is a variant, it is considered to be somatic mutation. If a variant is considered as both somatic and germline variant, it means that both possibilities exist.

If there is a control sample, the frequency of variation in the control sample needs to be less than the value of parameter “-nf” or .2 times frequency in the tumor sample. The variants that are not in control sample or supporting reads in the same direction are directly output. The LOD value of a variation in the control sample whose frequency is less than .5 times that in the tumor is calculated. If the LOD value is less than 3.9, the variation will be output. If it is greater than or equal to 3.9, the significance p-value of the variation is calculated according to Fisher’s exact test. If the p-value is less than or equal to the parameter setting value, the variation is output, otherwise it is not output. The p-value of Fisher’s exact test of a variation is calculated as follows: We denote supporting variant reads in normal sample as n 1 , supporting reference reads in normal sample as n 0 , supporting variant reads in tumor sample as c 1 , supporting reference reads in normal sample as c 0 . The significance value p is given by p = n 1 + n 0 ! c 1 + c 0 ! c 1 + n 1 ! c 0 + n 0 ! n 1 ! n 0 ! c 1 ! c 0 ! n 1 + n 0 + c 1 + c 0 !

Next, according to the conditions 5, 6, 7. and 8 in Supplementary Table S1, variation filtering is performed, and then the final variations in VCF format are outputted. If two adjacent loci are mutated at the same time, which shows that they are mutated in the same reads, we call this type mutation a substitution mutation. According to this features, we can correctly output these type mutations. Delete reads with more than or equal to three mutations, most of which are caused by alignment errors.

The software applies Fisher’s exact test algorithm to detect InDel. After meeting the reads filtering conditions, reads with one or two gaps/insert sequences (inss) are retained. If SNV variants exist SNV in the same reads within five bases distance to gap/ins, they are combined into one variant. If the gap or the distance among the gaps or inss in the same reads is less than 10 bases, they are also combined into a variant. If the two mutations are in the different reads, they are considered to be two variants. These conditions ensure the detection of complex mutation correctly. Variant filter conditions 5, 6, 7, 8, 9, and 10 in Supplementary Table S1 are also applicable to InDel detection.

Due to PCR amplification, duplicate reads may occur at the variant site. Duplicate reads are counted as one in variation frequency calculation. In repeated reads, if the proportion of mutated reads is greater than .8, it is considered as a read supporting variation. The maximum base quality value of mutated reads at the variation site is considered as the base quality value of this site. Only reads that cover all the InDel sites are counted in InDel frequency calculation. Final variation frequency VAF is given by V A F = t o t a l   m u t a t i o n   r e a d s t o t a l   r e a d s × 100 %

To evaluate the performance of DeteX, we used DeteX, VarScan, MuTect and VarDict to detect SNV and DeteX, VarScan, Gatk and VarDict to detect InDel. In the detection of SNVs and InDels of three groups of datasets, DeteX and Mutect were used the default value of the parameters. We added parameters of “--min-avg-qual 13, --min-tumor-freq .01, --max-normal-freq .02” to VarScan, “-filter T_INDEL_F < .01” to Gatk, “-m 5 -O 40 -V .02 -x 0 -k 0 -X 3 -c 1 -S 2 -E 3 -g 4” to VarDict.

We use sensitivity, precision and F1 value to evaluate the performance of the software. Sensitivity is the proportion of detected true variantions in all the true positive variations. The higher the sensitivity, the lower the rate of missed detection. Precision is the proportion of detected true variantions in all the observed variations. The higher the precision is, the accuracy is higher. The value of F1 is the products of sensitivity and precision, which comprehensively reflects the performance of variation detection of softwares.

Benchmark variant data is very important for assessing the performance of variant detection software. A few typical real samples containing all types of mutations, and simulation data usually can not contain all kinds of random errors in sequencing, public authority samples and standard samples can be used to evaluate software performance. In this paper, three groups of samples are selected. Each group of sample has different ways to get the benchmark variant data.

11 TCGA samples. The Mutect2 software can obtain the most reliable variant results. So we used variants detected by Mutect2 as the final benchmark variant dataset. The mutations with sequencing depth below eight and less than 2 variant reads are dropped. Finally, 760 mutations, including 37 InDels and 723 SNVs (Supplementary Table S2) were found in 11 glioma samples with Mutect2-labeled. Because substitution and complex mutations are considered as two mutations in the public mutation database, 763 variants are displayed in the table. We put the substitution mutations into the SNV set and the complex mutations into the InDel set.

28 real clinical samples. All SNVs and InDels with variant frequency ≥1% were detected. Those detected by at least two softwares were put into the final benchmark variant set. The categorization method is the same as TCGA sample variants, with substitution mutations categorized as SNV and complex mutations categorized as InDel. There are 1,481 variants (SSupplementary Table S3) including 1,045 SNVs and 436 InDel in the benchmark variant set.

NCCL-EQA external quality assessment data. There is a set of standard variables to assess the testing capabilities of laboratories across China. 94 variants are available in 2017 sample (Supplementary Table S4) and 23 variants are available in 2019 sample (Supplementary Table S5). These standard datasets enable accurately evaluate the performance of the software.

We compared the accuracy and sensitivity of each software in detecting SNVs and InDels in 11 pairs of TCGA samples (Table 1). In the table, in the SNV detection, VarScan.hc and MuTect have the best precision but at the expense of sensitivity. In contrast, VarScan has the best performance in sensitivity, but it contains a large number of false positive variants, resulting in the lowest accuracy. VarDict and DeteX show relatively balanced in both metrics, but DeteX slightly outperforms vardict in all metrics. So the score is the highest among all the software when it measured by F1 values. In the InDel detection, VarScan has a similar performance in detecting SNV with the lowest accuracy. VarScan.hc has a relatively high precision but has the lowest sensitivity. DeteX has great advantage over other software in all metrics, and it can reach a maximum F1 of one.

We also compared the results of each software for SNVs and InDels in 28 pairs of clinically targeted sequencing samples (Table 2; Figure 2). As shown in Table 2, DeteX and VarDict have similar sensitivity in SNV detection and they are significantly higher than VarScan and MuTect. The accuracy of DeteX is lower than MuTect, but higher than VarDict and VarScan.Finally DeteX has the highest F1 value. In the InDel detection, DeteX and VarDict have similar sensitivity, which is significantly higher than VarScan and Gatk. The accuracy of DeteX is moderate. The final the F1 value of DeteX is also the highest. Figure 2 shows the sensitivity and accuracy of each software in detecting SNV and InDel in 28 samples in more intuitive and detailed way. In some samples, there is no variant detected by MuTect. The precision of it in these samples is marked 0, so the status of 1 and 0 are appeared in Figure 2A. All variants marked somatic including “StrongSomatic” and “LikelySomatic” are counted in the VarDict results. Therefore, in InDel detection, the precision of VarDict performs the worst (Figure 2C). DeteX and VarDict are significantly better than the other two softwares in terms of sensitivity both in detecting SNV and InDel (Figures 2B, D).

We also analyzed the variant results of each software for the NCCL-EQA external quality assessment samples. Each software could detect SNVs accurately (Supplementary Table S6). For InDel, substitution and complex mutations (Table 4), only DeteX detected all of them correctly, VarDict detected 22 and both VarScan and Gatk detected 16.

It shows clearly in Table 3 and Figure 3 to detect substitution and complex mutations in TCGA samples by each software. From these results, it can be seen that for substitution mutations such as the variant one in Table 3 (Figure 3A), two adjacent bases are mutated in the same reads, and the standard result is chr10:28409253–28409254, CA- > AG, which was detected as two adjacent mutations by VarScan and MuTect, but accurately detected by VarDict and DeteX. If two adjacent bases are mutated, but one of them also mutated in the normal sample, it is not a substitution mutation. For example, the standard result of the variant 2 in Table3 (Figure 3B) is chr21:46074201, C- > T. VarDict detected as CA- > TG, the other softwares detected correctly. The above two cases indicate that VarScan, MuTect and VarDict have certain detection errors for the mutations occurring in adjacent sites. For complex mutations, variant three in Table3 (Figure 3C) is a deletion accompanied by a SNV variant. The standard result is chr8:145540703, GG>A.The results show that only DeteX could detect it correctly. VarDict detected it as two mutations. VarScan and Gatk detected a deletion only. MuTect did not detect the mutation. The variant 4 in Table 3 (Figure 3D) is similar to the variant 2 except that the two mutanted loci are not adjacent to each other. So the result is the same as the variant 2.

The results of NCCL-EQA (Table 4) once again confirmed the excellent detection ability of DeteX for complex mutations and substitution mutations. Only DeteX detected all of them. VarDict could accurately detect substitution mutations, while complex mutations can only be partially detected or not. The other two software could partially detect or fail to detect these two kinds of mutations. The detection ability of DeteX is excellent especially in complex mutations which contain SNV and gaps. For example, four events AAdel, TTmap, AAGAGAAGCAdel, and A- > C occurred simultaneously in the mutation chr7: 55242467–55242481 AATTAAGAGAAGCAA- > TTC (Figure 4). It is difficult to output it accurately without considering the adjacent InDel and SNV together.

We analyzed the SNVs and InDels results of DeteX and the HaplotypeCaller (DePristo et al., 2011) program in Gatk in tumor sample from the 2017 NCCL-EQA external quality assessment (Table 5). DeteX detected all the 94 variants in the standard variant set, but only 29 variation were detected by HaplotypeCaller, most of which were more than 10%. This indicates that DeteX is more sensitive than HaplotypeCaller in detecting somatic variants in a single sample.

Good accuracy is often at the expense of computing time. DeteX increases computing speed by splitting the targeted capture interval by setting the threads. We calculated the time taken by these four softwares to detect SNVs and InDels in 28 pairs of clinical samples (Figure 5A). The samples in Figure 5A are arranged from shortest to longest according to VarDict’s running time. It takes much less time than VarDict. Due to the high complexity of the algorithm, it takes slightly longer than VarScan and MuTect. DeteX processes the bam data linearly first and then it outputs the sequencing reads of the variant site in the case sample. Therefore, the more variations exist, the deeper the sequencing depth, and the longer it takes. The number of variants is not only related to the sample itself, but also to the sequencing quality. The higher the quality, the faster the speed is. Figure 5B shows the relationship between the running time of DeteX and the average sequencing depth. It can be seen that there is a high linear relationship between them, and the correlation coefficient reached .8.