Skip to content
Tech Interview Prep
Technical interview guide

Variant Calling & Genomic File Formats

The FASTQ-to-VCF pipeline: how raw reads become a filtered list of called variants, and the file formats each stage produces.

Read
50 min
Practice MCQs
25
Interview Q&A
25
Edition
v4
Editorial status
Reviewed

Scope: VCF 4.5 / BCF 2.2, SAM/BAM and CRAM 3.1 specifications; current GATK, bcftools, samtools, GIAB, vcfeval and GA4GH VRS guidance reviewed 2026-09-04.

Curated: · Written: · Reviewed:

25 questions

MCQ-2A germline pipeline discards reads whose mean base quality falls below 20. A reviewer asks what the Q20 on a single base in the FASTQ actually asserts about that base call. Which interpretation is right?
MCQ-4A read aligns with CIGAR 3M2I4M. A pipeline reports the read's 7th base (0-based read offset 6) at reference offset 6, but the alignment places it at reference offset 4. Which property of the CIGAR string did the pipeline fail to apply?
MCQ-13A pipeline converts VCF records to BED for a coverage tool. A deletion record has POS=100, REF=AT, ALT=A. Which statement about the VCF POS field gives the correct BED start?
MCQ-1A pipeline joins two single-sample VCFs on (CHROM, POS) and finds `chr1:1000` carrying REF=A in one file and REF=G in the other. Both files were written by spec-conformant tools and each validates against its own header. What explains the mismatch?
MCQ-3A sequencing lab archives 40,000 CRAM files (about 90 TB) written against GRCh38, and deletes the local copy of the reference FASTA to reclaim space. Eighteen months later a collaborator requests re-analysis, and decoding fails on every file. The lab still holds the original BAMs for only 3% of the cohort. What is the most likely cause of the failure?
MCQ-18A caller run with no filtering step emits 1,200,000 records whose FILTER column is `.` throughout, and the header defines no filter IDs other than PASS. A QC script that counts every record with `FILTER != 'PASS'` as failed reports 1,200,000 failures, and the pipeline drops the file. What is the correct diagnosis?
MCQ-10A variant caller emits, for one multiallelic site in a diploid sample, `PL` = 255,255,255,0,255,255 and `GL` = -25.5,-25.5,-25.5,0,-25.5,-25.5. The header lists ALT alleles A and T in that order. Which genotype do these likelihood fields most strongly support?
MCQ-17A phasing pipeline emits `GT=0|1` for one sample at two loci 40 kb apart, and neither record carries a PS (phase set) field. A teammate's script reads both pipes as a single chromosome-scale haplotype and reports the two alt alleles in cis. What does the pipe actually assert?
MCQ-20A pipeline reads a VCF and, for each record whose ALT contains `[` or `]`, extracts POS and the mate position from the ALT string. It discards the bracket characters and classifies the event as a deletion if the mate position is greater than POS, and as a duplication otherwise. On a validation set, a known 5 kb tandem duplication at chr5:1000000-1005000 is reported as a deletion, even though a deletion of the same region would have the same two breakpoints. What must the parser do to classify this event correctly?
MCQ-16A pipeline ingests two VCFs for one sample. The pinned reference at chr7:199-206 is G C A A A A A G (positions 201-205 are five A's). Caller X emits POS=202 REF=AA ALT=A and caller Y emits POS=204 REF=AA ALT=A for the same one-base deletion. After left-aligning and trimming against the pinned reference, what is the canonical record?
MCQ-14A pipeline normalizes a multi-allelic VCF record by splitting it into one biallelic record per ALT allele, each keeping REF and recoding the sample's genotype to the new allele indexes. The input record is `chr2 100 . C A,T . PASS GT 2/2`. In the split record whose ALT is T, what GT should the sample carry?
MCQ-15A joint-called VCF holds one multiallelic record: CHROM=7, POS=117199644, REF=C, ALT=G,T, INFO AN=125;AC=3,5;AF=0.024,0.04, with sample NA12878 carrying FORMAT AD:DP = 14,9,6:29. The header declares AC and AF as Number=A and AD as Number=R. A pipeline splits the record into two biallelic records, one per ALT allele. Which AD, AC and AF values belong on the record that keeps ALT=G?
MCQ-12You jointly genotype 300 exomes with GATK4 (HaplotypeCaller -ERC GVCF per sample, then GenomicsDBImport and GenotypeGVCFs). At chr5:1,500,000, 90 samples carry a SNV. Sample S7 has 35x coverage there and its gVCF holds a reference block with END=1,500,200 and GQ=42 spanning the position. Sample S8 also has 35x coverage, but its gVCF has no record covering chr5:1,500,000. The nearest block ends at 1,499,999. What does the joint VCF report for the two samples at chr5:1,500,000?
MCQ-24A germline SNV caller is benchmarked against a truth set of 13,000 variants inside the confident regions. It emits 12,400 calls, 11,900 of which match a truth variant. Which figure is the callset's recall?
MCQ-23A team benchmarks a caller against a GIAB-style truth set whose confident BED covers about 92% of the autosome; the excluded 8% is mostly segmental duplications and low-mappability sequence. Genome-wide the caller scores precision 0.912 and recall 0.958; inside the confident regions, precision 0.997 and recall 0.991. Which action do these numbers support?
MCQ-25A clinical lab's germline pipeline writes a run manifest containing the sample and library IDs, SHA-256 checksums of both FASTQ files and of the BQSR known-sites bundle, container image digests for BWA-MEM 0.7.17 and GATK 4.4.0.0, the full command line, and the final PASS VCF. The reference is passed as `-R /refs/GRCh38.fa` from a shared NFS mount that is re-synced nightly from an upstream mirror. Six months after a report was issued, the lab re-runs the identical command on the identical FASTQ files and gets the same PASS set plus 12 extra variants, all inside segmental duplications on chromosomes 1, 16 and 22. Which is the most likely cause?
MCQ-22A clinical lab's March 2021 run produced 4,812,000 variant records from one BAM, and its report classified 37 of them as variants of uncertain significance. In April 2025 the lab reannotates the unchanged calls against the ClinVar 2025-04 release: those same 37 alleles are now classified pathogenic, and no other record changed. An auditor asks to see both reports and to be able to reproduce each one from stored artifacts. Which storage design supports that?
MCQ-5A 30× WGS sample is aligned with BWA-MEM and called with `bcftools mpileup | bcftools call` at default settings. In a 2 kb segmental duplication, the call set contains 40 candidate heterozygous SNVs, every one supported only by reads with MAPQ 0–3; the sample's genome-wide median MAPQ is 60. The team's only post-call filter is QUAL < 20. What is the most defensible next step?
MCQ-8A 30× WGS BAM carries a true heterozygous 6 bp deletion and, 3 bp downstream on the same haplotype, a true heterozygous G>A SNP. A per-site pileup caller that does not reassemble the region reports the deletion at 0.47 allele fraction but reports the SNP at 0.04 (6 of 141 reads), below its 0.20 threshold, so it drops the SNP. HaplotypeCaller on the same BAM emits both variants at roughly 0.5. Which mechanism accounts for the recovered SNP?
MCQ-7A team sequences a non-model fish at 30× on NovaSeq. They run GATK 4.4 BaseRecalibrator with human dbSNP and Mills indel VCFs as --known-sites, then ApplyBQSR, then HaplotypeCaller. Against a truth set of 10,000 true SNVs, they call 9,600 with no BQSR (96% sensitivity) and 8,800 with BQSR (88% sensitivity); precision is 99% in both runs. They then rerun the same pipeline from the original aligned BAM, adding a fish-specific dbSNP VCF that contains 6,000 of the 10,000 true SNVs to --known-sites alongside the human VCFs. Assume the 6,000 true SNVs in the fish VCF are a uniform random sample of the 10,000 true SNVs. How many true SNVs do they now call, in expectation?
MCQ-6A tumor-only hybrid-capture panel is validated at 1,000x mean on-target depth, and the lab's accreditation requires that the reported coverage be reproducible from the BAM it delivers. The pipeline runs Picard MarkDuplicates with defaults (duplicates flagged, reads retained) into a somatic caller configured to skip 0x400 reads, plus an in-house coverage report that counts every read in that BAM. On one run, mean depth is 1,300x and 40% of reads carry the duplicate flag. To shrink the delivered BAM, an engineer inserts `samtools view -F 0x400` ahead of both consumers; the variant calls come back identical, but the coverage report now shows 780x and the sample fails the 1,000x gate. Which change should the team make?
MCQ-9A tumor-only panel (no matched normal) sequences a sample with estimated purity 30%. At a locus with local copy number 4, a somatic SNV is present on two of the four tumor copies. The variant is reported at 23% VAF. A diploid germline caller flags it as a likely artifact because it is far from the 50% VAF expected for a heterozygous germline variant. Which conclusion is correct?
MCQ-11A tumor-normal pair is called with default settings, so both samples use sample ploidy 2. At chr7:55,190,000 the tumor pileup is 20 REF reads and 40 ALT reads, and the matched normal is 30 REF with 0 ALT. A copy-number caller reports three copies at that segment in the tumor, with ALT on two of them. Which genotype does the variant caller emit for the tumor at this site?
MCQ-21A tumor-normal WGS pair is sequenced to 30×. In a 2 Mb region on chr7, tumor depth is 8× and matched normal depth is 7×; genome-wide tumor depth is 30×. The region has 28% GC and mappability 0.4. At 12 heterozygous SNPs in the region, tumor B-allele fractions are 0.48–0.52; tumor purity is 80%. The caller reports a somatic heterozygous deletion. What is the best next step?
MCQ-19An academic lab has 8 exomes from a rare-disease cohort (GRCh38, GATK4 HaplotypeCaller, joint-genotyped). The raw callset has 61,400 SNPs and 7,900 indels. The lab has the standard human training/truth resources and wants the most defensible filter for this discovery cohort now; a 200-exome replication cohort is expected in six months and will be processed through the same pipeline. Which filtering plan should they use now?