From $19, paid once and yours to keep. See what your file reads first. Check your file →
DeutschFrançaisEspañolPortuguês Sample report Support Log in
aimosti Check your file

Learn Under the hood

A missing line in a VCF is not a clean result

A variant file lists the places where a genome differs from the reference, and most software assumes the person matches it everywhere else. That is usually right. On real files we have watched it fail in three different ways.

Key takeaways

  • In a plain VCF a missing line can mean the reference base or a position never read. Two real 30x plain VCFs confirmed 0.16 and 0.17 percent of the bases our report examines[4].
  • A gVCF's reference blocks say where the caller looked. Our report counts a position as examined only with a called genotype at a depth of 10 or more.
  • Depth is not a call: blocks with 30 reads and the genotype ./. once made our coverage look complete.
  • GATK's CombineGVCFs blanks every genotype and keeps the likelihoods[6]; where they clearly favour one genotype, the report now restores it.
  • At rs4430796 GRCh37 has G and GRCh38 has A[9, 10], so a missing record there means opposite genotypes in the two assemblies.

A typical genome differs from the reference sequence at 4.1 to 5.0 million sites[1], out of about 3.1 billion bases in one set of the reference chromosomes[2]. A VCF is built to list those differences and nothing else: one line per position where the caller found something[3]. So almost every position in the genome has no line at all, and the file leaves the reader to decide what that silence means.

Silence has two meanings

When a plain VCF has no line at a position, the person may carry the reference base there on both copies, which a caller has no reason to write down. Or the sequencing may not have read the position well enough to decide anything, which a plain VCF cannot write down either. Both look like a gap between one record and the next. At the eye-colour marker rs12913832 the reference base is A, the brown-associated letter, so a silent file reads as brown eyes (what that marker does).

For a common position in a good 30x genome the first reading is far more likely, which is why most software assumes it. The trouble is that a plain VCF never lets you check, and for one gene the silence is the result itself: reading blood type from DNA follows it. We cut two real 30x plain VCFs to the 25.5 million bases our report examines and counted the bases each one positively confirmed[4].

0.17%

of the bases our report examines that a real 30x plain VCF positively confirmed; a second confirmed 0.16%[4]

99.0%

of the same bases confirmed by a complete 30x gVCF, which records where the reads matched[4]

Both plain VCFs held more than 41,000 variants in those regions, about what a good genome should hold. Nothing was wrong with them as files. A plain VCF only ever says where something differs.

What a gVCF adds, and the gate we put on it

A gVCF keeps the variant lines and fills the gaps between them with reference blocks: records that cover a run of positions where the reads matched the reference, with an END field marking where the run stops and a placeholder alternative allele, <*> in the specification and <NON_REF> in GATK's files[3]. Each block carries the read depth and a genotype quality. Here is the public 30x gVCF of the reference sample NA12878 at rs4430796 on chromosome 17, a site this article comes back to.

gVCF

#CHROM  POS       REF  ALT          INFO          FORMAT               NA12878
chr17   37738039  A    <NON_REF>    END=37738046  GT:DP:GQ:MIN_DP:PL   0/0:30:81:29:0,81,1215
chr17   37738047  G    <NON_REF>    END=37738048  GT:DP:GQ:MIN_DP:PL   0/0:30:78:29:0,78,1170
chr17   37738049  A    G,<NON_REF>  DP=29;...     GT:AB:AD:DP:GQ:PL:SB 0/1:0.570:16,12,0:28:99:333,0,472,381,508,889:8,8,5,7
chr17   37738050  A    <NON_REF>    END=37738056  GT:DP:GQ:MIN_DP:PL   0/0:29:81:28:0,81,1215
Four consecutive records from the 1000 Genomes Project's 30x gVCF of NA12878 on GRCh38, with the empty ID, QUAL and FILTER columns left out and the variant's INFO field shortened[5]. The third record is rs4430796: 16 reads show A and 12 show G, and the caller writes 0/1.

The blocks either side say that every position from 37,738,039 to 37,738,056 was read about 30 times and matched the reference, apart from the one that did not. A plain VCF made from the same calls would keep only the third line. In the 200,000 bases around rs4430796, NA12878's gVCF has a record over every base[5].

Table 1. NA12878's 30x gVCF across chr17:37,600,001 to 37,800,000 (GRCh38), counted by us
What the file holdsCount
Records32,814
Reference blocks32,520
Records that propose an alternative allele294
Of those, genotyped as a variant (0/1, 1/1, 1/2 or 2/3)242
Bases covered by some record200,000
Bases with a called genotype at depth 10 or more199,840

Source: Counted on 10 October 2026 from the public IGSR file, reading the region through its index[5]. That a plain VCF of this window would hold about 242 lines is our estimate.

A block saying 0/0 is still only as good as the reads behind it, so our report does not take the genotype field on trust. A position counts as examined only when the record covering it has a called genotype, anything but ./., and a depth of at least 10 reads. A gene counts as examined when at least 90 percent of its bases pass; below that its positions stay labelled inferences.

Aligned reads get the same floor. Deep Read measures depth across the panel from the reads themselves, so a gene examined from a BAM and one examined from a gVCF mean the same thing. A plain VCF never earns the word.

Figure 1. How the report treats a position that has no variant record, by file type.

Depth is not a call

On 1 September 2026 a customer's whole-genome gVCF, almost 4 GB, produced a report that looked complete and was wrong. Its blocks carried a depth around 30 and the genotype ./. at every position the report queried. Our coverage code then counted any block with enough depth as covered. About 99 percent of the polygenic-score positions looked covered while not a single variant was observed, so every score sat at the very bottom of its range, and the coverage panel listed 1,042 genes as examined.

Both errors turned "we could not tell" into "we looked and you are ordinary". The fix was one condition: a record whose genotype is ./. contributes nothing to coverage, whatever its depth. The same file then measured zero callable bases, and a gVCF in that state is now reported as unreadable, section by section, with no common genotype shown as a result.

Three fields this article relies on

GT
The genotype: 0/0 for two reference copies, 0/1 for one of each, 1/1 for two copies of the first alternative, and ./. when no call could be made[3].
DP
Read depth: how many reads covered the position for this sample[3].
PL
Phred-scaled likelihoods, one per possible genotype, rounded to whole numbers; the most likely genotype scores 0[3]. A gap of 20 between the best and the second best means the best is about 100 times more likely.

The file that was never finished

That diagnosis was incomplete. Five days later we read the file's header properly. It named GATK 3.7's HaplotypeCaller, run one chromosome at a time, and then CombineGVCFs, the tool that merges gVCFs so that a later step, GenotypeGVCFs, can genotype them. CombineGVCFs replaces every genotype with a no-call and keeps the likelihoods; the GATK 3.7 code builds each merged genotype from no-call alleles and carries the PL and AD values over[6]. The later step was never run. What shipped as the finished gVCF was an intermediate file.

By default CombineGVCFs reverts all genotypes to no-calls, but calls can be made if specified

GATK 4.6.2.0 source code, CombineGVCFs.java[7]

The genotypes had never been written, and the evidence for them was intact. In the part of the file our panel covers there were 2,417,662 records, every one reading ./., and 2,338,326 of them had at least 10 reads and a PL margin of 20 or more. Among the records that proposed a real alternative allele, 52,718 favoured a non-reference genotype by that margin: 33,753 heterozygous and 18,965 homozygous.

Same record, genotype blanked

#CHROM  POS       REF  ALT          FORMAT               NA12878
chr17   37738049  A    G,<NON_REF>  GT:AB:AD:DP:GQ:PL:SB ./.:0.570:16,12,0:28:99:333,0,472,381,508,889:8,8,5,7
The NA12878 record from above with only its genotype replaced by ./., the shape every record in a CombineGVCFs file has. This line is our edit for illustration, not a record from any file[5]. The PL still puts 0/1 at 0 and the next genotype at 333.

We added a narrow pass that restores genotypes from likelihoods before anything else reads the file.

What the restoring pass does to each record

  1. Only a blank genotype. A called genotype is never touched.
  2. Only a clear winner. The PL must have a single lowest value at least 20 below the runner-up. A tie, a missing PL or a count of values that fits no ploidy leaves ./. in place.
  3. Never the placeholder allele. If the winner names <NON_REF> or <*>, an allele nobody observed, the record stays a no-call.
  4. Nothing else changes. Every other field is kept byte for byte, and one header line records the pass.

The pass runs only when the file's own records show the shape: every genotype blank, and enough decisive PLs to prove the evidence is there. One called genotype in the window closes the gate, since a file that genotypes at all has real no-calls. A header naming CombineGVCFs is neither required nor enough.

Our first version asked for half of the first 2,000 records to have a decisive PL. A week later we tested it on a real whole-genome gVCF whose genotypes we had blanked, and it failed: a whole genome opens near the chromosome 1 telomere, where blocks rest on one or two reads.

Figure 2. Share of records with a decisive PL at the head of a real whole-genome gVCF whose genotypes we blanked, measured on 13 September 2026. The 2,000-record window, the one the gate used, sat at 49 percent: one point under the threshold, on a file that was nothing but recoverable calls.

The gate now counts instead of taking a share. It needs at least 100 blank genotypes and at least 50 decisive PLs among the first 2,000 records. A file that is blank and undecided throughout, a real no-call file, falls short of 50 and keeps the unreadable treatment from the previous section.

The reference moved

Even a perfect plain VCF only says "matches the reference", which leaves the question of which one. GRCh38 changed thousands of single bases from GRCh37, including 8,248 sites its authors judged to be errors or rare alleles[8]. Where the base differs, a GRCh37 file is silent exactly where a GRCh38 file is not.

We found this on 22 September by comparing one donor's GRCh37 whole genome with the same person's chip exports. All four chips that carried rs4430796 read G/G. The whole genome had no record there, and after the lift our pipeline credited two copies of A. The chips were right and the whole genome was wrong. When we re-ran that genome with the fix, every score containing a site like this moved, and the one score without any, breast cancer, did not move at all.

Table 2. One site, three genotypes, two assemblies: what a VCF writes at rs4430796
GenotypeVCF called on GRCh37 (reference G)VCF called on GRCh38 (reference A)Copies of A
G/GNo record1/1, alternative G0
A/G0/1, alternative A0/1, alternative G1
A/A1/1, alternative ANo record2

Source: Reference bases from NCBI's GRCh37 and GRCh38 chromosome 17 sequences[9, 10]; the genotype rows follow from the VCF genotype coding[3].

Our checks missed it. The script that fetches reference bases verifies that the base is one of the two alleles a score names, and at rs4430796 the GRCh38 base is the effect allele. The coordinates were right too, so our liftover guard had nothing to flag.

Figure 3. The same silence read against two references. Only the reference the file was called against gives the right genotype.

The scoring variants now carry both reference bases, fetched from Ensembl by rsID, and a GRCh37 file's silence is read against the GRCh37 one. Today the two assemblies disagree at 22 of the 1,699 variants in our seventeen main polygenic scores. Where a variant's GRCh37 base cannot be pinned down, or the file's assembly is unknown at a site where the two disagree, the variant is left out of the score, and the score's coverage line counts it as unread.

What the report says, file by file

Each section's count line in the report states which kind of reading it is.

Table 3. What a position without a variant record becomes in an Aimosti report
FileWhat silence can meanWhat the report shows
Chip exportNot on this array; every probe has a row, reference or not.A readable reference row counts as read. A missing probe, a no-call row or an A/T or C/G site that cannot be oriented is not examined.
Plain VCFReference, or never read."Read from your file: X of N", the rest labelled an inference. No gene is called examined; a polygenic score is a rounded estimate.
gVCFReference, if a called block at depth 10 covers it; otherwise unread."Examined: X of N" for genes 90 percent called. A gVCF whose blocks call nothing is reported as not read.
BAM or CRAM (Deep Read)No records; the reads are counted.Depth from the reads at the same floor of 10; pharmacogenes, CYP2D6 copy number included, called from them.

Source: Aimosti's report wording and coverage rules as of 10 October 2026. The measured plain-VCF and gVCF figures are on our data page[4]; record conventions from the VCF specification[3].

The difference shows most in the medication section. A plain VCF is read for 9 pharmacogenes, a gVCF for 25, because a gVCF's blocks let a missing defining position count as reference, and Deep Read calls 29 from the reads. The reason for that line is a measurement: run on a variant-only call set of the public sample HG002, the star-allele caller printed TPMT *1/*1, Normal Metabolizer, with none of the gene's 45 defining positions present in the file.

Chips had their own version of the problem. A chip export has a row for every probe, but a homozygous-reference row produces no variant record once it is translated, and until 30 August our chip coverage was counted from the records. On a real 18-file corpus that understated every polygenic score's coverage by 30 to 75 percent of its true value and suppressed 97.2 percent of file and score pairs. A reference row is a reading, and it is now counted as one.

Our guide to VCF, gVCF, BAM and CRAM goes through what each file holds, and the data page has the measurements for fourteen real files. The free file check says in the browser whether a file is a plain VCF or a gVCF, before anything is uploaded.

What Aimosti would (and wouldn't) show you

From a gVCF or aligned reads, a gene counts as examined when 90 percent of its bases carry a called genotype at a depth of 10 or more. From a plain VCF, every position without a record is labelled as read as the common (reference) genotype, an inference. A GRCh37 file's silence is read against the GRCh37 base, and a gVCF whose blocks call nothing is reported as not read.

What we won't claim

We won't read a missing record without knowing which reference the file was called against, count depth without a genotype as examined, or call a gene examined on the strength of a plain VCF.

Bottom line. No record means "matches the reference the file was called against, if this position was read at all". Before a report turns that silence into a genotype, it has to know which reference, and whether the position was read.

Questions people ask

If my VCF has no line for a variant, do I carry the reference version?

Probably, for a common variant in a good 30x genome, but a plain VCF cannot show it. The same gap appears where no read covered the position[3]. That is why our report labels those positions as inferred on a plain VCF.

How can I tell whether my file is a gVCF?

A gVCF has reference blocks: records whose alternative allele is <NON_REF> or <*> and which carry an END= field[3]. File names are not a reliable guide. The free file check looks for those records in the browser.

My gVCF shows ./. everywhere. Is the sequencing broken?

Not necessarily. GATK's CombineGVCFs writes every genotype as a no-call and keeps the likelihoods, leaving genotyping to a later step[7]. If the PL values are present and decisive, the calls can be recovered from them, and our report does that when the whole file has this shape.

Does lifting a GRCh37 file to GRCh38 fix the reference problem?

No. A liftover moves the records that exist to new coordinates; it cannot add a record where the person matched GRCh37 and the caller wrote nothing. At sites where the two assemblies carry different bases, that silence still means the GRCh37 base[9, 10].

References

  1. The 1000 Genomes Project Consortium. A global reference for human genetic variation. Nature, 2015. doi:10.1038/nature15393
  2. Genome assembly GRCh38.p14. NCBI Datasets, 2022. Total sequence length 3,099,441,038 bases.
  3. The Variant Call Format Specification, VCFv4.3. GA4GH / hts-specs, 2025. Sections 1.6.2 (GT, DP, GQ, PL; './.' when a call cannot be made) and 5.5 (reference blocks, END, <*>).
  4. What your DNA file can actually read: measured on 14 real files. Aimosti, 2026. Measured 6 October 2026 on public Personal Genome Project files; data under CC BY 4.0.
  5. 1000 Genomes 30x on GRCh38: per-sample HaplotypeCaller gVCFs (NA12878). International Genome Sample Resource, 2021. NA12878.haplotypeCalls.er.raw.vcf.gz, region chr17:37,600,001-37,800,000 read through its index on 10 October 2026: 32,814 records, 32,520 reference blocks, 294 records with an alternative allele, 242 of them genotyped as a variant, 199,840 bases with a called genotype at depth 10 or more.
  6. GATK 3.7 source: CombineGVCFs.java and ReferenceConfidenceVariantContextMerger.java. Broad Institute, GitHub. mergeRefConfidenceGenotypes builds each genotype with noCallAlleles and then sets PL and AD; CombineGVCFs.referenceBlockMerge does the same for reference blocks.
  7. GATK 4.6.2.0 source: CombineGVCFs.java. Broad Institute, GitHub.
  8. Schneider VA, Graves-Lindsay T, Howe K, et al. Evaluation of GRCh38 and de novo haploid genome assemblies demonstrates the enduring quality of the reference assembly. Genome Research, 2017. doi:10.1101/gr.213611.116
  9. Homo sapiens chromosome 17, GRCh37.p13 Primary Assembly (NC_000017.10). NCBI Nucleotide. Position 36,098,040 is G (read through E-utilities efetch, 10 October 2026).
  10. Homo sapiens chromosome 17, GRCh38.p14 Primary Assembly (NC_000017.11). NCBI Nucleotide. Position 37,738,049 is A (read through E-utilities efetch, 10 October 2026).
  11. dbSNP: rs4430796. NCBI dbSNP. Placed in HNF1B at NC_000017.11:37,738,049 (reference A) and NC_000017.10:36,098,040 (reference G).
  12. gnomAD v4.1 variant 17-37738049-A-G (rs4430796). Genome Aggregation Database, 2024. Genomes: G allele count 77,664 of 151,996 (51.1%); Finnish 3,603 of 10,568 (34.1%). Read through the gnomAD API on 10 October 2026.

Last reviewed . Every number on this page links to the source it comes from; if one of them has moved, tell us.

Free · in your browser

Find out what your own file can read.

The free check opens your DNA file on your device and shows which format you have and what a report could read from it. The file is never uploaded.

Check your fileSee a sample report