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

Lifting a genome from GRCh37 to GRCh38: what CrossMap quietly changes

Many genome files are still delivered on GRCh37 and are translated to GRCh38 before anything is read. On a real gVCF, CrossMap dropped an allele from about one variant call in nine, most of them deletions, and left the genotype pointing at an allele that was no longer there. Its log counts only records it cannot place, so none of this showed up there. Here is the mechanism, and the guard we run around it.

Key takeaways

  • GRCh37 and GRCh38 give the same DNA different coordinates, and at thousands of positions a different reference base[1].
  • CrossMap sizes a record by its whole ALT column. In a gVCF, where every call also lists <NON_REF>, a deletion is read as an insertion and its allele is thrown away[9].
  • Where GRCh38 corrected the reference base, CrossMap removes the ALT allele equal to the new base without renumbering the genotype[9].
  • Split, trimmed and stamped before the lift, a public GRCh37 gVCF lost 0.64 percent of its calls instead of 10.4 percent.
  • Above 3 percent of calls dropped, the file is refused. Healthy files measured 0.14 to 1.8 percent.

The human reference genome has been rebuilt more than once. GRCh37 dates from 2009; GRCh38, its first coordinate-changing successor, was submitted in December 2013[1]. The two number most of the genome differently, so a variant file made against GRCh37 has to be translated before it can be compared with anything built on GRCh38. That translation is a liftover, and CrossMap is one of the two tools most often used for it[2, 3]. On most records it is right. The failures described here are narrow, but they change genotypes instead of dropping records, and a pipeline notices a changed genotype only if someone counts.

Two maps of one genome

A reference assembly is a coordinate system as well as a sequence. Chromosome 1 is 249,250,621 bases long in GRCh37 and 248,956,422 in GRCh38[4, 5], and a position on it in one assembly is usually a different number in the other. GRCh38 resolved roughly a thousand problems in its predecessor, from megabase-scale reorganisations of the sequence to thousands of single-base changes[1].

GRCh38's builders looked for GRCh37 bases that nobody in the 1000 Genomes data carried, or that looked like sequencing errors, checked them, and replaced 8,248 of them[1]. At such a site most people carry the GRCh38 base, which in their GRCh37 file was written as the variant.

8,248

single bases replaced when GRCh38 was built, most where the GRCh37 base was rare or an error[1]

294,199

bases fewer on chromosome 1 in GRCh38 than in GRCh37[4, 5]

275 of 638

chains in the Ensembl file, on the primary chromosomes, that match the reverse strand[6]

Terms this piece relies on

Genotype
Two numbers that point into the allele columns: 0 is the REF, 1 the first ALT, 2 the second[7]. It is only as right as the allele order it points at.
Chain file
A file of chains: alignments saying which stretch of one assembly matches which of the other, with gaps on both sides[8]. A chain can match the reverse strand.
<NON_REF>
A placeholder allele a gVCF adds to every record for any alternative not listed, so a two-allele call there carries two ALT alleles.
Figure 1. How a chain file connects the two assemblies, schematic and not to scale. A record at a corrected base comes out with a different REF than it went in with[8, 1].

We lift with Ensembl's chain file. Its reverse-strand chains are many but short: together they cover 8.1 million of the 2.84 billion bases it aligns on the primary chromosomes[6].

What a liftover has to rewrite

Moving a record means more than renumbering it. REF is read again from GRCh38, which at a corrected base gives a different letter. If the alleles are then reordered, every field indexed by allele has to follow: the genotype, the depth per allele (Number=R in the specification) and the likelihood per genotype (Number=G)[7]. Change REF and ALT and leave those alone, and the record is well formed in places and wrong in others.

In the 2024 comparison by the authors of BCFtools/liftover, CrossMap, tested at version 0.6.6, was the only one of six tools that could not swap the alleles of a single-base variant, and one of four that could not handle a record with more than one ALT allele[3]. On 45,595,458 two-allele SNVs from the 1000 Genomes Project, lifted from GRCh37 to GRCh38, CrossMap dropped 46,607 and BCFtools/liftover 12,582[3]. Those are variant sites from a cohort, not one person's calls, and they were lifted with the UCSC chain file, not the Ensembl one.

Regardless of the increased accuracy of BCFtools/liftover compared to the other tools, we warn that the liftover process is in general a lossy procedure and should not be regarded as a substitute for realigning sequences against a different genome assembly for datasets with available raw data as realignment and re-calling of variants will always generate better results.

Genovese et al., Bioinformatics, 2024[3]

A deletion that CrossMap reads as an insertion

CrossMap decides whether a record is a substitution, an insertion or a deletion by subtracting the length of the REF column from the length of the ALT column[9]. It measures the whole ALT column, commas included. A two-base deletion in a gVCF, AT to A, arrives with ALT written as A,<NON_REF>: eleven characters against a REF of two. By that arithmetic it is an insertion.

For an insertion CrossMap reads a single reference base from GRCh38, here A. It then deletes every ALT allele equal to the new REF[9]. The deletion allele is A, so it goes. The record survives as A to <NON_REF>, and its genotype, which pointed at allele 1, the deletion, now points at the placeholder.

Figure 2. One gVCF deletion through CrossMap 0.7.3. The genotype column is copied unchanged, so allele 1 now names the placeholder instead of the deletion[9].

gVCF

#             POS   REF  ALT          FORMAT  SAMPLE
in            1000  AT   A,<NON_REF>  GT:AD   0/1:14,12,0
out           1000  A    <NON_REF>    GT:AD   0/1:14,12,0
CrossMap log  Variants failed to map: 0
An invented record on an invented reference, run through the real CrossMap 0.7.3 on 10 October 2026, columns trimmed for reading. The read depths still list three alleles for a record that now has two, and the log reports no failure.

On real files this was not rare. Before we changed anything, CrossMap removed an allele from 11.3 percent of the variant calls in the first two million lines of a public GRCh37 gVCF, the 1000 Genomes sample NA07357 as reanalysed with DRAGEN 3.7.6[10], and from 4.7 percent in the indel file of a Dante Labs genome from the Personal Genome Project. We paired each record dropped from the indel file with its original: all 1,863 were multi-allelic records carrying a deletion, returned with a one-base REF. Those are our measurements of 18 September 2026.

We run CrossMap 0.7.3. The two lines responsible are unchanged in 0.7.4, released on 31 July 2026, and on the project's main branch[9, 11].

A call deleted where GRCh38 fixed the reference

The same removal step does the second kind of damage. Take a corrected site where GRCh37 reads G and GRCh38 reads T, and a person who carries one of each. Their GRCh37 file says REF G, ALT T, genotype 0/1. CrossMap writes the GRCh38 base into REF, finds that the ALT allele now equals it, and deletes the ALT[9]. It does not renumber the genotype or the per-allele fields.

gVCF

#     POS   REF  ALT          FORMAT  SAMPLE
in    2000  G    T,<NON_REF>  GT:AD   0/1:11,13,0
out   2000  T    <NON_REF>    GT:AD   0/1:11,13,0
An invented record at a position where the invented GRCh38 base is T, through the real CrossMap 0.7.3. The person carried G and T; the record now says they carry T and the placeholder.

A two-allele record is left with an empty ALT column, a malformed line. A multi-allelic one keeps a genotype that names the wrong alleles, with depth and likelihood lists sized for alleles that are gone. That shape stopped a customer's analysis on 29 August 2026, when the next tool refused the mismatched depth field. Read instead, it would have been a confident wrong genotype.

Unlike the deletion defect, this loss tracks the person, not the file format. With the GRCh38 sequence as the check, all 899 single-base calls CrossMap dropped from the start of chromosome 1 in the Dante Labs SNV file were sites where the person carried the base GRCh38 had corrected to. A tool that swaps REF and ALT at such a site, as BCFtools/liftover does, keeps the call[3].

The guard around the lift

Once CrossMap has rewritten a record, the removed allele is gone and the genotype cannot be repaired from what is left. So the guard works on both sides of the lift: each record goes in a form CrossMap handles correctly, carrying a note of its own shape, and anything that comes back in a different shape is dropped.

What a GRCh37 VCF or gVCF goes through

  1. Split. bcftools norm -m -any turns every multi-allelic record into two-allele records.
  2. Trim and stamp. Each record is trimmed to its shortest form, and its allele count and REF length are written into INFO as AIMOSTI_OAC and AIMOSTI_ORL. CrossMap rewrites only the END field there and passes the rest through[9].
  3. Lift. CrossMap translates the file with the Ensembl chain. A gVCF goes through its gvcf mode, because the vcf mode rewrites each reference block's END as if the block were one base long[9].
  4. Compare. Any record whose allele count or REF length changed is dropped and counted. The stamps are then removed.
  5. Gate. If more than 3 percent of the variant calls were dropped, the analysis stops. Reference blocks and the placeholder halves left by the split are not counted.
  6. Normalise and sort. CrossMap writes a reverse-strand chain's records in reverse order, so the lifted file is normalised against GRCh38 and sorted before it can be indexed.

On the two invented records above, split first, the guard keeps the deletion and drops the switched call; one call in three dropped is a refusal.

Where the 3 percent line comes from

A healthy lift still drops calls at corrected sites, so the gate cannot be zero, and how many depends on the person and the stretch of genome. We measured slices and whole files through the real CrossMap in our worker image.

Table 1. Variant calls dropped in the lift from GRCh37 to GRCh38, before and after the split and trim
File and stretchBeforeAfter
NA07357 gVCF, whole file507,600 of 4,878,626 (10.4%)32,602 of 5,107,199 (0.64%)
NA07357 gVCF, first 2 million lines4,931 (11.3%)843 (1.8%)
NA07357 gVCF, chromosome 4 into 5not measured79 (0.14%)
Dante Labs indel VCF, start of chromosome 11,863 (4.7%)0
Dante Labs SNV VCF, start of chromosome 1899 (1.5%)883 (1.5%)
Dante Labs SNV and indel files, whole genomenot measured30,215 of 4,866,617 (0.62%)

Source: Aimosti measurements, 17 to 18 September 2026, CrossMap 0.7.3 with the Ensembl chain[6]. NA07357 is the DRAGEN 3.7.6 hg19 gVCF of the public 1000 Genomes reanalysis[10]; the Dante Labs files are a public Personal Genome Project genome. The first NA07357 figure is the run refused at the gate then in force, 1 percent.

Figure 3. The table as a chart. Every measurement after the change falls under the 3 percent gate.

The gate sits at 3 percent, above every healthy rate we measured, the highest 1.8 percent. A broken run would sit far above it: if the chain and the reference sequence did not belong together, every base CrossMap read would be unrelated to the call, and by our arithmetic about a quarter of single-base calls would match their new REF by chance and be dropped. The gate was 1 percent until the deletion defect was fixed at its source. That defect, not the limit, was what refused the GRCh37 genomes, and the limit moved only after it was gone.

The tool's own line, with its counts, goes to our records. The refusal is permanent, because the same bytes lift the same way every time.

What the report does with a GRCh37 file

Only variant files go through CrossMap.

Table 2. How each kind of GRCh37 file is read
FileHow it reaches the reportWhat does not come through
Chip exportNot lifted. Each call is matched to our panel by rsID or GRCh37 position and placed against the panel's GRCh38 alleles.Indels, and A/T or C/G positions, are not read from a chip.
Plain VCFSplit, stamped, lifted by CrossMap and checked by the guard; refused above 3 percent.A dropped call leaves a silent position, labelled as inferred like any other.
gVCFThe same, through CrossMap's gvcf mode, so reference blocks keep their length.A dropped call is not reported. An hg19 mitochondrion is dropped whole and the maternal-line card says why.
BAM or CRAMNot lifted. Deep Read reads the alignments against a GRCh37 reference and runs the 29-target pharmacogene panel in full.The clinical-panel coverage map and the recurrent-deletion screen need GRCh38; the report says they were not measured.

Source: Aimosti's pipeline as of 10 October 2026: the chip reader, the normalise step and liftover guard, and the Deep Read task.

The mitochondrion is a separate case. A file on hg19, UCSC's version of GRCh37, numbers it against a 16,571-base sequence, while GRCh38 and the maternal haplogroup tree use the 16,569-base rCRS[12, 13]. The chain maps the mitochondrion as if the two were identical, so its records would come through with shifted positions and no error. We drop that contig and say so on the card.

A plain VCF has one more trap that no liftover touches. A person who matches GRCh37 at a corrected site has no record there, and a lift cannot create one. Read against GRCh38, that silence would say they carry the GRCh38 base. Our polygenic scores read a silent position against the base of the assembly the file was made on; the two assemblies disagree at 20 of the 1,677 positions those scores use.

What is still lost

The guard trades a wrong call for a missing one, and the missing ones are real: at each corrected site where the person carried both bases, the call is dropped. In a gVCF that does not always leave a clean gap. The split turns G to T,<NON_REF> into a call and a placeholder half, G to <NON_REF> with the genotype 0/0. The half lifts without changing shape, so the guard keeps it, and in our reproduction it kept the call's read depth and our coverage code counted the position as read. A person who is G/T there can come out of the lift reading as T/T.

BCFtools/liftover swaps REF and ALT at corrected sites, reorders the per-allele fields and handles multi-allelic records, which would keep the calls the guard drops[3]. We have not switched: it is a different lift engine, unmeasured on our files. Its authors also found that the Ensembl chain, the one we use, dropped many more variants than the UCSC chain[3]. We have not compared the two.

Where the reads exist, the lift can be avoided. That is the point of the warning quoted above, and it is why a GRCh37 BAM or CRAM is read on its own assembly instead.

What Aimosti would (and wouldn't) show you

A GRCh37 VCF or gVCF is lifted to GRCh38 before anything in it is read, through a guard that drops every record whose alleles changed shape in the lift. If more than 3 percent of the variant calls are dropped, the analysis stops and the file is refused rather than reported on. A chip export is not lifted at all: its calls are matched to our panel by rsID. A GRCh37 BAM or CRAM is read against a GRCh37 reference, and the two Deep Read parts that need GRCh38 are reported as not measured.

What we won't claim

We won't report a genotype from a record whose alleles changed during the lift, or build a report from a file the guard refused. We won't call the lift lossless either: the guard drops real calls where GRCh38 corrected the reference, and at some of those a gVCF can still read as matching it.

Bottom line. A liftover can change a genotype and still report success. Count each record's alleles and its reference length before the lift, compare them after, and treat any record that changed as unread.

Questions people ask

Is a GRCh37 file less accurate than a GRCh38 one?

A lift costs a small share of its calls: 0.62 and 0.64 percent of the variant calls in the two whole genomes we measured. On the stretches we checked call by call, every one was at a site where the two assemblies disagree.

Does a 23andMe or AncestryDNA file go through CrossMap?

No. Each position in a chip export is matched to our panel by its rsID or its GRCh37 position, and its two letters are placed against the panel's GRCh38 alleles. Nothing in a chip file is lifted.

How can I tell which assembly my file is on?

A VCF's header usually says. Its ##contig lines give each chromosome's length, and chromosome 1 is 249,250,621 bases in GRCh37 and 248,956,422 in GRCh38[4, 5]. Our free file check reads the header in your browser and names the assembly without uploading anything.

What happens if my file is refused?

The analysis stops before any report is built, and you are told that the file changed too much in translation to be trusted. The same file would be refused again, so the message points to a GRCh38 copy of the same data if your provider has one, and otherwise to us.

References

  1. 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 GRCh38 submitted to INSDC in December 2013, the first coordinate-changing update since 2009; 8,248 sites updated by targeted base fixes.
  2. Zhao H, Sun Z, Wang J, Huang H, Kocher JP, Wang L. CrossMap: a versatile tool for coordinate conversion between genome assemblies. Bioinformatics, 2014. doi:10.1093/bioinformatics/btt730
  3. Genovese G, Rockweiler NB, Gorman BR, et al. BCFtools/liftover: an accurate and comprehensive tool to convert genetic variants across genome assemblies. Bioinformatics, 2024. doi:10.1093/bioinformatics/btae038
  4. Homo sapiens chromosome 1, GRCh37.p13 Primary Assembly (NC_000001.10). NCBI Nucleotide. Length 249,250,621 bases.
  5. Homo sapiens chromosome 1, GRCh38.p14 Primary Assembly (NC_000001.11). NCBI Nucleotide. Length 248,956,422 bases.
  6. GRCh37_to_GRCh38.chain.gz: Ensembl assembly mapping for Homo sapiens. Ensembl. 285,250 bytes, read 10 October 2026. Our count: 638 chains on the primary chromosomes, 275 of them on the reverse strand, covering 8,104,065 of 2,844,848,882 aligned bases.
  7. The Variant Call Format (VCF) Version 4.2 Specification. GA4GH / hts-specs.
  8. Chain format. UCSC Genome Browser documentation.
  9. CrossMap source code: cmmodule/mapvcf.py and cmmodule/mapgvcf.py. GitHub. alt_allele_size = len(fields[4]) and the removal of any ALT equal to REF, read in releases 0.7.3 and 0.7.4 and on the master branch, 10 October 2026.
  10. Illumina. 1000 Genomes Phase 3 Reanalysis with DRAGEN 3.5, 3.7, 4.0, 4.2, and 4.4. Registry of Open Data on AWS. NA07357.hard-filtered.gvcf.gz in the DRAGEN 3.7.6 hg19 set, 4,310,532,750 bytes.
  11. CrossMap release history. Python Package Index. 0.7.3 uploaded 18 July 2024; 0.7.4 uploaded 31 July 2026.
  12. Homo sapiens mitochondrion, complete genome (NC_001807.4, the hg19 chrM). NCBI Nucleotide. Length 16,571 bases.
  13. Homo sapiens mitochondrion, complete genome (NC_012920.1, the rCRS). NCBI Nucleotide. Length 16,569 bases.

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