Skip to content

v2.7.1

Latest

Choose a tag to compare

@dial481 dial481 released this 19 Nov 20:42

Fixed

  • A two-ALT site written over two or more VCF records no longer hides the
    ClinVar and GWAS Catalog findings 2.6.3 reported. 2.7.0 read such
    records as one call only when they were the exact complementary pair
    with the same ID (1/0 + 0/1, or 1|0 + 0|1 in one phase set), and
    gave every other shape no call at all. Now:
    • The pair is also 1/. + ./1 (and 1|. + .|1 in one phase set),
      as vt decompose and bcftools --multi-overlaps . write a 1/2
      call, and a missing, positional or other ID never keeps two records
      from pairing: only two different rsIDs do. The 1000 Genomes files,
      which write one positional ID per allele, now pair: HG00096's eight
      split sites on chromosome 22 give their eight calls. The pair's rsID
      is decided per ALT: each record keeps the rsID it names, and a record
      naming none has its ALT looked up alone by position. Two different
      rsIDs (both found by position, or one written and one found) make the
      records no pair, as if they named those rsIDs: each ALT is a half call
      under its own rsID. Exactly one rsID names the full call, even where
      the other ALT belongs to no known rsID. Either way the GWAS Catalog
      finding 2.6.3 gave is back, and extract finds the rsID. ClinVar
      then shows two half calls (A/., C/.) at a two-rsID site instead
      of the person's AC.
    • Records that agree but do not make one genotype (0/1 + 0/1
      unphased, different phase sets, a half call beside a heterozygote, two
      different rsIDs, a passing record beside a filtered, low-quality or
      uncalled one) give each passing record's allele as a half call
      (A/.): it goes only to the sources that read one allele (ClinVar and
      the GWAS Catalog, where they run), not to ClinPGx or SNPedia. A
      rejected sibling does not cancel a passing record's ALT; the site still
      follows the half-call rule, so the passing record's full genotype is
      not kept.
    • Records that contradict each other (more than two ALT copies, a
      haploid beside another call, or one haplotype carrying two ALTs) still
      give no call, and the site is now named in high_value_unchecked
      (split_site): under each high-value rsID it stands for, written by
      its records or placed at its position with its REF (whatever IDs or
      ALTs its records carry, as for a single record), and where ClinVar
      holds a Pathogenic or Likely pathogenic row at its position for an ALT
      the records carry, whatever rsIDs the records name. A row without an
      rsID is given as evidence by its locus and gene.
    • split_site_calls still counts every record of a site that is not
      the pair; its warning now says that where such records agree, their
      alleles go only to the sources that read one allele. A pair split only
      because its ALTs belong to two different rsIDs has its own sentence:
      the genotype there is known, and the alleles are assessed apart. The
      half-call note counts standalone half calls left after the split-site
      decision; the split-site alleles have their own count and line. Both
      are given only to the sources that read one allele, where those
      sources run.
      docs/formats.md describes the rule; ADR-0062 is amended.
  • extract shows the genotype analyze reports at a site written over
    several VCF records. 2.7.0's extract read each record on its own, so
    for T>A 1/0 + T>C 0/1 it showed the first record's A/T where
    analyze reports AC. It now reads the exact pair as analyze does,
    on the sequential and the indexed (tabix) path: one full call. Records
    that contradict each other give no genotype; at a site whose records
    only agree, extract shows one of their half calls, while analyze
    reports each. compare, stats and export plink still read every
    record on its own.
  • extract no longer shows -/- where analyze reports a split site's
    half call. A rejected record (FILTER or the sample's FT) written before
    the passing one gave a no-call that won on file order, so T>A 1/0
    rejected + T>C 0/1 passing showed -/- instead of C/-. At a
    placement, a call with a known allele, full or half, now beats a full
    no-call; placement still comes first, and file order breaks other ties.
  • Allelix reads its own annotated VCF again. --vcf-out wrote the rsID
    it recovered for an unnamed split pair on the first record only, so a
    second run saw two IDs and gave no findings. Both records of a merged
    pair now carry its rsID and their own ALLELIX_ORIGINAL_ID.
  • --diff against a baseline from the other genome build no longer
    reports every finding as removed and new. 2.7.0 paired findings by
    position as well as variant, and positions differ between GRCh37 and
    GRCh38, so a chip baseline against a whole-genome run of the same person
    looked like a complete change. When the builds differ, findings named by
    an rsID are now paired by variant and allele, real additions, removals
    and reclassifications are still reported, and the run prints one note
    naming both builds. A finding known only by its position (a ClinVar
    variant without an rsID) can still show as removed and new across
    builds. A baseline whose build is the same, missing or unknown is paired
    by position as before.
  • The HTML report's ClinVar review stars follow ClinVar's rubric. They
    were guessed from the wording of the review status, so a submission
    reviewed by an expert panel showed four stars instead of three, and
    other statuses came out one to four stars off. Each status ClinVar
    defines now has its own star count: four for a practice guideline,
    three for an expert panel, two for multiple submitters with no
    conflicts (or, for a somatic clinical impact, multiple submitters), one
    for a single submitter or conflicting classifications, and none for the
    rest. A status ClinVar does not define is shown as its
    text, with no stars.
    The JSON report, the console, scores and the zero-star rule (ADR-0052)
    were never affected.
  • An HTML report with no findings has a working dark-mode button, and no
    longer shows a search box and filter buttons that did nothing. The page
    script stopped before wiring any control when there was no findings
    table.
  • The closed detail panel of the HTML report no longer leaves a faint
    shadow strip down the window's right edge; the panel casts its shadow
    only while it is open.
  • A long cache path in the HTML report's "Sources not used" list wraps
    instead of making the page scroll sideways on a phone.
  • A position above 2^63 - 1, a VCF POS below 1 or a negative array
    position is a malformed row. A VCF record or an array row at a position
    of 2^63 or more ended analyze, pharmacogenomics, methylation,
    extract, stats and export plink with a Python traceback
    (OverflowError), and a VCF POS of 0 or below was read as a variant.
    Every positioned reader now skips such a row with a warning, and the VCF
    reader counts it in malformed_rows, as it does a POS that is not a
    number. Positions 1 to 2^63 - 1 are read as before, and an rsID list
    without positions is unchanged. An array row at position 0 is still read
    as before, as an rsID call the vendor did not place, with no warning: the
    sources that match by rsID (ClinPGx, the GWAS Catalog, SNPedia) still
    report it.
  • --vcf-out on a file system without hard links (FAT, exFAT, some
    network shares) is refused at start-up. 2.7.0 refused it only after the
    whole analysis, with the JSON or HTML report already written, so a
    whole-genome run on an external drive could end this way after an hour.
    The run now makes a test link in the destination's folder first and
    stops there, with the same message and nothing left in the folder.

Changed

  • Reading a VCF for analyze, methylation and pharmacogenomics is
    faster again. 2.7.0's split-site reading checked every single-base
    record for a second record at its position, adding about 40% to the
    parse. It now checks only when the next record has the same chromosome
    and position, with the same results: on a 600,000-record VCF the
    parse's extra cost fell from 43-46% to 0-17% (two rounds each).
  • JSON reports use schema 15 (2.7.0 wrote 14), and --diff reads
    schemas 1 to 15. Schema 15 adds input.vcf_quality.split_site_half_calls
    and the high_value_unchecked reason split_site. half_calls now
    counts standalone half calls left after the split-site decision; unlike
    schema 14, it excludes half calls belonging to a split site. A merged
    exact pair contributes to neither count. Split alleles eligible for
    half-call assessment have the separate split_site_half_calls count.