6b285a53b036b309e3c7a9b61d3741731088a172
lrnassar
  Fri Jun 12 02:35:01 2026 -0700
varFreqs: switch affectedAF/backgroundAF from max-across-cohorts to pooled
sum(AC)/sum(AN) so the rate matches the carrier count scale.

Per-arm AN is derived as round(AC/AF) when both are reported. An optional
"default_an" column was added to databases.tsv so AF-only cohorts (ABraOM,
ALFA) can synthesize a denominator from their cohort size; without it
those cohorts had been silently dropped from the pooled rate.

New affectedAN and backgroundAN columns expose the pool denominator. The
mouseOver now reads "Affected AC/AN: 33238 / 213153" so the ratio is
visible. Per-arm cohorts that ship only AC and no default_an (MGRB,
GREGoR AC_AFFECTED/UNAFFECTED/UNKNOWN, AllOfUs per-population) are still
listed in affectedCohorts/backgroundSources but contribute 0 to the
pool, preserving the invariant pool_AF <= 1.

The build pipeline is unchanged: re-run vcfToBigBed.py --split-affected
against the existing merged.annotated.vcf.gz. refs #36642

diff --git src/hg/makeDb/trackDb/human/varFreqsAffected.html src/hg/makeDb/trackDb/human/varFreqsAffected.html
index 0099d2fda9d..36af10b8e43 100644
--- src/hg/makeDb/trackDb/human/varFreqsAffected.html
+++ src/hg/makeDb/trackDb/human/varFreqsAffected.html
@@ -30,55 +30,68 @@
 <tr><th>Color</th><th>Consequence class</th><th>Examples</th></tr>
 <tr><th style="background-color:#FF0000;width:2em">&nbsp;</th>
     <td>Protein-truncating / loss-of-function</td>
     <td>stop_gained, frameshift, splice_donor, splice_acceptor, stop_lost, start_lost</td></tr>
 <tr><th style="background-color:#1F77B4;width:2em">&nbsp;</th>
     <td>Missense / in-frame</td>
     <td>missense, inframe_insertion, inframe_deletion, protein_altering</td></tr>
 <tr><th style="background-color:#008000;width:2em">&nbsp;</th>
     <td>Synonymous</td>
     <td>synonymous, stop_retained</td></tr>
 <tr><th style="background-color:#808080;width:2em">&nbsp;</th>
     <td>Non-coding / intergenic</td>
     <td>intron, non_coding, intergenic, UTR</td></tr>
 </table>
 <p>
-The score (used for shading) is the affected/case allele frequency times 1000. Variants
-contributed only by a cohort that reports allele counts but no allele frequency (GREGoR)
-have a score of 0 but are still drawn in their consequence color.
+The score (used for shading) is the pooled affected/case allele frequency times 1000.
+</p>
+
+<h3>Pooled allele frequency</h3>
+<p>
+<b>Affected AF</b> is the pooled rate across contributing affected arms:
+<code>affectedAF = sum(AC) / sum(AN)</code>, where <b>affectedAC</b> sums the allele counts
+and <b>affectedAN</b> sums the allele numbers across each cohort/arm that ships both AC and
+AF (the per-arm AN is derived as <code>round(AC / AF)</code>). Cohorts that publish only AF
+contribute via a configured <code>default_an</code> in the build configuration. Cohorts
+that publish only AC and have no <code>default_an</code> set (currently GREGoR's per-arm
+AC_AFFECTED/UNAFFECTED/UNKNOWN) are listed in <b>affectedCohorts</b> but do not contribute
+to the pool numerator or denominator; their carriers are visible in the per-database AC
+column instead. The pooled rate is preferred over a max-across-cohorts statistic so a
+small cohort with a high local AF cannot dominate the displayed frequency.
 </p>
 
 <h3>Finding case-enriched loss-of-function variants</h3>
 <p>
 To look for protein-truncating variants that are common in affected individuals but rare
 in the background, set the Consequence filter to Stop Gained, Frameshift, Splice Donor and
 Splice Acceptor (these appear red), then add an upper limit on the
 <b>Background AF</b> filter. Each variant here carries both its affected frequency and its
 background frequency, so this isolates variants seen in cases with little or no presence in
 the population/unaffected set. Comparing visually against the
 <a href="hgTrackUi?g=varFreqsBackground">Population + Unaffected</a> track shows the same
 contrast across a whole gene.
 </p>
 
 <h2>Filters</h2>
 <ul>
   <li><b>Variant Type</b> and <b>Consequence</b>: restrict to SNV/insertion/deletion/MNV
       and to predicted consequence classes (the Consequence filter uses OR logic over the
       comma-separated tokens on each variant).</li>
-  <li><b>Affected/case AF</b> and <b>Affected/case AC</b>: the maximum allele frequency and
-      summed allele count across the affected arms.</li>
-  <li><b>Background AF</b> and <b>Background AC</b>: the same variant's frequency in the
+  <li><b>Affected/case AF</b>, <b>AC</b>, <b>AN</b>: the pooled allele frequency
+      (sum AC / sum AN), summed allele count, and summed allele number across the
+      contributing affected arms. See &quot;Pooled allele frequency&quot; above.</li>
+  <li><b>Background AF</b>, <b>AC</b>, <b>AN</b>: the same triple computed across the
       population + unaffected background, for filtering case-enriched variants.</li>
   <li><b>Affected/case cohort</b>: restrict to variants seen in specific disease cohorts
       (for example, only the two autism cohorts).</li>
   <li><b>Reference/Alternate Length</b> and <b>Length Change</b>: filter by allele length.</li>
 </ul>
 
 <h2>Methods</h2>
 <p>
 Variant-frequency VCFs from the contributing cohorts were stripped of unneeded INFO fields,
 normalized with <code>bcftools norm</code> (splitting multi-allelic sites), and merged with
 <code>bcftools merge</code>. The merged callset was annotated with predicted protein
 consequences using <a href="https://samtools.github.io/bcftools/howtos/csq-calling.html"
 target="_blank">bcftools csq</a> against the
 <a href="https://www.ensembl.org/info/data/ftp/index.html" target="_blank">Ensembl</a>
 GRCh38 release 115 gene models.