8e053a77fec3df57d9bfa44fafbb07c0cdadd1a7
max
  Fri Sep 4 15:47:07 2026 -0700
AlphaGenome Variant Impact (AVI) score track for hg38

New track under the Deleteriousness Predictions container showing Google
DeepMind's AlphaGenome Variant Impact score for every possible single-base
substitution, as four bigWigs, one per alternate allele.

alphaGenomeToWig splits the 83 GB source table into one fixedStep wig per
alternate allele. It copies the PHRED column through as text so no precision
is lost, and zero-fills gaps up to -maxGap so a fixedStep block can span a
long contiguous run. The input has no row where alt == ref, so each per-allele
file is missing roughly every fourth position; without the fill the wig
fragments into blocks of about 3.5 values and the bigWig doubles in size.
All 8,812,917,339 input lines were used, none skipped.

The data cannot be redistributed, so the stanzas set tableBrowser off and the
description page sends users to the AlphaGenome Atlas. Alpha only for now.
No reference yet, the score is unpublished.

refs #38261

diff --git src/hg/makeDb/doc/hg38/alphaGenome.txt src/hg/makeDb/doc/hg38/alphaGenome.txt
new file mode 100644
index 00000000000..529b9bb9429
--- /dev/null
+++ src/hg/makeDb/doc/hg38/alphaGenome.txt
@@ -0,0 +1,57 @@
+# AlphaGenome Variant Impact (AVI) score track, refs #38261
+# Fri Sep  4 14:20:00 PDT 2026 (Claude max)
+
+# Google DeepMind provided alphagenome_variant_impact_score_snvs.tsv.gz (83 GB, tabix
+# indexed) directly, along with terms.txt, their terms of service. The same data is on
+# https://deepmind.google.com/science/alphagenome/atlas
+# The file is a tab-separated table of chrom, pos, ref, alt, raw_score, PHRED covering
+# every possible single-base substitution in GRCh38. It holds no indels, despite what
+# the AlphaGenome Atlas offers, and no row where alt == ref.
+
+mkdir -p /hive/data/genomes/hg38/bed/alphaGenome
+cd /hive/data/genomes/hg38/bed/alphaGenome
+
+# We cannot redistribute this data, so the trackDb stanzas carry "tableBrowser off"
+# and the description page points users at Google for the download.
+
+# Split into one wig per alternate allele. The program copies the PHRED column through
+# as text, so no precision is lost, and it prints a full accounting of lines read,
+# lines skipped and values written. Took 28 minutes and 10.7 MB of memory.
+# Source: ~/kent/src/hg/oneShot/alphaGenomeToWig/alphaGenomeToWig.c
+alphaGenomeToWig alphagenome_variant_impact_score_snvs.tsv.gz
+
+# What it reported. Nothing was skipped: every input line became a score, and the
+# real scores written (values minus zero fills) equal the input line count exactly.
+#   read 8812917339 data lines, used 8812917339
+#     A: 2931187386 values in 399706 sections (860697142 zero fills)
+#     C: 2937603476 values in   3556 sections (599003079 zero fills)
+#     G: 2937603388 values in   3562 sections (601477231 zero fills)
+#     T: 2931143258 values in 402430 sections (863442717 zero fills)
+#   wrote 11737537508 values in 809254 sections, 2924620169 of them zero fills
+# 11737537508 - 2924620169 = 8812917339, matching the 8,812,917,339 data lines in
+# the input (wc -l minus the header). A and T have far more sections than C and G
+# because unscored stretches longer than maxGap fall more often between A/T scores.
+
+# On the number of fixedStep sections: because no row has alt == ref, each per-allele
+# file is missing roughly every fourth position. alphaGenomeToWig fills gaps of up to
+# -maxGap=10 bases with zero so one fixedStep block can span a long contiguous run.
+# Without this the file fragments into blocks of about 3.5 values each and the
+# resulting bigWig is roughly twice as large. Measured on 30 M lines of chr21:
+#   fixedStep wig  ->  34.5 MB bigWig
+#   bedGraph       ->  74.0 MB bigWig
+
+for n in A C G T; do
+    wigToBigWig alphaGenome$n.wig ../../chrom.sizes $(echo $n | tr A-Z a-z).bw &
+done
+wait
+
+# Check the value range, this is what viewLimits/viewLimitsMax in trackDb are based on
+for n in a c g t; do bigWigInfo $n.bw | egrep "^(min|max|mean):"; done
+
+# Sanity check that the 1-based input maps to the right 0-based bigWig position.
+# chr21:5010001 G>A has PHRED 6.58432 in the source file:
+bigWigToBedGraph -chrom=chr21 -start=5010000 -end=5010004 a.bw stdout
+# chr21 5010000 5010001 6.58432   <- the G>A score
+# chr21 5010001 5010002 0         <- ref is A here, so no A>A row, zero filled
+
+# The wig files are only intermediates and can be deleted once the bigWigs check out.