ChromBPNet: correct the CHIP profile/count transform + drop 33 mouse models
Two corrections, published together. 786 -> 753 rows.
CHIP transform (chorus PR #120). The 744 BPNet/CHIP rows were built with a
per-strand softmax andexpm1of the pooled log-count. bpnet-refactor'sorig_multi_loss=Falsebranch β which the shipped weights uniquely pin
(logcounts_predictions/kernelis(2,1);Truegives(3,2)and refuses
to load) β takes ONE multinomial over the flattened both-strand vector and
builds its count target as a per-tracklog1ppooled withreduce_logsumexp,
so the target islog(2 + total)and the inverse isexp(C) - 2, notexpm1(C). Corroborated verbatim by bpnet-lite'sBPNetdocstring.
Both strands now pool into one CDF row, so CHIP sample counts double to
37,344 / 68,008 / 2,176,256. CHIP values fall (median effect ratio 1.86x);
the old rows inflated CHIP effect percentiles by ~30% at the median.Mouse removal (chorus PR #121). 33 of the 42 ATAC/DNASE rows were mm10
models from ENCODE's mouse developmental atlas, but the builder openshg38.faand uses the hg38 DHS vocabulary β those CDFs were built by
pushing human sequence through mouse models. Removed by pure row
subsetting; no model was re-run.
The 9 surviving human ATAC/DNASE rows (K562, HepG2, GM12878, IMR-90 ATAC;
HepG2, IMR-90, GM12878, K562, H1 DNase) are BIT-IDENTICAL to the previous
revision, with sample counts unchanged at 18,672 / 34,004 / 1,088,128.
sha256 76f267dc862edc86052f2b25a2a8520e960dd193ad39c4ecd19e32b8a8546553
All CDFs verified finite, monotone, and non-degenerate.
Verified before merge: 753 rows (744 CHIP + 9 human ATAC/DNASE), the 9 accessibility rows bit-identical to the previous revision, sample counts unchanged, all CDFs finite/monotone/non-degenerate, sha256 76f267dc.