[glennhickey]

Summary

Speeds up all three of bar's base aligners, makes minipoa as accurate as abPOA, and makes pecan practical again for progressive alignment. The default aligner does not change (<bar baseAligner="abpoa">), and abPOA's output is identical to before.

	bar on 8 Mb of human/mouse, 8 threads	evolver mammals (mafComparator)
abPOA (default)	59 s, 6.9 GB → 51 s, 5.5 GB	0.853 / 0.652, unchanged
minipoa	29 s, 3.1 GB → 12 s, 4 GB (new defaults)	0.846 / 0.646 → 0.854 / 0.652
pecan	26 s, 0.5 GB	0.859 / 0.654 → 0.869 / 0.666

pecan: cPecan master (55920ef, ComparativeGenomicsToolkit/cPecan#9)

• Exact probability-space forward-backward, 5–11× faster per pairwise alignment. Large unanchored chunks are recomputed in blocks rather than held (3.9 GB → 116 MB for a 10 kb pair). Array-based progressive MSA: same output, 2× faster, half the memory.
• In-process seed anchoring, now shipped as <pecan anchorMethod="seed">. The old MUM anchors accept "unique" matches of about a dozen bases; at 0.3 substitutions per site per branch, half of them are off the true alignment. That, not the HMM, is why pecan degraded on deep branches.
• <bar deepAligner deepAlignerDivergence> can switch the deepest ancestors to pecan. It is off by default.
• On evolver-like inputs pecan still spends about twice abPOA's CPU in bar.

minipoa: cactus-integration (d8050a1, ComparativeGenomicsToolkit/minipoa#2)

• Speed, with identical output: vectorised horizontal-gap scan; DP scratch kept across windows (it was 49 M page faults); a better-centred band; gap states stored in 16 bits (vertical) and 2 bits (horizontal); matrices written with streaming stores.
• Accuracy: an optional convex gap (minipoaGapOpenPenalty2/minipoaGapExtensionPenalty2), and the guide tree fixed (below 8 sequences it kept the input order). <minipoa> now ships abPOA's convex gaps (400/30 + 1200/1) with the guide tree on.
• last-train writes its trained single gap into minipoa and turns the second piece off.
• Includes ofirr's fix for the "backtrack found no predecessor" crash (Fix traceback at the end row: "backtrack found no predecessor" minipoa#1), ported to the convex traceback.

abPOA: ComparativeGenomicsToolkit fork, main (36110ac, ComparativeGenomicsToolkit/abPOA#12)

• The submodule URL changes from yangao07/abPOA to ComparativeGenomicsToolkit/abPOA. The pinned commit is upstream's current main, which aligns identically to the v1.5.6 we had, plus one change: each DP row stores only its band, packed together, instead of the full query width.

bar

• Aligning only the dominant end is now taken only when no adjacency exceeds bandingLimit; it allowed twice that, leaving the far part of such adjacencies unaligned.
• The log line names pecan's anchor method. New evolver tests run pecan on primates and mammals.

Updating an existing clone

The abPOA URL changed, so after pulling run git submodule sync && git submodule update --init --recursive, or the update will look for the new commit upstream and fail.

Testing

• Identical output where claimed. abPOA and minipoa (with its old settings) match their previous versions on simulated truth sets, on bar's 8 Mb human/mouse output, and on evolver.
• abPOA was also compared across all three gap models and local/extend/global modes (305 comparisons), and passes its own tests/run_all.sh.
• Evolver mammals and primates for all three aligners, and bar's unit tests.
• minipoa's DP was run under ASan/UBSan and independently reviewed cell by cell against a scalar reference.

Before merging

• Cluster test (running)
• Merge the tool PRs with merge commits rather than squash: Faster, exact forward-backward; in-process seed anchoring; array-based progressive MSA cPecan#9, Keep only each DP row's band, packed together abPOA#12, Faster DP, an optional convex gap, and a guide-tree fix; includes #1 minipoa#2
• Re-pin the three submodules to those merge commits (edd9142).
• Merge master in (d429818), so CI runs on the combined code. Its 23 commits merged without conflicts.

🤖 Generated with Claude Code