bdgdiff: do not truncate region scores to integers (fixes #715)

[cindykrafft]

Fixes #715.

Describe the bug

macs3 bdgdiff writes, as the score of every region in the cond1, cond2 and common BED files, the length-weighted mean of the log10 likelihood ratios of the intervals inside the region (as documented). Since the pure-Python-mode conversion, TwoConditionScores.mean_from_peakcontent in MACS3/Signal/ScoreTrack.py declares the per-interval score tmp_v and the accumulator sum_v as cython.long, so each log10LR is cut to its integer part before it is summed. In MACS2 2.2.9.1 and MACS3 3.0.0 both were float64_t ("for better precision"); 3.0.1 through current main have cython.long.

Consequences, exactly as reported in #715: with a cutoff below 1 (the reporter used -C 0.2) every score in common.bed becomes 0, and cond1/cond2 scores can be below the cutoff; with the default -C 3 every score is biased low by up to 1 (e.g. 24.9521 instead of 25.5431).

To Reproduce

Four one-interval-per-region bedGraphs where the cond1 intervals have log10LR(t1 vs t2) = 0.3271 and 6.1784 (see repro.py in the link below):

macs3 bdgdiff --t1 t1.bdg --c1 c1.bdg --t2 t2.bdg --c2 c2.bdg --d1 1 --d2 1 -C 0.2 -l 200 -g 100 --o-prefix syn

main: syn_c0.2_cond1.bed score 3.0 (= (0 + 6) / 2). Fixed: 3.25276 (= (0.3271 + 6.1784) / 2).

The project's own test/cmdlinetest step 10 (bdgdiff on the CTCF chr22 50k test data) shows the same: on main, 638 of 638 cond1 scores differ from test/standard_results_bdgdiff/run_bdgdiff_prefix_c3.0_cond1.bed (first three: 24.9521 / 36.903 / 48.8178 vs 25.5431 / 37.3104 / 49.3307). The reference files were evidently produced by a float-correct build; cmdlinetest does not catch this because BED files that differ are accepted when their Jaccard index is > 0.99, which ignores the score column.

The fix

Declare tmp_v and sum_v as cython.double again (two lines). With the fix, the bdgdiff output on the test data matches test/standard_results_bdgdiff exactly (0 of 638 scores differ), so no reference file needs regenerating.

Tests

• test/test_ScoreTrack.py::Test_TwoConditionScores::test_call_peaks_score_is_mean_logLR (new): fails on main with AssertionError: 3.0 != 3.2527 within 3 places, passes with the fix.
• The Test_TwoConditionScores.setUp fixture called bedGraphTrackI.safe_add_loc, which no longer exists; it now uses add_loc so the class can hold tests again.
• pytest test/test_ScoreTrack.py test/test_BedGraph.py test/test_PeakIO.py: 8 passed before, 9 passed after.
• Full pytest --runxfail (excluding test_online.py, which needs network): 102 passed, 4 skipped on main; same plus the new test with the fix.
• ruff check / flake8 report no findings on the changed lines (the files' pre-existing findings are unchanged).

Found while working through open issues in Mytochondria, a volunteer project that verifies fixes for the software behind published results (methods and harnesses: https://github.com/cindykrafft/mytochondria/tree/main/audits/macs2)

---

Generated by Claude Code

[taoliu]

Thanks for the fixes!