Improved peak-calling with MACS2

Key results

1. ShownIn default single-end mode, MACS2 drops R2 reads and extends each R1 read to a fixed length, so fragment coverage computed on paired-end ATAC-seq is often badly wrong.

   In the Corces et al. Omni-ATAC data run with '--nomodel', reads were extended to 200bp, yet only 21.7% of fragments were within 10% of 200bp. Over 40% of positions with true coverage of at least five had computed coverage off by at least 50%; in SRR5427884, at positions with true coverage of at least five, coverage was calculated correctly only 12% of the time. · Abstract; Figure legends

2. ShownThe coverage errors carry through to peaks: single-end mode calls many peak bases with little real fragment support; paired-end mode ('-f BAMPE') reduces these low-coverage peak bases while also adding some peaks.

   Of 17.5Mbp of single-end peak bases (sample SRR5427884), 32.2% had true fragment coverage below three. Paired-end mode gave fewer peaks overall and fewer low-coverage peak bases, and called nearly 1Mbp of additional peaks at coverage above two. Example loci show single-end peaks without apparent enrichment (a 257bp peak near MYRF). · Abstract; Figure legends

3. SuggestedPaired-end mode can miss real signal because original MACS2 sets the minimum peak length to the average fragment length; the modified version exposes '--min-length' (default 100bp) and '--max-gap' (default 50bp) as user parameters.

   In SRR5427884 a significant 129bp region at the TTF1 transcription start site was discarded because the minimum length was 174bp. With '--min-length 100', overlap between single-end and paired-end peak sets increased (Fig. 1D, Supplementary Fig. 6); the text reports no overlap counts. · Abstract; MACS2 modifications; Figure legends

4. SuggestedAn unresolved hash collision in MACS2 miscalculates p-scores and q-scores in 'callpeak' and 'bdgcmp', which can falsely extend peaks; the modified version fixes it.

   In SRR5427886 (run with dynamic lambda), a 1968bp peak over the CTCF transcription start site included a leading 175bp with no significant positions, attached through a single miscalculated q-value caused by the collision (Supplementary Fig. 7). · MACS2 modifications; Figure legends

5. ClaimedFurther fixes in v2.1.2_dev: corrected average fragment length in paired-end 'callpeak', reduced p-score errors at large control lambda, and double-precision scoring via 'bdgcmp' functions instead of single precision.

   Listed among the code changes to MACS2 v2.1.1.20160309; the text presents no data on how these fixes alter peak calls. · MACS2 modifications