Crazy log-likelihood values from MACS2 `bdgdiff`.

Question

[ariel.balter]

I'm trying out differential binding analysis with macs2 bdgdiff. I've called peaks and gotten very reasonable values for fold-changes, p-values, q-values, etc. For example:

treat1_chip_repB.bed

1   10006   10499   treat1_chip_repB_peak_1 109 .   2.80877 14.06537    10.98392  
1   46683   47046   treat1_chip_repB_peak_2 46  .   3.94851 7.41653 4.69998  
1   180768  181015  treat1_chip_repB_peak_3 32  .   2.89084 5.85297 3.29158  
1   199870  200048  treat1_chip_repB_peak_4 26  .   3.09339 5.16864 2.68865  
1   358746  359816  treat1_chip_repB_peak_5 44  .   4.02818 7.07086 4.41624  
1   394998  395443  treat1_chip_repB_peak_6 19  .   3.15028 4.30817 1.91544  
1   450607  451066  treat1_chip_repB_peak_7 18  .   3.09737 4.20707 1.85041  
1   588143  588508  treat1_chip_repB_peak_8 28  .   3.55979 5.34418 2.82556  
1   590955  591428  treat1_chip_repB_peak_9 32  .   3.20037 5.81651 3.26295  
1   623994  624155  treat1_chip_repB_peak_10    23  .   3.24114 4.74840 2.32049  
. . .    
. . .

treat2_chip_repB.bed

1   9982    10520   treat2_chip_repB_peak_1 325 .   4.51491 35.86348    32.59083   
1   11213   11689   treat2_chip_repB_peak_2 31  .   3.22501 5.38701 3.18507   
1   13482   14198   treat2_chip_repB_peak_3 32  .   2.81719 5.44192 3.23710   
1   29155   29325   treat2_chip_repB_peak_4 12  .   2.53058 3.23237 1.29448   
1   46687   47138   treat2_chip_repB_peak_5 123 .   5.60019 15.02273    12.31674  
1   89078   89436   treat2_chip_repB_peak_6 15  .   2.65434 3.51344 1.52353  
1   91051   91364   treat2_chip_repB_peak_7 13  .   2.66162 3.34514 1.36895  
1   180765  181094  treat2_chip_repB_peak_8 112 .   4.36022 13.94062    11.26981  
1   184508  184657  treat2_chip_repB_peak_9 38  .   2.76260 6.11458 3.85310  
1   199058  200002  treat2_chip_repB_peak_10    26  .   3.03284 4.81645 2.68143  
. . .    
. . .

However, I'm getting crazy values for the log-likelihoods produced by macs2 bdgdiff.

diff_treat1_repB-vs-treat2_repB_c3.0_common.bed

1 9999 10495 diff_treat1_chip_repB-vs-treat2_chip_repB_common_1 -5626780945305257054830592.00000  
1 46752 47031 diff_treat1_chip_repB-vs-treat2_chip_repB_common_2 -8612834383639687743930368.00000  
1 180771 180951 diff_treat1_chip_repB-vs-treat2_chip_repB_common_3 -14910970167988796344238080.00000  
1 359329 359497 diff_treat1_chip_repB-vs-treat2_chip_repB_common_4 -14372019716837756622602240.00000  
1 590976 591413 diff_treat1_chip_repB-vs-treat2_chip_repB_common_5 -7164430050156399653027840.00000  
1 779550 779705 diff_treat1_chip_repB-vs-treat2_chip_repB_common_6 -16340788830357069758464000.00000  
1 804823 804947 diff_treat1_chip_repB-vs-treat2_chip_repB_common_7 -20745697224011715050471424.00000  
1 812089 812380 diff_treat1_chip_repB-vs-treat2_chip_repB_common_8 -16801092741071353413632000.00000  
1 815171 815897 diff_treat1_chip_repB-vs-treat2_chip_repB_common_9 -3286164094130363687763968.00000   
. . .    
. . .

[moreno.colaiacovo]

Hi

I am getting the same weird results. It's a very strange thing, since the bdgdiff function should have a default threshold for this ratio (logLR > 3). So negative values should be removed by default.

Have you managed to solve the issue? I'm interested too!

Best, Moreno

Answer 1

[yinyeya]

It is indeed an overflow issue most probably by the incorrect-compiled-dependenies. I used macs2.1.0 and it works perfectly one year before. But recently I met the same issue. I was troubleshooting several days until I noticed that I have unmatched numpy and python from miniconda env installed several months before (macs2 was installed before by root without miniconda env). Now I fixed this issue by creating a new conda env and installing macs2 there. See the link: https://github.com/taoliu/MACS/issues/156