Question: XP-nSL maf filtering and interpretation of normalisation

[erwan-029]

Dear Prof. Szpiech,

Thank you very much for the release of version v2.1.0 of selscan, very fast indeed, useful for rerunning analyses when attempting to figure out what going on.

Despite reading some related issues, I would still appreciate your input on few questions regarding XP-nSL.

I understand that XP-nSL, by default, doesn't use any MAF filtering (re issue #59) and it seems the flag --maf 0.05 is not having any impact either. Is there now any native option in selscan for doing that, à la nSL? I'm asking because I've ran XP-nSL on 1KGP data, comparing BEB to CEU, and one of the top hits has an allele count of 1 in BEB, which seems quite unlikely to be anything truly genuine.

I’ve used the same parameters as in the Szpiech et al. 2021 for normalisation (see an example below of the output), but maybe I haven’t looked at the results in the right way. I’ve calculated the top 1% cutoff from the “fracAbove” <= 0 values distribution and then, within windows with a “fracAbove” above that threshold, I looked only at variants that had a “normedxpnsl” above the top 1% cutoff from its absolute distribution.

In short, it’s probably advisable to filter rare/very rare alleles, isn’t it and what is the proper way to filter the hits?

Thank you in advance for your understanding.

norm v1.3.1
You have provided 22 output files for joint normalization.
[…]

Total loci: 59560256
Reading all data.
Calculating mean and variance:

num mean variance
58809178 -0.0400236 0.00764802
[…]

Analyzing BP windows:

26740 windows with nSNPs >= 10.

High Scores
nSNPs 1.0 5.0
1875.9 0.464058 0.168618
1988 0.447101 0.132186
2067 0.400989 0.109858
2139 0.350682 0.114374
2203 0.347796 0.132649
2274 0.334948 0.119122
2356 0.324269 0.111103
2457 0.295938 0.108517
2635 0.323416 0.128354
11742 0.312676 0.118949

Low Scores
nSNPs 1.0 5.0
1875.9 0.466708 0.140141
1988 0.435886 0.178326
2067 0.495667 0.187139
2139 0.461146 0.185582
2203 0.495022 0.186689
2274 0.491401 0.182324
2356 0.442287 0.16983
2457 0.454321 0.176724
2635 0.429901 0.195696
11742 0.448345 0.187873

[szpiech]

Hello, So, in my own personal testing some time ago, I found that filtering for MAF when applying the two population statistics (XP-EHH/XP-nSL) hindered power quite a bit. This is why I didn't implement it in selscan. Regarding your observation of a top score having a MAC of 1. It isn't clear to me which allele is the derived allele in this example, but, more to the point, these statistics perform best when considering windows enriched for extreme scores, instead of considering individual outliers. It is possible that the major allele is the sweeping mutation or is in LD with a sweeping mutation. I'm not sure I follow your next observation/question about the window-based normalization output. Based on the output you provided, you likely have some output files that look like *.windows. In these files there are columns that indicate whether a window is in the top 1% of windows or not, for both positive scores and negative scores. Typically i recommend taking these windows as supporting evidence for positive selection in one vs the other population. Let me know if you have more questions. Zachary

…

On Wed, Jul 16, 2025 at 3:48 AM erwan-029 ***@***.***> wrote: *erwan-029* created an issue (szpiech/selscan#147) <#147> Dear Prof. Szpiech, Thank you very much for the release of version v2.1.0 of selscan, very fast indeed, useful for rerunning analyses when attempting to figure out what going on. Despite reading some related issues, I would still appreciate your input on few questions regarding XP-nSL. I understand that XP-nSL, by default, doesn't use any MAF filtering (re issue #59 <#59>) and it seems the flag --maf 0.05 is not having any impact either. Is there now any native option in selscan for doing that, à la nSL? I'm asking because I've ran XP-nSL on 1KGP data, comparing BEB to CEU, and one of the top hits has an allele count of 1 in BEB, which seems quite unlikely to be anything truly genuine. I’ve used the same parameters as in the Szpiech et al. 2021 for normalisation (see an example below of the output), but maybe I haven’t looked at the results in the right way. I’ve calculated the top 1% cutoff from the “fracAbove” <= 0 values distribution and then, within windows with a “fracAbove” above that threshold, I looked only at variants that had a “normedxpnsl” above the top 1% cutoff from its absolute distribution. In short, it’s probably advisable to filter rare/very rare alleles, isn’t it and what is the proper way to filter the hits? Thank you in advance for your understanding. norm v1.3.1 You have provided 22 output files for joint normalization. […] Total loci: 59560256 Reading all data. Calculating mean and variance: num mean variance 58809178 -0.0400236 0.00764802 […] Analyzing BP windows: 26740 windows with nSNPs >= 10. High Scores nSNPs 1.0 5.0 1875.9 0.464058 0.168618 1988 0.447101 0.132186 2067 0.400989 0.109858 2139 0.350682 0.114374 2203 0.347796 0.132649 2274 0.334948 0.119122 2356 0.324269 0.111103 2457 0.295938 0.108517 2635 0.323416 0.128354 11742 0.312676 0.118949 Low Scores nSNPs 1.0 5.0 1875.9 0.466708 0.140141 1988 0.435886 0.178326 2067 0.495667 0.187139 2139 0.461146 0.185582 2203 0.495022 0.186689 2274 0.491401 0.182324 2356 0.442287 0.16983 2457 0.454321 0.176724 2635 0.429901 0.195696 11742 0.448345 0.187873 — Reply to this email directly, view it on GitHub <#147>, or unsubscribe <https://github.com/notifications/unsubscribe-auth/ABAKRQVR2TUWURS5K7Q43MT3IX7UNAVCNFSM6AAAAACBUCTX32VHI2DSMVQWIX3LMV43ASLTON2WKOZTGIZTIOBXGY2TQOI> . You are receiving this because you are subscribed to this thread.Message ID: ***@***.***>

[erwan-029]

Thank you for your prompt reply. I'll try to articulate my questions/observations better.

I think it all boils down to how to figure out what are the windows of interest. And upon further inspection, I suspect my problem lay in that I mislabeled the columns from the normalisation output. Indeed columns 6 and 7 report 100, 5, 1 and -1 values. Is it correct to assume then that column 6 is for positive values and 7 for negative ones? (Edit: issue #51 gave me the answer, they are so).

Eventually, I would also be interested in attempting to fine-map the results, and that was the rationale for looking at outliers so to speak. That said, your point about LD is even more valid in this context. Should I then understand that I shall simply take more of an eagle view of the region(s) rather than attempting to "pin" some causal variant?

And simply for sake of completeness, the 1 MAC was on the following window (with my wrongly inferred header)
win_start win_end nb_snps fracAbove fracBelow percAbove percBelow maxScore minScore
98200001 98300001 2081 0.802499 0.000000 1 100 4.24560 -0.114865
With the variant itself having an normedxpnsl of 2.95149, but being a stopgain, it was an interesting one.

[szpiech]

Hello, So typically I suggest taking the top 1% windows, this would be in the percAbove or percBelow columns with a 1 in them (which you've correctly identified). The extreme windows marked by percAbove imply a signal in the non-ref population and the extreme windows marked by percBelow imply a signal in the ref population. For localizing the signal, you will need to use a method like SWIF(r) ( https://www.nature.com/articles/s41467-018-03100-7), as the statistics implemented in selscan do not provide enough resolution individually. Zach

…

On Wed, Jul 16, 2025 at 1:55 PM erwan-029 ***@***.***> wrote: *erwan-029* left a comment (szpiech/selscan#147) <#147 (comment)> Thank you for your prompt reply. I'll try to articulate my questions/observations better. I think it all boils down to how to figure out what are the windows of interest. And upon further inspection, I suspect my problem lay in that I mislabeled the columns from the normalisation output. Indeed columns 6 and 7 report 100, 5, 1 and -1 values. Is it correct to assume then that column 6 is for positive values and 7 for negative ones? Eventually, I would also be interested in attempting to fine-map the results, and that was the rationale for looking at outliers so to speak. That said, your point about LD is even more valid in this context. Should I then understand that I shall simply take more of an eagle view of the region(s) rather than attempting to "pin" some causal variant? And simply for sake of completeness, the 1 MAC was on the following window (with my wrongly inferred header) win_start win_end nb_snps fracAbove fracBelow percAbove percBelow maxScore minScore 98200001 98300001 2081 0.802499 0.000000 1 100 4.24560 -0.114865 With the variant itself having an normedxpnsl of 2.95149. — Reply to this email directly, view it on GitHub <#147 (comment)>, or unsubscribe <https://github.com/notifications/unsubscribe-auth/ABAKRQVU3AATIUMLKLWARZD3I2GYBAVCNFSM6AAAAACBUCTX32VHI2DSMVQWIX3LMV43OSLTON2WKQ3PNVWWK3TUHMZTANZZGY3TEMJWGA> . You are receiving this because you commented.Message ID: ***@***.***>

[erwan-029]

Thank you for confirming the columns as well as to how to interpret the side where the signal is present.

Thank you as well for bringing to my attention the SWIF(r) method and package, I'll get cracking with.

Merci beaucoup !

Erwan