Interpretation of MSMC2/PSMC/betaPSMC/ngsPSMC

Question

[vap768]

Hello, I am new to demographic history inference using software like https://github.com/lh3/psmc and https://github.com/stschiff/msmc2. Since I am not very familiar with the algorithms that underlie these programs and have mostly a conceptual understanding of how they work, I have tried a variety of different programs on my data (7 high coverage (35x+) genomes from a non-model species and 190+ low coverage genomes). Since this is a non-model species - genetic maps etc. are not available, and I am unsure that imputing from our 7 high coverage genomes makes any sense since we just have 1 or 2 individuals sequenced at high coverage per population.

For imputing demographic history i have tried MSMC2, PSMC, beta-PSMC (https://github.com/ChenHuaLab/Beta-PSMC) (for more recent history), and ngsPSMC(https://github.com/ANGSD/ngsPSMC), for the low coverage genomes (5x). I was hoping to convince myself of true demographic history with similar results from each of these programs. However, I have quite dramatically different results for different programs (and also using different -p flags to set intervals). For each program I have tried to set similar time segmentation patterns and check for overfitting following some advice here for psmc (https://github.com/lh3/psmc/issues/45), and following this guidance the models don't seem overfit. My question is, how should I interpret differences such as initial rising Ne with my red populations, for instance (this using -p 16+581 and 25 iterations) vs declines in these populations with other paramateers (ie -N25 -p -r5 "262+4+7+1 or betaPSMC parameters -p 16+58*1 -i 25). See attachments for examples. Are these artifacts or potentially real changes that occur because I have increased my resolution. How can I check this? Is the big difference in Ne values (ie sometimes 10x greater) between MSMC and PSMC expected?

other notes: currently using unphased genomes (one at a time in MSMC2), with mask files created as documented in MSMC instructions, but also planning to try phasing genomes (with whatshap since we have no map files for shape_it) in the two-sample mode to look at divergence times between two populations of interest.

Thanks! and please forgive my ignorance and confusion.

Answer 1

[csarabia]

Hi,

Your question is very intriguing (no ignorance or confusion to pardon!). You are pointing out exactly the reason why PSMC is normally not trustable for time intervals below 20-30kyr ago. However, ngsPSMC seems to work well for more recent times (we tested it here https://onlinelibrary.wiley.com/doi/full/10.1111/mec.15784).

How to get rid of those later extreme bumps in Ne in PSMC? Check out Hilgers et al. 2025 (https://www.sciencedirect.com/science/article/pii/S0960982224012399) who seem to have answered your question.

What they suggest is using, as you have done in your first plot, a smaller segment size in the beginning of the graph. Instead of the default recommended, "4+25x2+4+6", they use a “2+2+25x2+4+6” or a “1+1+1+1+25x2+4+6” and have much better results. Perhaps, in your case, I would test a modification of your third plot, "2+2+60x1". Let us know how it went!