Substack cometh, and lo it is good. (Pricing)

Assessing the utility of models in ancient DNA admixture analyses

Assessing the Performance of qpAdm: A Statistical Tool for Studying Population Admixture:

qpAdm is a statistical tool for studying the ancestry of populations with histories that involve admixture between two or more source populations. Using qpAdm, it is possible to identify plausible models of admixture that fit the population history of a group of interest and to calculate the relative proportion of ancestry that can be ascribed to each source population in the model. Although qpAdm is widely used in studies of population history of human (and non-human) groups, relatively little has been done to assess its performance. We performed a simulation study to assess the behavior of qpAdm under various scenarios in order to identify areas of potential weakness and establish recommended best practices for use. We find that qpAdm is a robust tool that yields accurate results in many cases, including when data coverage is low, there are high rates of missing data or ancient DNA damage, or when diploid calls cannot be made. However, we caution against co-analyzing ancient and present-day data, the inclusion of an extremely large number of reference populations in a single model, and analyzing population histories involving extended periods of gene flow. We provide a user guide suggesting best practices for the use of qpAdm.

The Reich lab provides its software and data. It’s really not that hard to replicate and tweak some of the analyses they do in their papers (check the supplements for the detailed specifications of the parameters). I’ve done many times when I got curious about a detail they hadn’t explored.

The preprint above is a valuable addition to the intuitions one can develop through using the packages.

9 thoughts on “Assessing the utility of models in ancient DNA admixture analyses

  1. One thing that this left me with an open question about is “co-analyzing ancient and present-day data”. This could many many of:

    A) Do not use both ancient and modern populations together in pright and pleft… but it is OK to use ancients in pright and moderns in pleft (or vice versa).
    B) Use only either ancients or moderns in both pright and pleft.
    C) Where ancients are target only use ancients in both pright and pleft, or for moderns vice versa.
    D) You can use combinations of ancients and moderns where-ever you wish, however do not complete runs where you use the same set up but vary the target to be modern or ancient.

    Some of these would contradict some of the modelling completed in published papers (for an instance, Narasimhan’s paper used AHG as a vital population, and co-modeled ancients and moderns with the same qpAdm setups, etc.)

    So what exactly does this mean?

  2. So, can it be summarized as don’t mix ancients and moderns (as Matt said) and don’t overfit?

  3. Possibly a bit off-topic here but someone took info from Reich to model Covid-19 rates by ancestry, finding that increased WHG ancestry was associated with increased mortality.

    https://www.medrxiv.org/content/10.1101/2020.04.05.20054627v1

    …We found significant positive correlation (p=0.03) between European Mesolithic hunter gatherers (WHG) ancestral fractions and COVID-19 death/recovery ratio…

    …SNPs associated with Interferon stimulated antiviral response, Interferon-stimulated gene 15 mediated antiviral mechanism and 2′-5′ oligoadenylate synthase mediated antiviral response show large differences in allele frequencies between Europeans and East Asians…

  4. Thanks for highlighting our paper Razib! I’m glad it is of interest!

    To try to answer a question raised in the comments, in this study we explored the effect of having C-to-T damage in a subset of populations included in a model (see Figures 3D and 7 in the paper). In cases where all the Left Pops (i.e. target and sources) have the same rate of damage (either 5% or 0%), the admixture proportion estimates and p-values appear unaffected, but in cases where the target and source populations have different damage rates, the admixture proportion estimates are biased and the p-values assigned to models that should otherwise be deemed plausible are reduced. So by mixing ancient and present-day populations in the Left population list you run the risk of rejecting models that should otherwise be accepted.

    These issues are likely the result of attraction between Right and Left populations that have similar damage rates, so while we didn’t explicitly explore the effect of having Right populations with different damage rates, the effect is likely to be similar. Based on these results, in general, we recommend that users avoid mixing populations with very different damage rates (as would be observed between ancient and present-day populations) in either the Left or Right population sets, or that users restrict analyses to transversions, which aren’t susceptible to ancient damage.

  5. Yes, that’s a great answer, thanks.

    One related question, if possible, it might be completely outside the scope of your paper, but are there any specific issues that would be expected with using shotgun and capture samples together in either / both of pleft / pright?

    Specific as in beyond having higher rates of missing data / more damage in shotgun samples (e.g. even should shotgun and capture samples have matched missingness and damage).

    (Question prompted by recently say some shotgun samples vs 1240k capture European Neolithic samples of broadly the same cultures / region+time. It seemed to be the case that the capture samples had higher attraction to the Anatolian Neolithic than the shotgun, and further that attraction fell away faster in with additional HG admixture into shotgun samples than capture.
    This could all just be because higher damage / missing among shotgun samples, but if presuming, that’s not the case…)

  6. Thanks for your question, Matt!

    This isn’t something that we explicitly modeled, so I won’t attempt to speculate too much. In our analysis, we did find that the results of qpAdm are not affected significantly when you use a non-random ascertainment scheme to select which SNPs are analyzed. However, when we did this, we applied the same ascertainment schemes across all populations. It would definitely be interesting to use different ascertainment schemes to select data for different populations, mirroring the effect of having some data sequenced via capture and some via shotgun sequencing.

    My guess would be that the results of qpAdm will be affected if some of the populations are captured and some are shotgun sequenced, as populations that exhibit similar patterns of missing data may appear more or less similar than they actually are. So I would advise that users exercise similar caution when using these different types of data.

  7. Hi Ms Harney. Thanks for replying to Matts questions. They help clear a lot of doubts. I have a few additional doubts. It would be great if you could clear them.

    1. Is it correct to say that the options allsnps:YES is better than NO in all cases and will lead to lesser false positives? In my use of qpAdm, allsnps uses 115k SNPs whereas allsnps:NO hardly ever goes over 700k SNPs for 1240k dataset samples.

    2. I see that when missing SNPs was 25%, no non optimal model passed the threshold (with +ve coefficients) and only models with sources 5 & 9 passed. Can this be taken as a firm conclusion? will it be fair to say that when all samples in left & right pops have >900k SNPs, the result will never give a false positive?

    I am also confused a bit because as per table 1 in supplementary excel, model with sources 4 & 9 gives pvalue>0.05 for 4622/5000 replicates. Whereas in table 13, in the case when 25% SNPs are set to missing, 0/45 replicates with sources 4 & 9 have pvalues > 0.05. Please explain what I am missing.

    3. The preprint states “However, when
    455 the target and source populations have differential rates of damage, this optimal model is almost
    456 always deemed implausible.”

    I do have few models with a mixture of ancient and modern sources in pleft that are successful, most do not work. would that mean that those successful models are optimal?

    4. Would it be possible to treat modern samples and create an alternate eigenstrat database in such a way that the damage replicates that of aDna making it possible to use both types of samples in qpAdm?

    Thank you.

Comments are closed.