Thanks to all of our SMaHT colleagues and especially to @sedlazeck.bsky.social who led the hackathon which spawned the prototype of this pipeline!

MosaicSim offers a realistic, scalable approach for assessing detection limits, with immediate applications to large sequencing efforts including those within the SMaHT Network, which was the springboard for this work.

A key (surprising) result was that ultra-high coverage (300×–450×) yields diminishing returns for mosaic variant detection. In many settings, 150× coverage performs comparably or better, highlighting opportunities for cost-effective study design.

Using MosaicSim, we benchmarked DRAGEN and found strong VAF- and depth-dependent performance limits. Sensitivity decreases sharply at low VAF, especially in complex genomic regions.

Detecting mosaic variants is challenging due to low VAFs and real sequencing noise. MosaicSim layers user-defined variants directly onto empirical WGS data, preserving true read-level properties while providing a controlled ground-truth set for benchmarking.

So since we only include >0.1% MAF variants in this article we can't address ultrarare, but check out Supp Fig 3; when comparing ancestry-specific AFs many variants deviate from the 1:1 line. We plotted this on the log₁₀(AF) scale to help magnify the low-frequency range.

To limit the noise from ultra-rare alleles we only looked at variants ≥0.1% MAF. Totally appreciate that's still quite low frequency, but even with that filter, we still saw the noted ancestry-specific frequency differences.

Great point; we thought about that too! Pragati stratified by whether variants were monomorphic or not to capture at least that aspect, but you’re right that the impact depends on where a variant sits on the SFS. Rare ones can show big fold-changes but small absolute shifts.

Thanks for the interest! The tutorial code is available to download as supplemental information of the paper, and has been deposited as a community workspace in the All of Us Researcher Workbench.

In summary, we present a replicable training model that empowers early-career researchers - including and especially those new to computational genomics - to responsibly leverage large-scale biobank data into their research programs and teaching.

From years 1–3, training outcomes reported by scholars to stem directly from this training included:
📊 17 conference presentations
🔬 Multiple funded research grants
🎓 Numerous genomics modules added in undergrad courses
🤝 Sustained collaborations across institutions

During the summit, scholars used real short-read WGS data to:
• Prepare phenotypes & covariates
• Run GWAS via Hail
• Visualize results with PCA, Manhattan & QQ plots
• Manage compute costs
All in ~4 hours with no prior coding required.

Our training was part of the All of Us Biomedical Researcher Scholars Program through @bcmgenetics.bsky.social focused on mentoring early-stage faculty in genomic data science. The curriculum launches with an intensive Faculty Summit, where scholars get hands-on experience working with genomic data.

Access to big genomic data is growing, but parallel access to skills needed to use it hasn’t kept up.
We created an accessible, cloud-based genomic analysis training bootcamp using real All of Us data, Jupyter notebooks, and the Hail framework to lower the barrier for early-career researchers.

Tractor-Mix builds on Tractor’s strengths to detect ancestry-enriched signals while adding power and robust false-positive control for relatedness via a GRM. By modeling both admixture and relatedness, it overcomes key GWAS barriers and enables more accurate, representative genomic discovery.

Tractor-Mix uses ancestry-specific genotypes as predictors, outputting ancestry-specific effect sizes and P values. We benchmark our new tool in simulations and apply it to multiple admixed cohorts (including UKBiobank and Mexico City Prospective Study), uncovering signals missed by standard GWAS.

In this work, we introduce Tractor-Mix, a new GWAS method that extends Tractor to handle related admixed samples. It combines a mixed model framework (like GMMAT) with local ancestry-aware genotypes (like Tractor) in a 2 d.o.f. test.

As biobanks and global cohorts grow, so does the inclusion of admixed individuals with close or cryptic relatedness. This introduces the statistical challenge of two interwoven sources of stratification: admixture and relatedness, which are rarely handled together.