henhaobio
Sign in

bioRxivBioinformaticsPreprint

Recomb-Mix Cloud: a cost-efficient and scalable cloud-friendly pipeline for local ancestry inference

Yuan WeiiD, Degui ZhiiD, Shaojie ZhangiD

PDFDOICodeData
In-depth full-text digest available · Jump to digest ↓

Abstract

Local ancestry inference (LAI) reveals population compositions of an individuals chromo-somes, which empowers population genetics and genetic association studies of diverse populations. The recent availability of large genomic datasets from biobanks offers unprecedented opportunities for LAI. However, such datasets are typically only accessible from cloud-based platforms, and the lack of efficient and cost-effective tooling makes LAI prohibitive for resource-constrained researchers. Here, we develop Recomb-Mix Cloud, a cloud-ready pipeline for LAI based on Recomb-Mix, a recently introduced efficient LAI method. Recomb-Mix focuses on individual variants while maintaining a lower computational burden than other LAI methods focusing on genomic windows. Recomb-Mix Cloud offers a cloud-friendly complete pipeline from panel sample selection, marker alignment, compact reference panel construction, Recomb-Mix threading, and generating LAI calls in common data formats. Recomb-Mix Cloud also offers a set of pre-computed compact population reference panels constructed based on data from the Human Genome Diversity Project and the 1000 Genomes Project. Our results showed Recomb-Mix is about one and two orders of magnitude cheaper and faster than FLARE and RFMix, two popular LAI methods, while offering largely consistent LAI calls on the All of Us Researcher Workbench. The estimated overall compute cost per All of Us sample for whole-genome LAI analysis using Recomb-Mix Cloud with a pre-computed reference panel is just $0.01, compared with $0.43 for FLARE and $4.58 for RFMix. Therefore, Recomb-Mix Cloud is well-positioned to democratize local ancestry-based genetics research on modern biobank-sized data. The Recomb-Mix Cloud pipeline and the compact reference panels are available at https://github.com/ucfcbb/Recomb-Mix.

In-depth summary

Claude Opus 5.5 · xhigh effort · 8 Nov 2026 · from the full text of version 1 (7 Oct 2026) · CC BY 4.0

On the All of Us workbench, a Recomb-Mix-based cloud pipeline runs whole-genome local ancestry inference for an estimated $0.01 per sample, versus $0.43 for FLARE and $4.58 for RFMix.

Biobank genomes such as All of Us are typically reachable only through cloud platforms, where local ancestry inference (LAI) is costly (a prior RFMix run on 53,944 individuals cost $400 for LAI alone), reference panels often must be self-built, and output formats differ by tool. This work packages the site-based Recomb-Mix method as a cloud pipeline with pre-computed compact Human Genome Diversity Project/1000 Genomes panels (GRCh38, GRCh37) and converters to ancestry dosage and state formats.

Key results

  1. ShownOn 200 All of Us chromosome 21 samples, Recomb-Mix ran local ancestry inference much faster and cheaper than FLARE or RFMix on the same cloud hardware.

    All methods ran on 32 CPUs and 208 GB RAM at $1.91/hour. Runtimes were 69 s for Recomb-Mix, 2,175 s for FLARE and 23,067 s for RFMix. Recomb-Mix cost $0.04, about 3% of FLARE's cost and 0.3% of RFMix's. The whole pipeline for these 200 samples, including $9.50 of PLINK 2 preprocessing, cost $9.54 ($0.05 per individual). · 3.4 Cost; 3.3 Marker alignment

  2. SuggestedScaled up to the whole genome and the full All of Us cohort, the estimated cost is about $0.01 per individual with Recomb-Mix Cloud, against $0.43 for FLARE and $4.58 for RFMix.

    Chromosome 21 costs were scaled linearly by variant count (658,219,451 variants; an estimated 63,569,360 overlapping sites) and by individual count (414,830). Estimated cohort totals are $4,859.43 for Recomb-Mix, $179,088.03 for FLARE and $1,900,632.53 for RFMix. Whole-genome preprocessing is estimated at $711.13. · 3.4 Cost; Abstract

  3. SuggestedIn simulations, Recomb-Mix was more accurate than FLARE and RFMix in all seven scenarios, including the smallest reference panel and the oldest admixture.

    The simulations used 200 admixed individuals from three-way African/East Asian/European chromosome 18 data. At 15 generations, panel sizes were 100, 250, 500 and 1,000; with 500 references, admixture was 50, 100 or 200 generations old. The metric was mean diploid squared Pearson r^2 across ancestries, as defined in the FLARE paper. The values are only in Supplementary Table S3. · 3.1 Accuracy

  4. ShownCompact reference panels and run-length encoded output use far less storage and still keep every per-position ancestry call.

    The eight-way whole-genome compact panels total 1.06 GB, 5.84% of the 18.14 GB original panels. For 200 simulated individuals (chromosome 18, three-way, 15 generations, 100 references), run-length output was 77 KB, the smallest format tested, with 100% of calls retained. FLARE's compressed VCF was larger, and RFMix's matrix needs the query file to recover per-position calls. · 3.3 Disk usage

  5. SuggestedOn real All of Us chromosome 21 data, the three methods produced similar ancestry tract-length distributions, roughly exponential in shape. Recomb-Mix was closer to RFMix than to FLARE.

    The data were 200 random individuals and 849,229 markers shared with the eight-way panel, in 40,000 bp bins. Histogram overlap was 0.8784 for Recomb-Mix vs FLARE, 0.9205 for Recomb-Mix vs RFMix and 0.8957 for FLARE vs RFMix. FLARE (min-mac=0, min-maf=0) labelled 95.91% of markers; the remaining markers took labels from neighbouring calls. · 3.5 Ancestry tract length distribution; Figure legends

Methods, data and reproducibility

A cloud pipeline that uses PLINK 2 to keep target samples and the biallelic sites shared with a compact reference panel. It then runs Recomb-Mix, a site-based approximation of the Li and Stephens hidden Markov model that uses haplotype-compressed panels, and converts the run-length encoded calls to dosage or state formats. It was benchmarked against FLARE and RFMix on simulated and All of Us data.

  • Human Genome Diversity Project (HGDP) and 1000 Genomes Project samples, merged, phased and imputed with Beagle 5.4 to build compact reference panels (GRCh38; lifted over to GRCh37) · human (Human Genome Diversity Project, 1000 Genomes Project) · n = 130,207,294 sites
  • Simulated three-way (African, East Asian, European) admixed chromosome 18 data; reference panels of 100 to 1,000 samples; 15 to 200 generations since admixture · human (African, East Asian, European ancestries) · n = 200 admixed individuals
  • All of Us phased genomic data, chromosome 21 (8,793,215 sites; 849,229 retained after overlap with the eight-way panel) · human (All of Us Research Program) · n = 200 randomly selected individuals
  • All of Us version 8 phased genomic dataset, used only for whole-genome and whole-cohort cost extrapolation · human (All of Us Research Program) · n = 658,219,451 variants; 414,830 individuals (cohort size in cost formula)
Methods & tools
Recomb-Mix, Recomb-Mix Cloud, Recomb-Mix Utility, PLINK 2, Beagle 5.4, CrossMap, FLARE 0.6, RFMix version 2, All of Us Researcher Workbench
Software & environment
Named versions: PLINK 2, Beagle 5.4, FLARE 0.6 and RFMix version 2. No Recomb-Mix version, container or dependency file is described.
Assay conditions & compute
All of Us cloud environment with 32 CPUs and 208 GB RAM at $1.91/hour, sized for RFMix (needed at least 200 GB). Chromosome 21 preprocessing took 17,910 s wall-clock and cost $9.50. LAI on 200 individuals took 69 s (Recomb-Mix), 2,175 s (FLARE) and 23,067 s (RFMix).
Reproducibility gaps
All of Us data: no route given beyond the Researcher Workbench. The simulation pipeline is only cited. Per-scenario r^2 values are only in Supplementary Table S3, and the file sizes of the other output formats and the per-method costs are in Tables 2 and 3, whose contents are not in the text.
Limitations (4)
  • The standard panels, built from HGDP and 1000 Genomes, miss many small, isolated or intra-continental populations (e.g. indigenous Americans). Such ancestries would get the labels of geographically nearby groups, which may be inaccurate. · stated by the authors
  • Results depend on how many variants the reference panel shares with the target. Only 9.66% of All of Us chromosome 21 variants overlapped the panel, so the calls cover a small part of the genome, and labels estimated for the remaining variants are preliminary. · stated by the authors
  • Accuracy was measured only in simulation (one chromosome, three continental ancestries, values only in a supplementary table). On real All of Us data there is no truth set, only tract-length distribution overlap, although the Figure 1 legend claims superior accuracy on All of Us data. · noted in this summary
  • Whole-genome and cohort-wide costs are linear extrapolations from one chromosome and 200 individuals; no full-genome or cohort-scale run was actually performed. · noted in this summary

Read this if you need local ancestry calls for many genomes inside All of Us or another cloud biobank and want a cheaper, packaged alternative to RFMix or FLARE with ready-made reference panels.

Citations
0

Reproducibility ledger

Nobody has reported trying to reproduce this paper yet. Did you replicate an experiment or run the code? Your report saves the next lab a week.

Sign in to file a reproducibility report.

Related papers

Similarity Map
SelectedSame fieldCross-field
THISWei '25Diem '26Browning '23Mauer '25Smetanin '20Tretmanis '26
77% semantic matchBioinformatics
Yuan Wei, Degui Zhi, Shaojie Zhang·Bioinformatics·Apr 26, 2025· 7 citations