henhaobio
Sign in

Molecular Ecology ResourcesGenomicsPeer-reviewed journal article

Dsuite - fast D-statistics and related admixture evidence from VCF files

Milan MalinskyiD, Michael MatschineriD, Hannes SvardaliD

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

Abstract

O_LIPattersons D, also known as the ABBA-BABA statistic, and related statistics such as the f4-ratio, are commonly used to assess evidence of gene flow between populations or closely related species. Currently available implementations require custom file formats and are impractical to evaluate all gene flow hypotheses across datasets with many populations or species. C_LIO_LIDsuite is a fast C++ implementation, allowing genome scale calculations of the D and f4-ratio statistics across all combinations of tens or hundreds of populations or species directly from a variant call format (VCF) file. Furthermore, the program can provide evidence of whether introgression is confined to specific loci and aid in interpretation of a system of f4-ratio results by implementing the f-branch method. C_LIO_LIDsuite is available at https://github.com/millanek/Dsuite, is straightforward to use, substantially more computationally efficient than other comparable programs, and presents a novel suite of tools and statistics, including some not previously available in any software package. C_LIO_LIThus, Dsuite facilitates assessment of evidence for gene flow, especially across large genomic datasets. C_LI

In-depth summary

Claude Opus 5.5 · xhigh effort · 8 Nov 2026 · from the full text of version 2 (13 Feb 2020) · CC BY 4.0

Dsuite, a C++ tool, computes D and f4-ratio statistics across all trios of tens to hundreds of taxa directly from VCF, plus windowed fd/fdM and f-branch.

Established D and f4-ratio programs (ADMIXTOOLS, HyDe, Comp-D) cannot read VCF, and quartet counts grow combinatorially with taxon number, raising compute costs and complicating interpretation. Dsuite works directly on VCF with far lower memory use, and adds windowed fd/fdM and the f-branch statistic to localize introgression and assign it to branches.

Key results

  1. ShownDsuite used the least memory on all three benchmark datasets and was fastest on the Malawi cichlid and large simulated data, although ADMIXTOOLS and HyDe ran faster on the small simulation.

    Large simulation (100 species, 100Mb): Dsuite needed <223MB RAM vs >1 Terabyte for ADMIXTOOLS and Comp-D and >18 Gigabytes for HyDe; memory gap vs ADMIXTOOLS and Comp-D stayed above two orders of magnitude on the other datasets; run-time gaps among Dsuite, ADMIXTOOLS and HyDe up to ~2-3 fold; Comp-D runs cancelled after 24hours. · Performance and results; Computational efficiency

  2. SuggestedOn the 20-species simulation, f-branch condensed correlated f4-ratio results into a few branch-level signals that recovered most branches involved in simulated gene flow.

    Ten of 568 f-branch signals exceeded 5%, vs 73 signals from raw trio analysis with BBAA ordering; these ten recovered seven of the nine gene-flow branches, and six pinpointed both branches of an event, e.g. (d, k), (e, j), (m, g), (c, b). · Results and interpretation; Figure legends

  3. SuggestedBecause trios share drift, a handful of gene flow events inflated D and f4-ratio across many trios, so a significant trio alone does not identify the admixing pair; BBAA-based topology inference can also err under gene flow.

    Small simulation (20 species, 1,140 trios, 5 gene-flow events): 546 (47.9%) D values significant even under the Dmin ordering; f4-ratio >5% for at least 48 trios; BBAA ordering disagreed with the true tree in 39 trios (3.4%), Dmin ordering in 124 (10.9%). · The Fbranch program; Results and interpretation; Figure legends

  4. ClaimedFbranch is presented as the first public implementation of the f-branch (fb(C)) statistic, which assigns gene flow evidence to possibly internal branches to tame the combinatorial growth of trio tests.

    With a fixed outgroup, trio count is n choose 3: 1,140 combinations for 20 taxa and 161,700 for 100 taxa; fb(C) aggregates f4-ratio results with A and B in P1/P2 and C in P3 on a user-supplied Newick tree; first-implementation status stated to the authors' knowledge. · The Fbranch program

  5. ClaimedBeyond genome-wide trio tests, Dsuite merges per-region runs with combined jackknife significance and computes fd and fdM in sliding windows to check whether admixture signal is localized.

    DtriosCombine sums site-pattern counts and f4-ratio denominators across Dtrios runs (e.g., per chromosome) and recomputes block-jackknife p-values; Dinvestigate windows hold a user-set number of informative SNPs; no windowed results are reported in the text. · The DtriosCombine program; The Dinvestigate program

Methods, data and reproducibility

Given only an outgroup, Dsuite sums allele-frequency-based BBAA, ABBA and BABA patterns for every trio, reports D and f4-ratio with block-jackknife Z-scores and p-values, computes fd/fdM in windows of informative SNPs, and maps correlated f4-ratios onto tree branches via f-branch.

  • Variants on the largest Metriaclima zebra reference scaffold (~16Mb) from a published Lake Malawi cichlid dataset (Malinsky et al. 2018) · Lake Malawi cichlid fishes · n = 73 species
  • Small simulated dataset generated with ms', 20Mb of sequence, 5 gene-flow events, diploids from paired haploid sequences, Ne = 1 outgroup · n = 20 species (1,140 trios)
  • Large simulated dataset, 100Mb of sequence, random directional admixture with proportions from a rescaled beta distribution (0% to 30%) · n = 100 species
Methods & tools
Dsuite (Dtrios, DtriosCombine, Dinvestigate, Fbranch), ADMIXTOOLS (with convertf), HyDe, Comp-D, ms' (simulation software), VCFtools v0.1.12b, vcf2phylip, Block-jackknife standard error estimation
Software & environment
Dsuite is implemented in C++; competitor inputs were prepared with VCFtools v0.1.12b (--plink), the ADMIXTOOLS convertf program (EIGENSTRAT), and vcf2phylip (PHYLIP).
Assay conditions & compute
Large simulation memory: Dsuite required less than 223MB vs over 1 Terabyte for ADMIXTOOLS and Comp-D and over 18 Gigabytes for HyDe; Comp-D runs cancelled after 24hours. Hardware not specified.
Reproducibility gaps
Benchmark hardware is not described; detailed timing, memory and simulation parameters are only in Tables 1 and 2; competitor format-conversion time was excluded from reported run times.
Limitations (5)
  • Significantly elevated D and f4-ratio values can arise from ancestral population structure or substitution rate variation rather than introgression, and reliably reconstructing full gene flow networks remains unresolved. · stated by the authors
  • f-branch does not reveal gene flow direction, and one event can yield several fb signals (a mirror-image signal plus correlated signals with relatives of the donor); the authors recommend 5-taxon tests for direction. · stated by the authors
  • Trio tests are not independent; users must correct p-values for multiple testing, and treating tests as independent under FDR control is only a conservative approximation. · stated by the authors
  • f-branch accuracy is evaluated on a single 20-species simulation without comparison to alternative approaches, and simulations omitted ancestral structure and rate variation, the very confounders the authors flag. · noted in this summary
  • The only empirical benchmark covered one ~16Mb scaffold of the cichlid data on unspecified hardware, so whole-genome runtime and memory scaling is extrapolated rather than measured. · noted in this summary

Read this if you need to test for admixture across many populations or species from VCF files and want help interpreting correlated D and f4-ratio results via f-branch.

Citations
998
FWCI
53.31
Field percentile
top 1%

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
THISLeppälä '24Elworth '18Zhang '26Leppälä '25Smith '26Soraggi '17
57% semantic matchGenetics
Kalle Leppälä, Flavio Augusto da Silva Coelho, Michaela Richter+2 more·Molecular Biology and Evolution·Sep 20, 2024· 6 citations