Skip to main content

DAGIP: alleviating cell-free DNA sequencing biases with optimal transport.

Passemiers A, Tuveri S, Jatsenko T, Vanderstichele A, Busschaert P, Coosemans A, Timmerman D, Tejpar S, Vandenberghe P, Lambrechts D, Raimondi D, Vermeesch JR, Moreau Y

Publication: Genome Biol (2025) • Vol 26 • Page 49PMID: 40055826View on PubMed

Purpose

This paper developed a method (DAGIP) to correct cell-free DNA (cfDNA) next-generation sequencing (NGS) data obtained using different workflows (that included different collection tubes, extraction methods, library preparation kits and sequencing methods) to allow for joint analysis of multiple studies. The developed method was compared to other methods of normalization (baseline approach, MappingTransport, DryClean, Optimal transport).

Conclusion

In the initial analysis, the authors calculated the Pairwise distances between 66 specimens analyzed using both the TruSeq Nano and Kapa HyperPrep Kits and found that only 13 of the 66 were paired properly by specimen. After data normalization using DryClean, center-and-scale, and optimal transport, only a portion of the 66 specimens were properly paired (13, 25 and 41 specimens, respectively). Analysis of all the different specimen pairs in the different studies found the best pairing for each study occurred when the data was normalized using the Optimal transport (50-100% of pairs correct). Optimal transport normalization was developed further to adapt for fragment length ratio, nucleosome position score, end motif, coverage profile, and or GC content to produce the DAGIP normalization. In DAGIP normalization, pairing accuracy was lowest when different library preparation methods were used and highest when the only difference was between sequencing instruments. DAGIP produced the highest pairing accuracy and coefficient of determination (R2, highest proximity between coverage profiles), and lowest mean absolute error and allowed for co-clustering of specimens from healthy individuals processed using very different workflows. Use of DAGIP also significantly improved copy number profile accuracy (87%). When all five cancer types were considered, the highest mean Area Under the Receiver Operating Characteristic curve (AUROC) was found when center and scale were used, followed by DAGIP, but the Matthews Correlation Coefficient (MCC) was highest when DAGIP was used.

Studies (1)

Purpose

This study developed a method to correct cell-free DNA (cfDNA) next-generation sequencing (NGS) data that was obtained using different workflows (that included different collection tubes, extraction methods, library preparation kits and sequencing methods) to allow for joint analysis of multiple studies and compared the method to other methods of normalization.  This study used sequencing data obtained during four prior studies: 563 plasma specimens from non-invasive prenatal testing; 179 cases of Hodgkin lymphoma, 37 cases of diffuse large B-cell lymphoma, and 22 cases of multiple myeloma and 499 controls from the HEMA study; 550 ovarian cancer and 118 control cases as part of the ovarian cancer (OV) study; and 74 female controls and 51 breast cancer cases from a fragmentomics study (FRAG). Most studies used multiple workflows. Blood was collected into Streck cfDNA BCT or Roche Cell-Free DNA Collection Tubes, and cfDNA was extracted using the QIAamp Circulating Nucleic Acid Kit or the Maxwell automated protocol. DNA was batch-sequenced using the Illumina HiSeq 2000, HiSeq 2500, HiSeq 4000, or NovaSeq 6000 platforms. Sequencing was single-end (1 × 36 bp, 1 × 50 bp) or paired-end (2 × 50 bp, or 2 × 150 bp). Bioinformatics pipelines differed among the studies. Data were corrected using the baseline approach, MappingTransport, DryClean, Optimal transport (correction based on Wasserstein distance), and a new method termed DAGIP that corrects Optimal transport.

Summary of Findings

In the initial analysis, the authors calculated the Pairwise distances between 66 specimens analyzed using both the TruSeq Nano and Kapa HyperPrep Kits and found that only 13 of the 66 were paired properly by specimen. After data normalization using DryClean, center-and-scale, and optimal transport, only a portion of the 66 specimens were properly paired (13, 25 and 41 specimens, respectively). Analysis of all the different specimen pairs in the different studies found the best pairing for each study occurred when the data was normalized using the Optimal transport (50-100% of pairs correct). Optimal transport normalization was further developed to adapt for fragment length ratio, nucleosome position score, end motif, coverage profile, and/or GC content to produce the DAGIP normalization method. In DAGIP normalization, pairing accuracy was lowest when different library preparation methods were used and highest when the only difference was between sequencing instruments. The pairing accuracy was highest when DAGIP normalization was applied (0.664), followed by center and scale (0.638), baseline (0.567), MappingTransport (0.462) and DryClean (0.180) normalization methods. Importantly, DAGIP pairing was significantly better than baseline, MappingTransport and DryClean (P<0.005, all).  Importantly, after DAGIP principal component analysis, specimens no longer clustered by protocol. The coefficient of determination (R2) among protocols was highest after DAGIP, indicating the highest similarity occurred between coverage profiles. Further, the coefficient of determination (R2) was significantly higher after DAGIP normalization than when normalization was performed with any other method (P<0.005).  Applying DAGIP to normalize data sets produced using very different library preparation methods (the KAPA HyperPrep Kit and the NEBNext Enzymatic Methyl-seq Kit), which produced very different end motif profiles, nucleosome positioning scores and proportions of long fragments, improving clustering of specimens from healthy individuals in principal component analysis.  The mean absolute error was also lowest when DAGIP was applied, which was attributed to DAGIP being less likely to under- or over- correct than the other methods evaluated. After correction with any of the other normalization methods, there was ≥20% inconsistency in copy number profiles; this was not correlated with tumor fraction. Use of DAGIP significantly improved copy number profile accuracy (87%) and tumor ploidy estimates when results were compared to center-and-scale, MappingTransport, and DryClean normalization methods (P<0.001, all). When compared across protocols, significant improvements in copy number accuracy and tumor ploidy estimates were limited to comparisons drawn with the DryClean method (P<0.001, both). When all five cancer types were considered, the highest mean AUROC occurred when center-and-scale (0.862, range 0.689-0.974) was used for normalization, followed by DAGIP (0.854, range 0.591-0.976), MappingTransport (0.846, range 0.604-0.967), baseline (0.659, range 0.395-0.842), and DryClean (0.589, range 0.494-0.670). Importantly, center-and-scale normalization was significantly better than the DAGIP method.  Matthews Correlation Coefficient (MCC) was highest following normalization with DAGIP (0.473), which was significantly greater than the MCC after center-and-scale normalization (0.460).

  • Biospecimens

    FluidBlood
    FluidPlasma
  • Preservation Types

    Streck/BCT
  • Diagnoses

    NeoplasticCarcinoma
    NeoplasticLymphoma
    Normal
    Pregnant
  • Platforms

    DNANext generation sequencing
  • Pre-analytical Factors

    ClassificationPre-Analytical FactorValue(s)
    Next generation sequencing SpecificData handling
    Baseline
    Center and Scale
    DAGIP
    DryClean
    MappingTransport
    Optimal Transport
    Next generation sequencing SpecificTechnology platform
    Different library preparations used
    Different sequencers used