Comprehensive Evaluation of Genotype Imputation Tools for Ultra-low-depth Whole-Genome Sequencing Data

Ultra-low-depth sequencing (ULDS), defined as sequencing coverage below 1x, has gained traction due to its low cost, broad genome coverage, and compatibility with diverse sample types. It has already shown clinical value in applications such as non-invasive prenatal testing (NIPT)¹, cancer monitoring², and chromosomal copy number variation (CNV) detection^3,4. Beyond clinical diagnostics, the decreasing cost of sequencing and rapid advances in bioinformatics have enabled ULDS to play a growing role in population genomics and complex trait research. By combining ULDS data with population-scale haplotype reference panels, genotype imputation enables the recovery of genome-wide variant information at the individual level. As a result, ULDS has emerged as a cost-effective alternative to traditional single-nucleotide polymorphism (SNP) arrays and high-depth whole-genome sequencing (WGS)⁵, particularly in large-scale studies such as genome-wide association studies (GWAS) and population structure analyses.

Previous research has demonstrated the feasibility of conducting various genetic studies using NIPT sequencing data, including variant calling, population history reconstruction, viral infection pattern inference, and GWAS⁶.

Despite these advantages, the extremely sparse nature of ULDS data presents unique challenges. At the variant level, many sites are entirely unobserved or represented by only a single allele per individual, leading to insufficient data quality for downstream analyses. Genotype imputation is therefore essential, leveraging haplotype structure from large reference panels (e.g., 1000 Genomes⁷ or population-specific resources) to statistically infer missing or uncertain genotypes. Previous work has shown that imputation from NIPT data can achieve high accuracy and retain robust statistical power in GWAS for identifying trait-associated variants⁸. Using the STITCH⁹ algorithm, NIPT data (mean depth ~0.15x) in a cohort of 20,900 Chinese pregnant women were successfully imputed, leading to the identification of pregnancy-associated loci. The imputed genotypes showed strong concordance with high-depth WGS data in GWAS results (Pearson R² > 0.8)¹⁰.

The success of ULDS-based analyses critically depends on imputation accuracy, which is influenced by sequencing depth, reference panel quality and population match, imputation algorithm performance, sample size, and allele frequency spectrum¹¹. Among these, the choice of reference panel is a major determinant of imputation accuracy. Commonly used panels include globally representative resources such as the 1000 Genomes Project (1KGP)⁷, TOPMed¹², and the Haplotype Reference Consortium (HRC)¹³, as well as increasingly available population- or region-specific panels such as Singapore 10,000 Genomes (SG10K)¹⁴, the China Kadoorie Biobank (CKB)¹⁵. Another key factor in imputation performance is the choice of algorithm. Several tools have been developed to accommodate the unique challenges of low-depth sequencing, significantly advancing the practical use of imputation in large-scale genetic research. While imputation methods such as Beagle (v5+)¹⁶, Minimac4¹⁷, and IMPUTE5¹¹ are widely used for SNP array and medium-to-high-depth WGS data, they often perform suboptimally in ULDS settings. More recently, specialized tools have been developed to address these challenges. STITCH⁹ infers haplotypes directly from low-depth sequencing reads, making it particularly suitable for large homogeneous cohorts. QUILT2¹⁸ employs a compressed haplotype library and localized likelihood model, enabling efficient imputation with massive reference panels and offering unique applications in prenatal genomics. GLIMPSE2¹⁹, an extension of the original GLIMPSE framework, provides further improvements in both accuracy and computational efficiency.

Although these tools represent major advances, their relative performance under different experimental designs (e.g., sequencing depth, cohort size, and reference panel choice) has not been systematically evaluated, leaving researchers without clear guidance on selecting the most appropriate strategy. To bridge this gap, three widely used ULDS imputation tools -- STITCH, QUILT2, and GLIMPSE2 -- were systematically benchmarked under multiple sequencing depths and sample sizes. Their performance was evaluated using two East Asian reference panels highly relevant to Chinese populations. The findings indicate that ULDS imputation is generally reliable at sequencing depths ≥0.5x, while depths <0.1x require substantially larger cohorts to achieve acceptable accuracy. Reference panel selection should be tailored to the study context, with population-matched panels such as CKB improving imputation accuracy. Moreover, these approaches are directly applicable to ultra-low-depth data generated in large-scale population studies and NIPT. This study thus establishes a practical framework for tool selection in ULDS-based research, providing methodological guidance for future applications in population genetics and complex trait analyses.

All participants provided written informed consent prior to participation. The study involving high-depth WGS data was reviewed and approved by the BGI Institutional Review Board (BGI-IRB 23058-T2), and approval for the collection of human genetic resources was obtained from the Human Genetic Resources Administration of China ([2023] CJ0262). The study involving ULDS data from NIPT was approved by the Institutional Review Board of Wuhan Children's Hospital (2021R062) and the BGI Institutional Review Board (BGI-IRB 21088), with additional approval from the Human Genetic Resources Administration of China ([2021] CJ2002).

NOTE: This study included two types of WGS data. The first type consisted of high-depth WGS data (30xx) obtained from blood samples of 500 individuals recruited from a natural population cohort in Shenzhen. These data were used to construct a high-quality ground truth dataset and for subsequent down-sampling and accuracy evaluations. The second type comprised ULDS data derived from the NIPT of 10,000 pregnant women from the Wuhan area.

1. High-depth whole genome sequencing data

1. Collect 500 peripheral blood samples (5 mL each) from a general population cohort after informed consent. Store samples in EDTA tubes and transport them at 2-8 °C.
2. Centrifuge blood at 1,600 x g for 10 min at 4 °C to separate plasma and buffy coat. Carefully collect the buffy coat and store it at -80 °C until DNA extraction.
3. Extract genomic DNA from buffy coat using a magnetic bead-based kit following the manufacturer's instructions.
4. Quantify DNA concentration using a fluorometric assay and assess DNA integrity by agarose gel electrophoresis. Select samples with total DNA yield ≥1 µg, concentration ≥12.5 ng/µL, and fragment length >20 kb without visible degradation for library preparation.
5. Shear 80-200 ng of high-quality genomic DNA to an average size of 350-400 bp by ultrasonication.
6. Perform end-repair at 20 °C for 30 min, adapter ligation at 20 °C for 15 min, and circularization at 37 °C for 30 min to construct PCR-free libraries. Generate DNA nanoballs (DNBs) using rolling circle amplification (RCA). Sequence paired-end libraries (PE100, read length 100 bp) on a DNBSEQ platform to a target depth of ~30x (average 100 Gb per sample). Store raw sequencing reads in FASTQ format for downstream analysis.
   NOTE: Handle all human-derived samples under BSL-2 laboratory conditions. Avoid repeated freeze-thaw cycles to prevent DNA degradation. Dispose of blood-derived materials as biohazardous waste; dispose of chemical reagents following institutional hazardous waste guidelines.

2. Ultra-low depth NIPT data (~0.1x WGS)

1. Collect 10,000 maternal blood samples (5 mL each) for routine non-invasive prenatal testing (NIPT). Use EDTA tubes and transport at 2-8 °C; process plasma within 8 h of collection.
2. For stabilized circulating DNA tubes (K-tubes or G-tubes), transport at 6-35 °C using temperature-controlled carriers and process within 96 h following the manufacturer's standard operating procedures.
3. Centrifuge blood at 1,600 x g for 10 min at 4 °C to separate plasma. Carefully collect the upper plasma layer without disturbing the buffy coat or cell pellet using a pipette and transfer it to a new tube. Centrifuge the recovered plasma again at 16,000 x g for 10 min at 4 °C to remove any residual cells or debris. Carefully transfer the clarified supernatant (cell-free plasma) to a fresh tube for DNA extraction.
4. Extract circulating cell-free DNA (cfDNA) from plasma using a nucleic acid extraction kit. Perform end-repair at 20 °C for 30 min, adapter ligation at 20 °C for 15 min, and PCR amplification (12 cycles, 98 °C denaturation 10 s, 60 °C annealing 30 s, 72 °C extension 30 s).
5. Purify PCR products and circularize libraries at 37 °C for 30 min. Generate DNBs through RCA. Sequence single-end libraries (SE35, read length 35 bp) on a BGISEQ-500 platform. Store raw sequencing data in FASTQ format.
   NOTE: Handle plasma samples as potentially infectious material under BSL-2 conditions. Minimize freeze-thaw cycles to reduce cfDNA degradation. Dispose of plasma waste and plastic consumables as biohazardous material.

3. Data preprocessing pipeline

1. To systematically evaluate the performance of genotype imputation tools under varying sequencing depths, carry out a standardized preprocessing workflow on both the original high-depth WGS data (30x) and the ultra-low-depth NIPT data (<0.1x), including simulated downsampling, quality control, read alignment, duplicate removal, and base quality score recalibration (BQSR).
   NOTE: The steps from this point to imputation accuracy evaluation constitute the Main Protocol (Figure 1) of this study. The specific code can be found in Supplementary File 1.
2. Downsampling
   1. Generate a series of downsampled datasets from the original 30x high-depth sequencing samples. Employ two strategies to realistically mimic the sequencing characteristics of NIPT data as described below.
   2. Random subsampling: Use seqtk v1.5 (https://github.com/lh3/seqtk) with a fixed random seed of 100 to create four levels of low-depth data (0.05x, 0.1x, 0.5x, and 1.0x).
   3. NIPT-like read structure simulation: Retain only the first read (R1) of each paired-end read and truncate all retained reads to 35 bp with seqtk trimfq -L 35, consistent with the typical single-end, short-read nature of ultra-low-depth NIPT sequencing.
3. Quality control
   1. Process all raw FASTQ files with fastp v0.23.4²⁰. Use the following parameters: --qualified_quality_phred=5 (base quality threshold), --unqualified_percent_limit=50 (maximum percentage of low-quality bases allowed), --n_base_limit=10 (maximum N bases per read), and custom adapter removal with --adapter_sequence=AAGTCGGAGGCCAAGCGGTCTTAG
      GAAGACAA (R1) and --adapter_sequence_r2=AAGTCGGATCGTAGCC
      ATGTCGTTCTGTGAG
      CCAAGGAGTTG (R2).
   2. Disable Poly-G tail trimming (--disable_trim_poly_g), and generate reports in both JSON and HTML formats for each sample.
4. Alignment and duplicate removal
   1. Align high-quality reads to the human reference genome GRCh38 (hg38)²¹ using BWA v0.7.16a-r1181²².
   2. Perform the alignment with the aln algorithm (-e 10 -t 4 -i 5 -q 0), followed by samse for single-end alignment with read group information.
   3. Convert the resulting SAM files to BAM, sorted (samtools sort -@ 8), and remove duplicates using SAMtools v1.3²³ (samtools rmdup). Index all BAM files.
5. Base quality score recalibration (BQSR)
   1. Perform BQSR using GATK v4.0.4.0²⁴. Train the recalibration model on three high-confidence variant datasets: dbSNP build 146²⁵, Mills and 1000G gold standard indels²¹, and the GATK resource bundle²⁴ known indels file for GRCh38. The bundle files used refer to the GATK official example (https://github.com/gatk-workflows/gatk4-data-processing/blob/master/processing-for-variant-discovery-gatk4.hg38.wgs.inputs.json). In total, download three files and their corresponding index files.
   2. Run BaseRecalibrator followed by ApplyBQSR to generate recalibrated BAM files. Index all BAMs using SAMtools v1.3.
      NOTE: All simulated datasets underwent identical preprocessing steps-downsampling, quality control, alignment, duplicate removal, and BQSR-to ensure consistency and comparability in subsequent imputation performance evaluations.

4. Genotype imputation

1. Data preparation
   1. Imputation dataset setup: Construct multiple evaluation datasets to systematically benchmark genotype imputation tools across different depths and sample sizes as described below. A total of nine combinations are formed based on the different sample sizes and sequencing depths mentioned above. The input files consist of the sequencing data BAM file lists (bamlist.txt) for the above-mentioned nine subsets after quality control, stored in corresponding bamlist.txt files. Other input files include the human reference genome (GRCh3821) and the genetic map from the 1000 Genomes Project⁷.
      1. Downsampled high-depth WGS data: Randomly select two subsets (200 and 500 samples) from 500 individuals sequenced at 30x depth. Downsample each subset to four depths (1x, 0.5x, 0.1x, and 0.05x) to generate eight experimental conditions.
      2. NIPT-based ULDS dataset: Combine 10,000 ultra-low-depth NIPT samples (average depth is 0.102x, Figure 2) with 50 high-depth samples downsampled to 0.1x.
   2. Analysis region specification: Restrict all analyses to the chromosome 1 region chr1:150,500,000-160,500,000 (10 Mb), with a 500 kb buffer for imputation, to ensure direct comparability across tools.
   3. Reference panel selection: Use two reference panels (Table 1): the CKB panel, constructed from Chinese population data, and the 1KGP-EAS panel, derived from the East Asian subset of the 1000 Genomes Project.
      NOTE: The CKB reference panel¹⁵ was constructed using high-depth (~15x) whole-genome sequencing data from 9,964 Chinese adults in the China Kadoorie Biobank, a large prospective cohort study. These samples are derived from a natural population with minimal phenotypic bias, homogeneous Han Chinese ancestry, and consistent population structure, making them particularly well-suited for genotype imputation in Chinese cohorts. Yu et al. ¹⁵ demonstrated that, in a real-phenotype GWAS for height, imputation using the CKB panel tripled the number of SNPs detected and doubled the number of genome-wide significant variants. The 1000 Genomes Project (1KGP)^7,26, the most widely used genomic reference, includes 585 individuals in its East Asian (EAS) Phase 3 subset. This subset covers five East Asian populations, which have a sequencing depth of approximately 30x, including Han Chinese in Beijing (CHB), Southern Han Chinese (CHS), Chinese Dai in Xishuangbanna (CDX), Kinh in Ho Chi Minh City, Vietnam (KHV), and Japanese in Tokyo (JPT).
2. Imputation tools
   1. Evaluate three imputation algorithms, chosen for their distinct modeling strategies and applicability to ultra-low-depth sequencing (ULDS) data.
      1. STITCH: STITCH (v1.6.6) is a reference-free haplotype-based imputation algorithm that can optionally incorporate external reference haplotypes. Include BAM lists, human reference genome (GRCh38) as input files. Prepare reference panel files (hap/legend/pos) when performing reference-based imputation. Include the following key parameters: method=diploid, buffer=500 kb, K=10 ancestral haplotypes, and nGen=4x sample size/K (as recommended in STITCH documentation). Generate output files containing per-SNP genotype dosages for all individuals.
         NOTE: As per the STITCH official documentation, K is the number of ancestral haplotypes in the model. A larger K improves imputation accuracy for larger samples and higher coverages, but it also increases computation time, and accuracy may decrease with lower coverage.
      2. QUILT2: QUILT2 applies a Bayesian reference-guided approach optimized for ULDS data. Perform reference panel preparation using the provided prepare_reference script, specifying the genetic map and region coordinates. Run imputation diploid mode with the same buffer size (500 kb) and nGen setting as STITCH to ensure comparability.
      3. GLIMPSE2: GLIMPSE2 is an HMM-based reference-driven imputation tool, designed for large-scale and very low-depth sequencing datasets. Execute imputation with GLIMPSE2_phase_static, specifying input BAM list, human reference VCF panel, genetic map, input region = chr1:150,000,000-161,000,000, output region = chr1:150,500,000-160,500,000 (to maintain a 500 kb buffer). The BAM list file entered here needs to contain two columns: one is the BAM path, and the second column is the sample name. If the second column is not entered, each BAM file name will be used as the sample name in the output VCF file.

5. Imputation accuracy evaluation

1. Truth set definition
   1. Select 50 individuals sequenced at 30x depth as the ground-truth dataset. Include these samples in the experimental downsampling conditions to ensure comparability.
   2. Perform variant calling for the truth set using the following pipeline: SOAPnuke²⁷ for QC, BWA for alignment, Picard for duplicate marking, GATK v4.0.4.0 BQSR and HaplotypeCaller for variant calling, DPGT (https://github.com/BGI-flexlab/DPGT) for joint calling, and BCFtools v1.11²³ for variant quality filtering.
   3. Retain only high-confidence PASS variants to generate a benchmark VCF file. Restrict the evaluation to chr1:150,500,000-160,500,000, consistent with the imputed datasets.
2. Data harmonization and filtering
   1. Process imputed VCF files from all tools using PLINK2.0²⁸. Extract dosage data and convert to pgen format.
   2. Apply SNP-level quality control with the following filters: Minor allele frequency (MAF) ≥ 0.05 (--maf 0.05), Hardy-Weinberg equilibrium (HWE) p-value ≥ 1e-6 (--hwe 1e-6), Biallelic SNPs only (--max-alleles 2).
   3. Export variants passing QC to traw format for downstream comparison.
3. Accuracy metrics
   1. Compare the imputed dosages with the ground-truth dosages for each SNP. Compute Pearson correlation coefficients (R) on a SNP-by-SNP basis, retain only sites common to both datasets, and calculate the mean of squared correlation coefficients (R²) across all evaluated SNPs to quantify overall imputation accuracy for each condition.
   2. Use this accuracy metric to capture the concordance of genotype dosage estimates between imputed and true genotypes.

Impact of sample size on imputation accuracy
Increasing the sample size from N = 200 to N = 500 improved the imputation accuracy of STITCH, particularly under low-coverage conditions. For example, with the CKB reference panel at 1x coverage, STITCH achieved an R2> of 0.916 (N=500) compared to 0.882 (N=200), representing a 3.4% increase (Figure 3; Supplementary File 2). Similarly, at 0.5x coverage, its accuracy rose from 0.800 to 0.868 (ΔR2> = 8.5%). In contrast, QUILT2 and GLIMPSE2 showed minimal sensitivity to sample size variations, with R2> fluctuations of less than 0.5% across all tested conditions. For instance, QUILT2 maintained stable performance under the CKB panel at 1x coverage (R2> = 0.970 for N = 200 vs. 0.971 for N = 500). This indicates that STITCH benefits more from larger sample sizes, consistent with its hidden Markov model (HMM)-based haplotype inference framework.

Reference panel compatibility
The choice of reference panel affected the imputation accuracy of QUILT2 and GLIMPSE2, while having little impact on STITCH (Figure 3; Supplementary File 2). When using the Chinese population-specific CKB panel, both QUILT2 and GLIMPSE2 achieved higher accuracy than with the general-purpose EAS panel. For example, at 1x coverage, the R2> of GLIMPSE2 increased from 0.956 with EAS to 0.974 with CKB (ΔR2> = 1.8%), while QUILT2 rose from 0.950 to 0.971 (ΔR2> = 2.1%). In contrast, STITCH showed almost no difference between the two panels (R2> = 0.916 versus 0.915), consistent with its independence from external reference panels. These results highlight the practical advantage of population-specific reference panels in precision medicine research.

Sequencing depth and accuracy trade-offs
The imputation accuracy of all tools increased with higher sequencing depth, but differences between methods became more evident under ultra-low coverage conditions (≤ 0.1x) (Figure 3; Supplementary File 2). Using the CKB panel as an example, GLIMPSE2 achieved an R2> of 0.676 at 0.05x coverage, which increased to 0.974 at 1x. Similarly, QUILT2 rose from 0.666 to 0.971 over the same range. In contrast, STITCH performed poorly at 0.05x coverage, with an R2> of 0.255, indicating reduced robustness under extremely low coverage. At 0.1x coverage, QUILT2 and GLIMPSE2 reached R2> values above 0.817 with the CKB panel, whereas STITCH reached only 0.532 (N = 500).

SNP count considerations and impact on tool selection
STITCH produced fewer total SNPs than QUILT2 and GLIMPSE2, but after quality control, the differences in effective SNP numbers among the tools were reduced. The total number of imputed SNPs varied across tools. For example, at a sample size of N = 500 and coverage of 1x, STITCH generated 21,829 SNPs with the CKB panel and 24,546 with the EAS panel, compared with 398,026 (CKB) and 205,711 (EAS) for QUILT2, and 467,754 (CKB) and 239,495 (EAS) for GLIMPSE2 under the same conditions. After applying standard quality control procedures, MAF and HWE filtering, the number of SNPs retained for downstream analysis became comparable across all tools (Figure 4; Supplementary File 2). This suggests that although the raw SNP counts differ across methods, the number of usable variants after QC is similar, and therefore total imputed SNP count should not be considered a decisive factor when selecting an imputation tool.

Computational resource consumption and tool selection recommendations
In this study, core-hours -- defined as the product of the number of CPU cores and runtime -- were used as a key metric to quantify the computational burden of genotype imputation tools. Under the same sample size and sequencing depth, imputation using the CKB reference panel consistently required more core-hours than using the EAS panel, suggesting that larger or more complex reference panels increase computational demand. Additionally, core-hour usage increased with both sample size and sequencing depth, highlighting the need for careful resource planning in large-scale studies (Figure 5; Supplementary File 3). Among the three tools evaluated, QUILT2 consumed substantially more core-hours than both STITCH and GLIMPSE2, especially at the largest sample size (N = 10,050), where QUILT2 required 361-496 core-hours per imputation run, compared to approximately 20-22 for STITCH and 87-145 for GLIMPSE2. Based on these observations, GLIMPSE2 is recommended for small- to medium-sized datasets, particularly when imputation accuracy is a primary concern, but computational resources are constrained. Conversely, STITCH should be considered the tool of choice in studies that lack high-quality reference panels or aim to handle very large cohorts, as it offers a favorable trade-off between accuracy and efficiency. Researchers are therefore advised to balance accuracy, reference panel availability, and computational costs when selecting imputation tools for ultra-low-depth WGS studies.

Figure 1: Main steps of the comprehensive evaluation of genotype imputation tools. This flowchart illustrates the three main steps of genotype imputation accuracy evaluation: Data Preprocessing Pipeline, Genotype Imputation Workflow, and Imputation Accuracy Evaluation. Please click here to view a larger version of this figure.

Figure 2: Sequencing depth distribution of 10,000 ultra-low-depth NIPT samples. Histogram showing the coverage depth distribution of 10,000 non-invasive prenatal testing (NIPT) samples sequenced at ultra-low depth. The red dashed line indicates the mean coverage depth (0.102x), and the green dashed line represents the median coverage depth (0.104x). The x-axis ranges from 0.05x to 0.35x, with the majority of samples concentrated around 0.10x. Please click here to view a larger version of this figure.

Figure 3: Evaluation of imputation accuracy for three genotype imputation tools under varying conditions. (A) Accuracy across varying coverage depths (0.05x, 0.1x, 0.5x, 1x) using two reference panels (CKB versus EAS), with sample size fixed at 500. (B) Accuracy across varying coverage depths under two sample sizes (200 versus 500), using the EAS reference panel. (C) Accuracy on 0.1x coverage depth data under different sample sizes (200, 500, and 10,050). Please click here to view a larger version of this figure.

Figure 4: Comparison of SNP numbers from different genotype imputation tools. A sample size of n = 500 and coverage depth of 1x: (A) The total number of SNPs predicted and the number of SNPs remaining after quality control (MAF > 0.05 and HWE filtering) by three genotype imputation tools (STITCH, QUILT2, and GLIMPSE2) in the EAS and CKB populations. (B) The number of SNPs after quality control and the number of effective SNPs used for calculating imputation accuracy. Please click here to view a larger version of this figure.

Figure 5: Comparison of core-hour consumption across tools under varying coverage depths, sample sizes, and reference panels. (A) Core-hour consumption changes with increasing coverage depth under a fixed sample size (N = 500), comparing three tools (STITCH, GLIMPSE2, and QUILT2) using the EAS and CKB reference panels. (B) Compares core-hour usage across different sample sizes (200 vs. 500) as coverage depth increases, under the EAS reference panel. (C) The impact of sample size (200, 500, 10,050) on core-hour consumption at a fixed coverage depth of 0.1x, under both EAS and CKB reference panels. Please click here to view a larger version of this figure.

Table 1: Information on reference panels. The table provides key details about the reference panels used for genotype imputation, including the name of the reference panel, sample size, sequencing depth, and ancestries.

Table 2: Recommended application scenarios for STITCH, QUILT2, and GLIMPSE2. Recommended application scenarios for STITCH, QUILT2, and GLIMPSE2. Recommendations are derived from comparative evaluations under varying sequencing depths, sample sizes, and reference panel conditions, aiming to guide tool selection in ultra-low-depth WGS studies.

Supplementary File 1: Original code for analyses. Please click here to download this File.

Supplementary File 2: Imputation accuracy and SNPs number. Please click here to download this File.

Supplementary File 3: Core-hour consumption of imputation. Please click here to download this File.

This study systematically assessed the performance of three widely used genotype imputation tools for ULDS, with high-depth WGS serving as the gold standard. A key methodological strength lies in the adoption of a unified preprocessing pipeline -- encompassing alignment, quality control, and base quality score recalibration -- that minimizes batch effects and ensures comparability across tools and conditions. By downsampling deeply sequenced samples, ultra-low-depth data were simulated under controlled settings, thereby providing an objective framework for benchmarking. Restricting analyses to a defined 10 Mb genomic interval on chromosome 1 ensured computational feasibility while retaining sufficient variant density. Comparative assessment of STITCH, QUILT2, and GLIMPSE2 reveals distinct strengths and weaknesses across imputation strategies.

Sequencing depth and sample size exerted predictable effects on imputation accuracy. Accuracy declined sharply below ~0.1x depth, reflecting insufficient information to reliably recover haplotypes from sparse read data. Larger sample sizes mitigated this limitation by enabling improved inference of population haplotype structure, consistent with theoretical expectations that more samples improve imputation performance²⁹. The results indicate that depths of at least 0.5x maintain acceptable accuracy for common variants, even in modest sample sizes, whereas depths below 0.1x require substantially larger cohorts to achieve comparable performance. These findings have practical implications for study design: sample size can partly compensate for sequencing depth, but only within certain limits.

Reference panel selection emerged as a decisive factor influencing imputation performance. 1KGP-EAS panel, widely used in global studies, provided robust coverage of common variants. By contrast, the CKB panel, derived from large-scale local sequencing, improved accuracy for population-specific variants. This underscores that population-specific reference panels, which capture a larger repertoire of haplotypes, substantially enhance imputation quality. This observation aligns with previous findings that leveraging local haplotype diversity increases the chance that any given allele is tagged by a representative haplotype^7,15. Future efforts to expand and diversify Chinese reference panels will be critical for further improving imputation accuracy. Comprehensive evaluation demonstrates distinct performance profiles for STITCH, QUILT2, and GLIMPSE2 under varying sequencing depths, sample sizes, and reference panel configurations. GLIMPSE2 achieved the highest imputation accuracy (R² = 0.974) when using the CKB reference panel at 1x coverage, closely followed by QUILT2 (R² = 0.971). Both tools exhibited greater robustness to changes in sample size and sequencing depth compared with STITCH. Although STITCH yielded lower overall accuracy, its minimal computational requirements and ability to function without an external reference panel make it particularly advantageous for large-scale cohorts (N ≥ 10,000) or populations where no well-matched reference panel is available.

Overall, these findings suggest that tool selection should be tailored to study design, dataset characteristics, reference panel availability, and computational constraints. A comparative summary of recommended applications is provided in Table 2. Notably, QUILT2 offers unique functionality for maternal-fetal genome separation, making it particularly applicable to ultra-low-depth NIPT datasets. GLIMPSE2 balances accuracy and computational efficiency, providing high-quality results with manageable resource usage.

Several technical issues were identified that can influence imputation outcomes. For GLIMPSE2, the BAM list file must be correctly formatted with two columns (BAM path and sample name); omission of the second column may lead to inconsistent sample identifiers in the output VCF¹¹. For STITCH, the parameter K (number of ancestral haplotypes) requires careful adjustment: larger values improve accuracy at higher depth and sample size, but can reduce performance under ultra-low-depth conditions⁹. In addition, STITCH occasionally fails during multi-threaded runs due to known instability in its parallelization library; switching between single- and multi-thread execution generally resolves this issue³⁰ (GitHub issue #86: https://github.com/rwdavies/STITCH/issues/86). Addressing these points was essential to ensure reproducibility and stable imputation results.

In this study, a representative genomic interval on chromosome 1 was selected to evaluate overall imputation performance, with a focus on common variants. The aim was to benchmark tool behavior under typical genomic contexts rather than highly complex regions such as the HLA locus, which involves structural variation and may not be equally well addressed by current imputation algorithms. This design provides a reliable overview of population-level imputation accuracy; however, it does not capture the challenges associated with rare variants or structurally complex loci, which remain important limitations. Future studies may focus on challenging genomic regions such as the HLA locus to further test tool performance. Moreover, while the CKB panel substantially improved accuracy for Chinese populations, continued expansion of population-specific reference panels will be essential to improve the resolution of low-frequency variants. Finally, emerging approaches such as deep learning-based imputation methods³¹ hold promise for further enhancing accuracy and robustness under ultra-low-depth conditions.