Benchmarking Q40 sequencing for sensitive and efficient detection of rare genomic variants Authors: Shumeng Duan, Yaqing Liu, Xiaorou Guo, Zhiyin An, Ruiwen Ma, Qiaochu Chen, Yanming Xie, Qingwang Chen, Ying Yu, Lianhua Dong, Leming Shi, Yuanting Zheng Published: 2026 ABSTRACT BackgroundPhred quality score (Q score) is critical for sequencing accuracy, yet the impact of Q40-achieving sequencing technologies (99.99% accuracy) on detecting subtle biological variations remains unvalidated.ResultsUsing a comprehensive set of well-established DNA/RNA reference materials (Quartet, NIST-RM8398, SEQC2-HCC1395/BL, MAQC, and ERCC), we benchmarked Q40 sequencing (Element AVITI) against the conventional Q30 standard (Illumina NovaSeq 6000). Q40 reduced required sequencing depth by 33.3% while maintaining accuracy for germline variants (20 × vs. 30 ×) and somatic single-nucleotide variant/insertion-deletion (SNV/InDel) (60 × vs. 90 ×). Crucially, Q40 enhanced sensitivity for low-frequency somatic mutations (variant allele frequency, VAF ≤ 0.2) by 33.3% and sixfold higher copy number variation (CNV) detection reproducibility (60.3% vs. 10.4%) with Q40 at 30 × depth, directly reducing per-sample volumes by 33.3–60% and theoretically reducing sequencing costs by 2.2–31.7%. In addition, Q40 improved the discriminatory resolution between biological samples with 13.1% signal-to-noise ratio (SNR) enhancement.ConclusionsTaken together, our findings establish the value of Q40 sequencing as a sensitive and cost-effective method for low-frequency variant detection. While this positions it as a promising tool for precision oncology, its performance in real-world clinical applications remains to be evaluated in future studies.Supplementary InformationThe online version contains supplementary material available at 10.1186/s13059-026-04146-3. Background Sequencing accuracy, quantified by the Phred quality scores(Q score, calculated by the formula \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$Q=-10\mathrm{log}{}_{10} P$$\end{document}Q=-10log10P, where PP is the probability of a sequencing error), underpins reliable genomic and transcriptomic analyses [1–3]. While Q30 (Phred quality score of 30, 99.9% accuracy) has long served as the benchmark [3, 4], its limitation in detecting ultra-low frequency mutations (variant allele frequency,VAF < 0.001) constrained applications in liquid biopsy and tumor heterogeneity studies [5, 6]. Recent advances have enabled Q40 (Phred quality score of 40, 99.99% accuracy), offering transformative potential for precision genomics, particularly in capturing low-frequency biological events. To fully leverage the potential of Q40, platforms such as Element AVITI [7], Illumina NovaSeq XLEAP-SBS [8], PacBio Onso, Ultima ppmSeq, and MGI DNBSEQ™ have introduced innovative chemistries. Among these, Element AVITI has achieved Q40 for over 85% of bases and has seen widespread adoption. However, thorough validation of Q40's performance in complex biological contexts is still limited. Previous studies [9] have primarily focused on single reference materials (e.g., Genome in a Bottle (GIAB) [10] or External RNA Controls Consortium (ERCC) [11]) or narrow gene panels, overlooking Q40's ability to detect subtle biological variations crucial for clinical genomics [5, 6]. Here, we address this gap using a combinatorial approach with well-characterized DNA/RNA reference materials (Quartet DNA/RNA [12–14], National Institute of Standards and Technology (NIST)-RM8398 [15], Sequencing Quality Control Phase 2 (SEQC2)-HCC1395/BL [16, 17], MicroArray Quality Control (MAQC) [18, 19], and ERCC [20]). We generated Q30 and Q40 datasets spanning germline (single nucleotide variant/insertion and deletion (SNV/InDel)), somatic mutations (low-frequency somatic variants (VAF ≤ 0.2)), somatic copy number variation (CNV), and gene-level quantification standards, to mimic real-world samples with diverse characteristics. By applying multiple quality metrics, we evaluated Q40's accuracy in detecting low-frequency mutations and gene-expression dynamics. Our work establishes a framework for deploying Q40 sequencing to resolve biologically nuanced variations in clinical and research settings. Results Overall study design To rigorously evaluate Q40 sequencing's accuracy in detecting genomic events under real-world conditions (Fig. 1a), we designed a combinatorial framework leveraging a comprehensive set of universal DNA/RNA reference materials with truth sets (Fig. 1b, c). For DNA-level variant assessment, we integrated: 1) NIST-RM8398 reference material [15] for germline SNV/InDel validation; 2) Quartet DNA reference materials (monozygotic twins D5/D6 and parents F7/M8) [12–14] to benchmark germline variants by Mendelian consistency; and 3) SEQC2-HCC1395/BL tumor-normal reference materials [16, 17] for low-frequency somatic variants (VAF ≤ 0.2) and CNV detection. For transcriptome analysis, RNA reference materials included: 1) Quartet RNA samples (D5/D6/F7/M8) and engineered mixtures T1 (M8:D6 = 3:1) and T2 (M8:D6 = 1:3) [21] for known titer differences, calculated as (T1—D6)/(M8—D6) = 0.75 and (T2—D6)/(M8—D6) = 0.25; 2) MAQC samples A/B [18, 19] with TaqMan-quantified expression baselines; and 3) ERCC spike-ins [20] for sensitivity thresholds via concentration gradients.Fig. 1Overview of study design. a Schematic of Q40 vs. Q30 performance evaluation in variant detection and gene quantification. b DNA reference materials: Quartet DNA, NIST-RM8398, and SEQC2-HCC1395/BL. c RNA reference materials: Quartet RNA, MAQC (Universal Human Reference RNA and Human Brain Reference RNA), and ERCC (spiked into Quartet M8 and D6). Mixed samples: T1 (M8:D6 = 3:1) and T2 (M8:D6 = 1:3). d Whole-exome sequencing (WES) and RNA sequencing (RNA-seq) datasets generated from DNA and RNA reference materials using Illumina NovaSeq 6000 (Q30) and Element AVITI (Q40) platforms. WES: 42 replicates (7 samples × 3 technical replicates × 2 platforms). RNA-seq: 48 replicates (8 samples × 3 technical replicates × 2 platforms) Overview of study design. a Schematic of Q40 vs. Q30 performance evaluation in variant detection and gene quantification. b DNA reference materials: Quartet DNA, NIST-RM8398, and SEQC2-HCC1395/BL. c RNA reference materials: Quartet RNA, MAQC (Universal Human Reference RNA and Human Brain Reference RNA), and ERCC (spiked into Quartet M8 and D6). Mixed samples: T1 (M8:D6 = 3:1) and T2 (M8:D6 = 1:3). d Whole-exome sequencing (WES) and RNA sequencing (RNA-seq) datasets generated from DNA and RNA reference materials using Illumina NovaSeq 6000 (Q30) and Element AVITI (Q40) platforms. WES: 42 replicates (7 samples × 3 technical replicates × 2 platforms). RNA-seq: 48 replicates (8 samples × 3 technical replicates × 2 platforms) The Illumina NovaSeq 6000 served as the Q30 benchmark platform (Fig. 1d), based on its: 1) dominant market presence; 2) well-characterized error profile (Q30: 1 error per 1,000 bases); and 3) FDA recognition as a predicate device for in vitro diagnostics (IVD) applications [22]. For Q40 assessment, we used the Element AVITI system [7], chosen for its: 1) unique single-chemistry design eliminating PCR-induced errors; and 2) native Q40 accuracy (Phred score ≥ 40; error rate ≤ 0.01%). DNA libraries were processed by whole-exome sequencing (WES), and RNA libraries by RNA sequencing (RNA-seq). All libraries underwent triplicate sequencing, generating 42 DNA and 48 RNA replicates. Performance was quantified through different metrics (Tables 1, 2, and 3), spanning: 1) variant-calling accuracy (germline/somatic SNV/InDel and CNV); and 2) gene-expression precision (relative quantification and spike-in recovery). Table 1Metrics for assessing data qualityMetricsDetailsReference materialsPhred Quality ScoreProbability of incorrect base calling, defined as defined as\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$Q=-10\mathrm{log}{}_{10} P$$\end{document}Q=-10log10P(PP= sequencing error probability)RM8398, Quartet DNA/RNA, HCC1395/BL, MAQCDuplication Rates (%)Proportion of duplicate reads from PCR amplification (duplicate reads/total mapped reads)RM8398, Quartet DNA/RNA, HCC1395/BL, MAQCMapping QualityReliability/uniqueness of read alignment (MAPQ score: 0–60; higher = more reliable)RM8398, Quartet DNA/RNA, HCC1395/BLAligned Reads (%)Percentage of reads successfully aligned to reference genome (excludes adapters, low-quality/contaminant reads)RM8398, Quartet DNA/RNA, HCC1395/BL, MAQC5'/3' BiasThe efficiency in capturing, amplifying, or detecting sequences at the 5' or 3' ends of RNA molecules during sequencing or analysisQuartet RNA, MAQCTable 2Metrics for assessing variant detectionTypeMetricsDetailsReference materialsReference-independentReproducibilityConsistency of variant detection across technical replicates (Jaccard index of shared variants: matching CHROM, POS, REF, ALT)RM8398, Quartet DNA, HCC1395/BLReference-dependentF1 ScoreHarmonic mean of precision [TP/(TP + FP)] and recall [TP/(TP + FN)] for variant detection accuracy in high-confidence genomic regionsRM8398, Quartet DNA, HCC1395/BLMendelian Consistency Rate (MCR)Proportion of variants conforming to Mendelian inheritance in Quartet DNA (identical twins + parents)Quartet DNATable 3Metrics for assessing gene quantificationTypeMetricsDetailsReference materialsReference-independentSignal-to-Noise Ratio (SNR)Calculated via PCA: Signal = RMS of inter-group distances; Noise = RMS of intra-replicate distances.(RMS means Root Mean Square)Quartet RNA, MAQCCoefficient of Variation (CV)CV of normalized gene expression values (FPKM), reflecting technical variation (lower = more stable)Quartet RNA, MAQCGene CountGenes detectable in ≥ 2/3 technical replicates (sample pair: e.g., D5/D6; FPKM > 0.1)Quartet RNA, MAQCReference-dependentPearson Correlation Coefficient (PCC)Linear correlation between experimental log₂FC and ground-truth values (range: −1 to 1; higher = more consistent)Quartet RNA, MAQCMatthews Correlation Coefficient (MCC)Balances sensitivity/specificity for DEG detection (range: −1 to 1; 1 = perfect consistency, 0 = random)Quartet RNA, MAQCTiter DifferenceTheoretical values (0.75 = (T1-D6)/(M8-D6); 0.25 = (T2-D6)/(M8-D6)) for validating relative expression in mixed Quartet RNA samplesQuartet RNAArea Under the Curve (AUC)DEG detection accuracy via ROC curves (true-positive vs. false-positive rate; range: 0.5–1.0; higher = better performance)ERCCLimit of Detection Ratio (LODR)Minimum expression signal for reliable DEG detection (modeled via ERCC spike-in concentration gradients and differential expression P-values)ERCC Table 1Metrics for assessing data qualityMetricsDetailsReference materialsPhred Quality ScoreProbability of incorrect base calling, defined as defined as\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$Q=-10\mathrm{log}{}_{10} P$$\end{document}Q=-10log10P(PP= sequencing error probability)RM8398, Quartet DNA/RNA, HCC1395/BL, MAQCDuplication Rates (%)Proportion of duplicate reads from PCR amplification (duplicate reads/total mapped reads)RM8398, Quartet DNA/RNA, HCC1395/BL, MAQCMapping QualityReliability/uniqueness of read alignment (MAPQ score: 0–60; higher = more reliable)RM8398, Quartet DNA/RNA, HCC1395/BLAligned Reads (%)Percentage of reads successfully aligned to reference genome (excludes adapters, low-quality/contaminant reads)RM8398, Quartet DNA/RNA, HCC1395/BL, MAQC5'/3' BiasThe efficiency in capturing, amplifying, or detecting sequences at the 5' or 3' ends of RNA molecules during sequencing or analysisQuartet RNA, MAQC Metrics for assessing data quality Table 2Metrics for assessing variant detectionTypeMetricsDetailsReference materialsReference-independentReproducibilityConsistency of variant detection across technical replicates (Jaccard index of shared variants: matching CHROM, POS, REF, ALT)RM8398, Quartet DNA, HCC1395/BLReference-dependentF1 ScoreHarmonic mean of precision [TP/(TP + FP)] and recall [TP/(TP + FN)] for variant detection accuracy in high-confidence genomic regionsRM8398, Quartet DNA, HCC1395/BLMendelian Consistency Rate (MCR)Proportion of variants conforming to Mendelian inheritance in Quartet DNA (identical twins + parents)Quartet DNA Metrics for assessing variant detection Table 3Metrics for assessing gene quantificationTypeMetricsDetailsReference materialsReference-independentSignal-to-Noise Ratio (SNR)Calculated via PCA: Signal = RMS of inter-group distances; Noise = RMS of intra-replicate distances.(RMS means Root Mean Square)Quartet RNA, MAQCCoefficient of Variation (CV)CV of normalized gene expression values (FPKM), reflecting technical variation (lower = more stable)Quartet RNA, MAQCGene CountGenes detectable in ≥ 2/3 technical replicates (sample pair: e.g., D5/D6; FPKM > 0.1)Quartet RNA, MAQCReference-dependentPearson Correlation Coefficient (PCC)Linear correlation between experimental log₂FC and ground-truth values (range: −1 to 1; higher = more consistent)Quartet RNA, MAQCMatthews Correlation Coefficient (MCC)Balances sensitivity/specificity for DEG detection (range: −1 to 1; 1 = perfect consistency, 0 = random)Quartet RNA, MAQCTiter DifferenceTheoretical values (0.75 = (T1-D6)/(M8-D6); 0.25 = (T2-D6)/(M8-D6)) for validating relative expression in mixed Quartet RNA samplesQuartet RNAArea Under the Curve (AUC)DEG detection accuracy via ROC curves (true-positive vs. false-positive rate; range: 0.5–1.0; higher = better performance)ERCCLimit of Detection Ratio (LODR)Minimum expression signal for reliable DEG detection (modeled via ERCC spike-in concentration gradients and differential expression P-values)ERCC Metrics for assessing gene quantification Higher Q score enhanced sequencing accuracy and data quality in both WES and RNA-seq Raw FASTQ datasets were downsampled to 10 gigabases (Gb) to eliminate the influence of sequencing depth on the comparison (Additional file 1: Table S1). Optimization percentage (defined as:\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\frac{|\textrm{Q}40-\textrm{Q}30|}{\mathrm{Q}30} \times 100\%$$\end{document}|Q40-Q30|Q30×100%) demonstrated improvement (Methods). As shown in Fig. 2a, the Q40 sequencer demonstrated significantly higher mean base quality than the Q30 sequencer in WES analyses (41.10 ± 0.14 vs. 35.83 ± 0.07, p < 0.001, Cohen's d = 29.44; Paired t-test). Figure 2b illustrates that Q40 reduced duplication rates by 55.3% in WES (15.30 ± 0.64% vs. 34.19 ± 0.72%, p < 0.001, Cohen's d = −21.75, Wilcoxon signed-rank test). Moreover, Q40 exhibited the increase in WES alignment accuracy (99.95 ± 0.01% vs. 99.89 ± 0.01%, p < 0.001, Cohen's d = 4, Paired t-test; Fig. 2c). The enhanced base quality directly improved mapping confidence, with Q40 showing significantly higher mean mapping quality (MAPQ) than Q30 (58.49 ± 0.01 vs. 58.41 ± 0.01, p < 0.001, Cohen's d = 30.59, Paired t-test; Fig. 2d).Fig. 2Comprehensive comparison of platforms with different sequencing quality scores (Q30 vs. Q40). a-d WES quality control metrics for datasets from all DNA reference materials: a Phred quality score, b duplication rates, c read alignment rates, and d mapping quality. e–h RNA-seq Quality control metrics for datasets from all RNA reference materials: e Phred quality score, f duplication rates, g read alignment rates, and h 5'/3' bias. Boxplots show the interquartile range, median (line within the box), and whiskers (min–max excluding outliers). Data points represent individual FASTQ files for raw data metrics (Phred score, duplication; n = 42 for WES, 48 for RNA-seq) and unique biological samples for post-alignment metrics (alignment rate, mapping quality, bias; n = 21 for WES, 24 for RNA-seq). Significance was determined using paired t-test or Wilcoxon signed-rank test (*p < 0.05, **p < 0.01, ***p < 0.001, ****p < 0.0001)., as appropriate, following normality assessment (Shapiro–Wilk test). For significant results, effect size (Cohen's d) and 95% confidence interval (CI) are provided. See Methods and Additional file 2: Table S2 for details Comprehensive comparison of platforms with different sequencing quality scores (Q30 vs. Q40). a-d WES quality control metrics for datasets from all DNA reference materials: a Phred quality score, b duplication rates, c read alignment rates, and d mapping quality. e–h RNA-seq Quality control metrics for datasets from all RNA reference materials: e Phred quality score, f duplication rates, g read alignment rates, and h 5'/3' bias. Boxplots show the interquartile range, median (line within the box), and whiskers (min–max excluding outliers). Data points represent individual FASTQ files for raw data metrics (Phred score, duplication; n = 42 for WES, 48 for RNA-seq) and unique biological samples for post-alignment metrics (alignment rate, mapping quality, bias; n = 21 for WES, 24 for RNA-seq). Significance was determined using paired t-test or Wilcoxon signed-rank test (*p < 0.05, **p < 0.01, ***p < 0.001, ****p < 0.0001)., as appropriate, following normality assessment (Shapiro–Wilk test). For significant results, effect size (Cohen's d) and 95% confidence interval (CI) are provided. See Methods and Additional file 2: Table S2 for details In RNA-seq analyses, Q40 similarly showed significantly higher mean base quality than Q30 (40.62 ± 1.38 vs. 35.74 ± 0.22, p < 0.001, Cohen's d = 3.88, Wilcoxon signed-rank test; Fig. 2e). As depicted in Fig. 2f, Q40 reduced duplication rates by 16.40% in RNA-seq (34.21 ± 3.30% vs. 40.85 ± 3.07%, p < 0.001, Cohen's d = −3.85, Wilcoxon signed-rank test). Alignment rates remained comparable between Q40 and Q30 in RNA-seq (96.88 ± 0.38% vs. 96.89 ± 0.26%, p > 0.05, Paired t-test; Fig. 2g). Notably, Q40 exhibited increased 5'/3' end sequence bias (0.05 ± 0.02 vs. 0.03 ± 0.02, p < 0.001, Cohen's d = 1.22, Paired t-test; Fig. 2h), likely attributable to enhanced resolution of GC-rich regions through avidity chemistry [7]. Detailed statistical information, including normality tests, sample sizes, means, mean differences with 95% confidence intervals, and Cohen's d values, is provided in Additional file 2: Table S2. A secondary observation revealed asymmetric read fidelity, with read1 exhibiting fewer errors than read2 for Q40 [11] (Fig. 2a, e). This disparity aligns with avidity chemistry's preferential stabilization of forward-strand synthesis during cluster amplification [7] (Table 2). The increased sequence bias observed in Fig. 2h correlates with improved detection of structurally challenging genomic regions, such as promoters and untranslated regions (UTRs), which are critical for isoform-level expression analysis. Higher Q score in WES enabled germline mutation detection at lower sequencing depths To assess variant detection performance across sequencing depths, we systematically subsampled WES datasets (Methods, Additional file 3: Table S3). We evaluated three context-specific metrics (Table 2): 1) genome-wide variant accuracy using Mendelian concordance rate (MCR) in family quartets; 2) precision in high-confidence regions via F1 score against benchmark variants (RM8398/Quartet DNA); and 3) accuracy in repetitive elements assessed by MCR. See Additional file 4: Table S4 for detailed metric values. Q40 significantly outperformed Q30 in germline mutation detection across the whole genome at varying sequencing coverages, even with lower sequencing depth. Notably, at 20 × coverage, Q40 achieved precision comparable to Q30 at 30 × for both SNV and InDel detection (SNV: Q40 at 20 × : 0.897 ± 0.001 vs. Q30 at 30 × : 0.884 ± 0.001, InDel: Q40 at 20 × : 0.693 ± 0.016 vs. Q30 at 30 × : 0.710 ± 0.012, Fig. 3a, b; Additional file 5: Fig. S1a, b, e). This represented a 33.3% reduction in required data depth while maintaining sensitivity, offering a substantial efficiency advantage (Additional file 5: Fig. S2, Additional file 6: Table S5). This advantage was evident at lower coverages (e.g., < 90 × for SNV, < 60 × for InDel), but the consistently small sample size (n = 3, as the MCR was calculated per Quartet pedigree sample set) limits definitive statistical conclusions. For instance, the large observed difference at 25 × (SNV: 0.928 ± 0.002 vs. 0.838 ± 0.001, Cohen's d = 44.87, Additional file 7: Table S6) was not statistically significant (p = 0.181), whereas significance was achieved at 30 × (p < 0.001), illustrating how limited sample size can constrain the detection of statistical significance even for substantial effects. Therefore, these results should be considered preliminary pending further validation.Fig. 3Germline SNV and InDel calling performance (Q30 vs. Q40). a, b Mendelian consistency rate (MCR) across the whole genome region at different sequencing depths. c, d F1 score in high-confidence regions at different sequencing depths. e, f F1 score in repeat regions (LINE: Long Interspersed Nuclear Element; LTR: Long Terminal Repeat; SINE: Short Interspersed Nuclear Element). Individual data points represent biological replicates of Quartet reference materials: sample-level for F1 score (n = 12, derived from three technical replicates each of samples D5, D6, F7, and M8) and family-level for MCR (n = 3, corresponding to three independent replicate groups of the complete pedigree). One sample with Q40 sequencing depth below 120 × was excluded from the 120 × and repeat region analyses, comparisons for these subsets were performed using the Wilcoxon rank-sum test. For all other analyses, statistical significance was determined using paired t-test or Wilcoxon signed-rank test (*p < 0.05, **p < 0.01, ***p < 0.001), as appropriate, following normality assessment (Shapiro–Wilk test). For significant results, effect size (Cohen's d) and 95% CI are provided. See Methods and Additional file 7: Table S6 for details Germline SNV and InDel calling performance (Q30 vs. Q40). a, b Mendelian consistency rate (MCR) across the whole genome region at different sequencing depths. c, d F1 score in high-confidence regions at different sequencing depths. e, f F1 score in repeat regions (LINE: Long Interspersed Nuclear Element; LTR: Long Terminal Repeat; SINE: Short Interspersed Nuclear Element). Individual data points represent biological replicates of Quartet reference materials: sample-level for F1 score (n = 12, derived from three technical replicates each of samples D5, D6, F7, and M8) and family-level for MCR (n = 3, corresponding to three independent replicate groups of the complete pedigree). One sample with Q40 sequencing depth below 120 × was excluded from the 120 × and repeat region analyses, comparisons for these subsets were performed using the Wilcoxon rank-sum test. For all other analyses, statistical significance was determined using paired t-test or Wilcoxon signed-rank test (*p < 0.05, **p < 0.01, ***p < 0.001), as appropriate, following normality assessment (Shapiro–Wilk test). For significant results, effect size (Cohen's d) and 95% CI are provided. See Methods and Additional file 7: Table S6 for details Notably, when assessed with a larger sample size (n = 12) in high-confidence genomic regions, this benefit remained clear and statistically robust. At 25 × depth, Q40 improved F1 scores by 3.5% for SNV detection (0.980 ± 0.001 vs. 0.947 ± 0.003, p < 0.0001, Cohen's d = 13.51, Paired t-test; Fig. 3c, Additional file 5: Fig. S1a) and by 5.6% for InDel detection (0.855 ± 0.009 vs. 0.809 ± 0.007, p < 0.001, Cohen's d = 3.98, Paired t-test; Fig. 3d, Additional file 5: Fig. S1b), demonstrating enhanced precision across variant types. At higher depths, however, a nuanced pattern emerged. Notably, at 90 × (p < 0.001, Cohen's d = −1.92, Paired t-test; Fig. 3c) and 120 × (p < 0.001, Cohen's d = −2.12, Paired t-test; Fig. 3c), Q40 yielded a slightly lower F1 score for SNV detection than Q30 (Additional file 7: Table S6). The apparent discrepancy stems from Q40's increased sensitivity, which detects additional variants whose validity is established by their Mendelian concordance within the Quartet pedigree rather than by inclusion in a reference dataset (Additional file 5: Fig. S3). However, as these additional variants are absent from the reference datasets used here—which were constructed from Q30-level data [14]—they were misclassified as false positives during evaluation. In most repetitive genomic regions, Q40 demonstrated superior (in low-complexity) or comparable detection accuracy to Q30 (Fig. 3e, f; Additional file 5: Fig. S1c, d). A nuanced pattern emerged specifically in simple repeat regions: here, Q40 detected a greater total number of InDels (392) than Q30 (381), yet a smaller proportion were Mendelian-concordant (Q40 vs. Q30: 311 vs. 314), resulting in a nominally lower MCR for Q40 (Q40 vs. Q30: 0.793 ± 0.034 vs. 0.830 ± 0.002, p = 0.4, Cohen's d = −1.45, Wilcoxon rank-sum test; Fig. 3f; Additional file 5: Fig. S1d; Fig. S3e, f), corroborating this region-specific discrepancy. The difference was not statistically significant, and owing to the limited sample size, its precise cause remains indeterminate. Collectively, these findings highlight simple repeats as a persistent challenge for variant assessment, due to their high mutability from polymerase slippage during DNA replication and technical artifacts introduced by PCR and sequencing that confound true variant calls [23, 24]. The specific error profiles of the Avidity (Element) and SBS (Illumina) chemistries in such complex genomic contexts may differ and contribute to the observed discrepancy. Higher Q score in WES enabled ultrasensitive detection of low-frequency tumor somatic variants To evaluate somatic variant detection performance across sequencing depths (Methods, Additional file 3: Table S3), three context-specific F1 score metrics were applied (Table 2): 1) precision in high-confidence regions; 2) accuracy in repetitive elements; and 3) accuracy for low-frequency somatic variants (VAF ≤ 0.2). Q40 sequencing demonstrated superior performance in somatic mutation detection across multiple metrics. Over a range of sequencing coverages, Q40 consistently achieved superior variant detection accuracy compared to Q30, with pronounced advantages at lower depths (≤ 60 ×, p < 0.01, Paired t-test; Fig. 4a, b). Notably, at 60 × sequencing depth, Q40 (SNV:0.833 ± 0.006, InDel: 0.676 ± 0.02) achieved comparable accuracy to Q30 at 90 × depth (SNV: 0.837 ± 0.003, InDel: 0.636 ± 0.019; Fig. 4a, b; Additional file 5: Fig. S4a), effectively reducing per-sample sequencing volume by 33.3% (Additional file 5: Fig. S2; Additional file 6: Table S5). Despite these advantages, both platforms showed limited detection accuracy in highly repetitive genomic regions (Fig. 4c, d; Additional file 5: Fig. S4c), underscoring persistent challenges in resolving complex structural variations. Similarly, in the analysis of somatic variants within simple repeat regions, Q40 exhibited a modest, non-significant decrease in detection performance compared to Q30, mirroring the pattern observed for germline variants in these challenging genomic contexts.Fig. 4Somatic SNV, InDel, and CNV calling performance (Q30 vs. Q40). a, b F1 score for somatic SNV and InDel calling in high-confidence genomic regions at different sequencing depths. c, d F1 score for somatic SNV and InDel calling in repeat regions. e, f F1 score for somatic SNV and InDel calling for low-frequency somatic variants (VAF ≤ 0.2). g Technical repeatability of CNV measured by Jaccard Index. h Correlation of copy numbers between technical replicates measured by Pearson correlation coefficient (PCC). Individual data points represent biological replicates: a–d (n = 3), g (n = 18, including CNV gain, loss, and LOH (Loss of Heterozygosity)), and Fig. 4 h (n = 6, including CNV gain–loss and loss). One sample with Q30 sequencing depth at 30 × lacked CNV detection in several genomic regions, comparisons for these subsets were performed using the Wilcoxon rank-sum test. For all other analyses, statistical significance was determined using paired t-test or Wilcoxon signed-rank test (*p < 0.05, **p < 0.01, ***p < 0.001), as appropriate, following normality assessment (Shapiro–Wilk test). For significant results, effect size (Cohen's d) and 95% CI are provided. See Methods and Additional file 7: Table S6 for details Somatic SNV, InDel, and CNV calling performance (Q30 vs. Q40). a, b F1 score for somatic SNV and InDel calling in high-confidence genomic regions at different sequencing depths. c, d F1 score for somatic SNV and InDel calling in repeat regions. e, f F1 score for somatic SNV and InDel calling for low-frequency somatic variants (VAF ≤ 0.2). g Technical repeatability of CNV measured by Jaccard Index. h Correlation of copy numbers between technical replicates measured by Pearson correlation coefficient (PCC). Individual data points represent biological replicates: a–d (n = 3), g (n = 18, including CNV gain, loss, and LOH (Loss of Heterozygosity)), and Fig. 4 h (n = 6, including CNV gain–loss and loss). One sample with Q30 sequencing depth at 30 × lacked CNV detection in several genomic regions, comparisons for these subsets were performed using the Wilcoxon rank-sum test. For all other analyses, statistical significance was determined using paired t-test or Wilcoxon signed-rank test (*p < 0.05, **p < 0.01, ***p < 0.001), as appropriate, following normality assessment (Shapiro–Wilk test). For significant results, effect size (Cohen's d) and 95% CI are provided. See Methods and Additional file 7: Table S6 for details Q40 exhibited no significant improvement was observed for low-frequency SNVvariants (Fig. 4e), but significantly enhanced sensitivity for clinically relevant low-frequency InDel (VAF ≤ 0.2), outperforming Q30 by 33.3% (p < 0.01, Cohen's d = 14.87, Paired t-test; Fig. 4f). Correlation analysis further revealed Q40's stronger linear association with reference VAF values, particularly within the low-frequency mutation range critical for early relapse monitoring (Additional file 5: Fig. S4b). Beyond SNV and InDel, CNV is also a significant driver of cancer progression [25, 26]. At 30 × sequencing depth, Q40 achieved 60.3% reproducibility in CNV detection (All CNV types), far exceeding Q30's reproducibility of 10.4% (p < 0.001, Cohen's d = 2.12, Wilcoxon rank-sum test; Fig. 4g; Additional file 5: Fig. S5a, c, e). At this lower sequencing depth, Q30 showed missing CNV detection in several genomic regions (Additional file 5: Fig. S5b, d, f), precluding paired statistical analysis, whereas Q40 maintained consistent detection and demonstrated superior regional resolution. In addition to its detection robustness, Q40 also showed a substantial cost advantage: while Q30 requires ≥ 90 × depth to reach 54.1% reproducibility, Q40 maintains 80.7% reproducibility at 60 × (Fig. 4g), reducing per-sample sequencing volume by 33.3% (Additional file 5: Fig. S2, Additional file 6: Table S5). Notably, Q40 also demonstrated a 148.6% higher Pearson correlation coefficient of copy number (cnPCC) at 30 × (0.875 ± 0.008 vs. 0.347 ± 0.346, p = 0.12, Cohen's d = 1.52, Paired t-test; Fig. 4h), although no significant improvement was observed across CNV types (Additional file 5: Fig. S5g, h). Despite these advancements, CNV loss detection remained challenging. Q40 achieved only 25.1% reproducibility for CNV loss (Additional file 5: Fig. S5c, d) versus 79.7% for CNV gain (Additional file 5: Fig. S5a, b) at 30 × depth. This discrepancy was likely attributable to reduced library complexity in deletion regions and ambiguous boundary signals in low-mappability regions [27]. Higher Q score in RNA-seq enhanced biological resolution To evaluate gene quantification accuracy of Q40 data across varying data volumes, we downsampled RNA-seq FASTQ datasets to equivalent total bases of 2 Gb, 4 Gb, 6 Gb, 8 Gb, and 10 Gb (Methods). We utilized a comprehensive set of metrics for performance evaluation (Table 3; Additional file 8: Table S7), with particular focus on Q40's capacity to detect minor biological variations in comparison to Q30 data. While Q40 and Q30 exhibited comparable gene quantification correlation with reference datasets (p > 0.05, Wilcoxon signed-rank test; Fig. 5a, b), Q40 demonstrated superior performance in differential expression analysis. Q40 showed significantly higher Matthews Correlation Coefficient for DEG detection (0.673 ± 0.023 vs. 0.657 ± 0.023, p < 0.01, Cohen's d = 1.48, Wilcoxon signed-rank test; Fig. 5e, f) and a 13.1% increase in Signal-to-Noise Ratio (32.76 ± 2.11vs. 29.01 ± 0.89, p = 0.05, Cohen's d = 3.07, Wilcoxon signed-rank test; Fig. 5c, d). These improvements in classification accuracy and signal clarity enabled better discrimination of biologically similar samples (e.g., Quartet twins D5 and D6, Additional file 5: Fig. S6) [21] even at reduced volumes (2–4 Gb), and were consistently observed across MAQC RNA (Additional file 5: Fig. S7a-d). It is also noted that Q40 performed comparably to Q30 in the ERCC spike-in control experiments (Additional file 5: Fig. S7e, f).Fig. 5Gene expression quantification performance at different sequencing depths (Q30 vs. Q40). a, b Pearson correlation coefficient (PCC) with Quartet reference datasets (D5/D6, F7/D6, and M8/D6). c, d Signal-to-noise ratio (SNR) for differentiating the four groups of Quartet samples. e, f Matthews correlation coefficient (MCC) with Quartet reference datasets (D5/D6, F7/D6, and M8/D6). g Coefficient of variation (CV) analysis of technical replicates. Genes in Quartet sample F7 were stratified into six bins by mean FPKM (Fragments Per Kilobase of transcript per Million mapped fragments): [0,2), [2,4), [4,6), [6,8), [8,10), ≥ 10. Individual data points represent biological replicates: Fig. 5b and d (n = 5), Fig. 5f (n = 15, corresponding to the three comparison groups in Fig. 5e), and g (n = number of genes detected in all three replicates). Comparisons of the coefficient of variation (CV) of FPKM were performed using the Wilcoxon rank-sum test, as informed by a Kolmogorov test. For all other analyses, statistical significance was determined using paired t-test or Wilcoxon signed-rank test (*p < 0.05, **p < 0.01, ***p < 0.001), as appropriate, following normality assessment (Shapiro–Wilk test). For significant results, effect size (Cohen's d) and 95% CI are provided. See Methods and Additional file 8: Table S7 for details Gene expression quantification performance at different sequencing depths (Q30 vs. Q40). a, b Pearson correlation coefficient (PCC) with Quartet reference datasets (D5/D6, F7/D6, and M8/D6). c, d Signal-to-noise ratio (SNR) for differentiating the four groups of Quartet samples. e, f Matthews correlation coefficient (MCC) with Quartet reference datasets (D5/D6, F7/D6, and M8/D6). g Coefficient of variation (CV) analysis of technical replicates. Genes in Quartet sample F7 were stratified into six bins by mean FPKM (Fragments Per Kilobase of transcript per Million mapped fragments): [0,2), [2,4), [4,6), [6,8), [8,10), ≥ 10. Individual data points represent biological replicates: Fig. 5b and d (n = 5), Fig. 5f (n = 15, corresponding to the three comparison groups in Fig. 5e), and g (n = number of genes detected in all three replicates). Comparisons of the coefficient of variation (CV) of FPKM were performed using the Wilcoxon rank-sum test, as informed by a Kolmogorov test. For all other analyses, statistical significance was determined using paired t-test or Wilcoxon signed-rank test (*p < 0.05, **p < 0.01, ***p < 0.001), as appropriate, following normality assessment (Shapiro–Wilk test). For significant results, effect size (Cohen's d) and 95% CI are provided. See Methods and Additional file 8: Table S7 for details The observed improvement stems fundamentally from Q40's expanded dynamic range. At matched sequencing depths, it detected more genes than Q30 (Additional file 5: Fig. S7g)—a capability particularly impactful for low-abundance transcripts (< 2 FPKM) where it reduced technical variation by 4.8% (p < 0.0001, Cohen’s d = −0.07, Wilcoxon rank-sum test; Fig. 5g), despite comparable quantification variability overall (Additional file 5: Fig. S7h). However, both platforms exhibited comparable accuracy in resolving titration-derived expression gradients (Additional file 5: Fig. S7i, j), suggesting that intrinsic biological noise, rather than sequencing accuracy, is the primary limiting factor in sensitivity. The improvement in quantification accuracy observed with Q40 data in this study may stem from the retention of more sequencing data during preprocessing [28, 29], owing to its higher initial base quality. This, in turn, further contributes to its 13.1% higher signal-to-noise ratio (SNR), which demonstrates critical value for detecting subtle biological variations. Moreover, Q40 with only 4 Gb of data achieved performance comparable to, or even superior to, that of Q30 with 10 Gb (Fig. 5a, c, e), effectively reducing the required data volume by 60% and highlighting its potential to lower sequencing costs (Additional file 5: Fig. S2; Additional file 6: Table S5). Discussion Our study demonstrates that sequencing under the Q40 quality standard (as exemplified by Element AVITI) achieves comprehensive improvements in data quality over the conventional Q30 standard (as exemplified by Illumina NovaSeq 6000). However, a key consideration in interpreting these results is the involvement of two distinct sequencing platforms employing different core chemistries—SBS (Illumina) versus Avidity (Element)—(Table 4). These inherent technological differences are known confounders that can independently influence metrics such as duplication rates and sequence bias [30]. Therefore, the observed performance advantages for "Q40" likely represent a composite effect, attributable to both the higher intrinsic base-calling accuracy (Phred Q score ≥ 40) and the specific attributes of the Avidity chemistry. Table 4Characteristics of Q40 and Q30 sequencing methodsQ40Q30SequencerElement AVITIIllumina NovaSeq6000Sequencing reagentAVITI 2 × 150 Sequencing Kit Cloudbreak LowNovaSeq 6000 S4 Reagent Kit v1.5 (300 cycles)Library preparationSingle-stranded circular DNA (ssCirDNA)DNA fragmentation and adapter ligationFragment amplificationRolling circle amplification increases the copy number of the target sequence by about a thousandfold, forming covalently linked DNA nanoballs on the chip (similar to a ball of yarn)Amplification of template DNA using bridge PCR to form DNA clustersBase recognitionSequencing with affinity probes, with a fluorescent group at the center and four arms extending outwards, each arm terminated with a baseSequencing with fluorescently labeled dNTPs Table 4Characteristics of Q40 and Q30 sequencing methodsQ40Q30SequencerElement AVITIIllumina NovaSeq6000Sequencing reagentAVITI 2 × 150 Sequencing Kit Cloudbreak LowNovaSeq 6000 S4 Reagent Kit v1.5 (300 cycles)Library preparationSingle-stranded circular DNA (ssCirDNA)DNA fragmentation and adapter ligationFragment amplificationRolling circle amplification increases the copy number of the target sequence by about a thousandfold, forming covalently linked DNA nanoballs on the chip (similar to a ball of yarn)Amplification of template DNA using bridge PCR to form DNA clustersBase recognitionSequencing with affinity probes, with a fluorescent group at the center and four arms extending outwards, each arm terminated with a baseSequencing with fluorescently labeled dNTPs Characteristics of Q40 and Q30 sequencing methods Collectively, the Q40 technology, as represented by the Element AVITI platform in this study, may set a new benchmark for genomic data performance, achieving a 40% higher base-calling accuracy and 14–50% lower duplication rates across RNA-seq and WES datasets (Additional file 2: Table S2). These advancements enabled reliable detection of biological changes at reduced sequencing depths and redefined quality thresholds for precision oncology applications [31]. While ensuring equivalent accuracy in variant detection with 33.3% lower sequencing depth than Q30 (Q40 at 20 × vs. Q30 at 30 × for germline variants; Q40 at 60 × vs. Q30 at 90 × for somatic variants), Q40 shows clear advantages over current NGS cost paradigms, with theoretical cost reductions of 2.2–5.8% for WES and 28.8–31.7% for WGS (Additional file 5: Fig. S2; Additional file 6: Table S5), while maintaining ≥ 90% accuracy (F1 score) for germline variants (Fig. 3a-d) and ≥ 80% (F1 score) for somatic variants (Fig. 4a, b). This breakthrough has transformative potential for clinical diagnostics: in oncology, it enables cost-effective tumor profiling from limited biopsy samples (e.g., liquid biopsies) with faster turnaround times; in genetic disease screening, it lowers the barrier for population-scale germline testing while preserving sensitivity for low-frequency pathogenic variants. The ability of Q40 technology to resolve low-frequency mutations (≤ 0.2) with a 30% sensitivity improvement (Fig. 4e) could redefine minimal residual disease (MRD) monitoring [32], where detecting rare tumor-derived mutations is critical for early relapse prediction. Furthermore, its enhanced overall CNV detection reproducibility at 30 × depth (Q40 60.3% vs. Q30 10.4%; Fig. 4g) provides a clinically actionable solution for identifying oncogenic amplifications [26], which are currently challenging to detect reliably in low-purity clinical samples. While Q40 demonstrates broad advantages in genomic variant detection, its performance in calling variants within simple repeat regions warrants further rigorous evaluation. For gene quantification, the superior initial base quality of Q40 data serves as the fundamental driver (Fig. 2e), allowing for the retention of more sequencing data during preprocessing due to less aggressive quality filtering. This enhanced data integrity improves detection stability for transcripts, which underlies the observed gain in quantification accuracy. Ultimately, these combined effects further contribute to the 13.1% higher SNR, suggesting a potential advantage for Q40 in detecting subtle biological variations. Moreover, given that Q40 with 4 Gb of data matches the performance of Q30 with 10 Gb, this theoretically translates to a 24% reduction in RNA-seq sequencing costs (Additional file 5: Fig. S3; Additional file 6: Table S5). However, several limitations warrant consideration. First, while the current study demonstrated a significant advantage of Q40 in detecting low-frequency somatic mutations (VAF ≤ 0.2) due to insufficient ultra-low frequency mutations (VAF < 0.001) in the reference dataset, further validation is required to assess Q40's detection accuracy for mutations within the MRD (range 0.001 to 0.01) [33]. Second, the conclusions about variant detection drawn here are primarily applicable to SNV/InDel and CNV. Owing to the inherent limitations of short-read WES in resolving structural variant (SV) and complex repetitive regions, dedicated studies using long-read platforms are needed to assess the performance of ultra-high-quality (Q40) data for SV detection [34]. Third, this benchmarking was conducted using well-characterized reference materials under controlled conditions. Performance in real-world, heterogeneous clinical samples may vary due to factors such as sample degradation, lower input mass, and higher background noise, and should be validated in dedicated clinical studies [35–37]. Finally, this study exclusively utilized data from one Q40 platform. To comprehensively evaluate performance enhancement achieved by increasing Q scores, future investigations should incorporate Q50 sequencing datasets for comparative analysis. Based on these findings and limitations, we propose several directions for future work. First, integrating Q40 with CRISPR-based enrichment strategies to achieve ultra-sensitive detection of ultra-low frequency mutations (VAF < 0.001) for early cancer screening [38, 39]. Second, developing Q score optimization frameworks tailored to specific clinical scenarios, such as refining the thresholds (Q30, Q40, and Q50) for germline and somatic analyses.Third, expanding the technological scope is essential. This includes leveraging long-read sequencing to benchmark Q40's utility for structural variant detection, and establishing public reference datasets with well-defined ultra-low-frequency variants to standardize validation across labs. Collectively, pursuing these directions will address key real-world diagnostic challenges. The enhanced sensitivity and precision of Q40 data will be crucial for reliably analyzing heterogeneous clinical samples and for positioning it as a foundational technology for robust next-generation molecular diagnostics. Conclusions Against the growing backdrop of genomic benchmarking, where established resources such as GIAB [40, 41], SEQC2 [42, 43], Platinum Pedigree [44] and Quartet [14] have set extensive benchmarks for germline variants, comprehensive benchmarks for somatic variants remain relatively scarce. The SEQC2 [17] project has made significant contributions in this area, and the community is actively working to detect structural variants at ultra-low allele frequencies [45] and to establish pangenome-informed resources for somatic variants [46]. Within this context, our findings on germline and somatic mutations using the Q40 platform (represented by Element AVITI) provide crucial complementary data and a new reference point for these broader community goals. In this study, we utilize a comprehensive set of well-characterized DNA/RNA reference materials to establish Q40 as a cost-effective solution for enhancing the detection of rare variants. Our results demonstrate that Q40 sequencing achieves comparable accuracy at 33–50% lower depth across variant types: germline variants (20 × vs. 30 ×), somatic SNV/InDel (60 × vs. 90 ×), and somatic CNV (30 × vs. 60 ×). Critically, Q40 outperforms Q30 in detecting low-frequency variants (VAF ≤ 0.2), showing 33.3% higher sensitivity, and exhibits dramatically improved reproducibility in CNV detection (60.3% vs. 10.4%) at 30 × depth. Furthermore, Q40 data yield 13.1% higher SNR in biological resolution and utilize only 4 Gb of data, demonstrated performance on par with or exceeding that of Q30 using 10 Gb. Collectively, these findings position Q40 sequencing as a highly promising and cost-effective approach for enhancing the detection of low- frequency variants, suggesting potential applicability in precision oncology. Its specific performance in clinical scenarios will necessitate future confirmation. Methods Reference materials and datasets Quartet RNA reference materials and reference datasets RNA reference materials were derived from Epstein-Barr virus-transformed B-lymphocyte cell lines of a family quartet (father: Quartet-F7; mother: Quartet-M8; identical twin daughters: Quartet-D5, Quartet-D6) within the Fudan Taizhou cohort [13]. These materials establish a ratio-based transcriptome "ground truth" for cross-platform and cross-laboratory benchmarking. Relative gene expression and differentially expressed genes (DEGs) were obtained from the Quartet Data Portal (https://chinese-quartet.org/) [47]. The dataset comprises 31,155 gene expression/DEG results from M8/D6, F7/D6, and D5/D6 comparisons. After intersection with Ensembl release-109 annotations, 30,976 results (5,036 DEGs) were retained. MAQC RNA reference materials and reference datasets The MAQC project conducted repetitive analyses across various gene-expression platforms and multiple test sites using four libraries composed of two types of RNA samples [19], namely the Universal Human Reference RNA (UHRR) from Stratagene (now Agilent) as Sample A and the Human Brain Reference RNA (HBRR) from Ambion (now Thermo Fisher) as Sample B. Normalized expression values for 1,044 genes (Gene Expression Omnibus (GEO): GSE5350) [48] were analyzed. After excluding genes with undetectable CT values (CT > 35 or underdermined), 830 genes were retained. DEGs between Samples A and B were identified by comparing mean expression across four technical replicates. ERCC spike-in reference materials and built-in truth ERCC has developed a set of 96 exogenous RNA spike-in controls that were recognized by the NIST as the Reference Material (SRM) 2374 and serve as a benchmark for technical replicates in differential gene expression studies [20, 49]. The ERCC RNA control set encompasses 92 distinct genes across Mix 1 and Mix 2, each with defined and diverse concentrations, providing a robust framework for assessing absolute expression levels. Furthermore, ERCC genes have four subgroups (a, b, c, and d), characterized by their specific concentration ratios of 4:1, 1:2, 2:3, and 1:1 between Mix 1 and Mix 2, respectively. This stratification enabled evaluation of relative expression level measurements precision. NIST DNA reference material RM8398 and reference datasets The NIST DNA reference material RM8398/HG001 was derived from the peripheral blood B lymphocytes of a Caucasian female, immortalized through Epstein-Barr virus (EBV) transformation [50]. The corresponding germline reference dataset for RM8398 encompasses 3,254,352 SNVs and 46,768 InDels, with a high-confidence interval spanning 2,512 megabases (Mb) of genomic regions, representing 83.3% of the reference genome (GRCh38, Chr1-22, X). Quartet DNA reference materials and reference datasets Genomic DNA samples from the Quartet family members (D5, D6, F7, and M8) [National Institute of Metrology, China/Fudan University] enable quality assessment via Mendelian inheritance patterns [14]. Each sample contains ~ 4.2 million high-confidence variants across 2,660 Mb (87.8% of GRCh38). HCC1395/BL DNA reference materials and reference datasets We used the SEQC2 consortium's HCC1395 (TNBC)/HCC1395BL (matched B-cell) paired whole-genome DNA reference materials [51, 52], which contain somatic mutation reference datasets consist of approximately 40,000 SNVs and around 2,000 InDels, covering 82% of genomic regions with high confidence [16, 17]. RNA sequencing RNA-seq library preparation RNA extracts from the four Quartet samples (D5, D6, F7, and M8) were supplemented with ERCC ExFold spike-in mixes (Mix 1 in M8; Mix 2 in D6; Thermo Fisher, 4,456,740) per manufacturer specifications. Samples T1 and T2 represent mixtures of samples M8 and D6 at the defined ratios of 3:1 and 1:3, respectively, and thus hold "built-in truths" of sample mixing ratios. MAQC reference RNAs (Universal Human Reference RNA [Agilent 740000]; Human Brain Reference RNA [Thermo Fisher QS0611]) were designated Samples A/B per consortium standards. Quartet D6 served as the differential expression control for M8/F7/D5/T1/T2. Twenty-four aliquots (8 samples × 3 technical replicates; 200 ng/μL in 8 μL volumes) were stored at − 80 °C. RNA sequencing by illumina and element platforms During the sequencing preparation, pooled libraries underwent high-throughput sequencing on the Illumina NovaSeq 6000 System, adopting a 2 × 150 base pair read configuration and integrating index reads according to the vendor's guidelines. The system harnesses reversible termination chemistry and fluorescently labeled dNTPs for precise SBS. Furthermore, the Element AVITI system sequenced quantified libraries with a 2 × 150 base pair read setup, employing avidity chemistry and RCA to minimize errors and provide a high-accuracy, cost-effective solution. RNA-seq data analysis workflow The RNA-seq data processing pipeline was designed with a focus on reproducibility and quality control, drawing inspiration from the Quartet project's framework for transcriptomic data assessment [21]. The key steps are summarized below, with all custom analytical scripts available in the GitHub repository (https://github.com/duanshumeng/Q40_DNA_RNA/). 1) Raw sequencing reads were processed using fastp (v0.19.6) [29] with default parameters to remove adapter sequences and low-quality bases. 2) Subsequent quality control was performed with FastQC (v0.11.5). 3) Clean reads were aligned to the human reference genome (Homo_sapiens.GRCh38.dna.primary_assembly.fa) and the ERCC spike-in sequences [20] using HISAT2 (v2.7.10). 4) The alignment employed non-default parameters as detailed in the provided script (rnaseq-fastq2bam.sh). 5) The quality of the resulting BAM files was assessed using Qualimap (v2.0.0) [53]. 6) Transcript abundance was estimated with StringTie (v2.1.7) [54] as detailed in the provided script (gene-quantification.sh). 7) Differential expression analysis was conducted using limma (v3.54.0) [55] and edgeR (v3.40.2) [56] within the R environment, following the workflow documented in the https://github.com/chinese-quartet/quartet-rseqc-report. Whole-exome sequencing WES library preparation DNA was enzymatically fragmented. A total of 300 ng of DNA was utilized to prepare libraries via PCR. Exonic regions were captured from 300 ng library using the Twist Exome 2.0 plus Comprehensive Exome Spike-in kit. Hybrid capture PCR was performed for 6 cycles. WES sequencing by illumina and element platforms Sequencing was conducted on the Illumina platform using the Novaseq 6000 instrument. The NovaSeq 6000 S4 Reagent Kit v1.5 (300 cycles) was employed for the sequencing process. The sequencing was carried out in a paired-end configuration with a read length of 150 bp per end, following the standard protocols provided by the manufacturer. Sequencing was conducted on the Element platform using the Element AVITI instrument. The AVITI 2 × 150 Sequencing Kit Cloudbreak Low was employed for the sequencing process. The sequencing was carried out in a paired-end configuration with a read length of 150 bp per end, following the standard protocols provided by the manufacturer. WES data analysis workflow The WES data processing pipeline followed established best practices for somatic and germline variant analysis [17, 57]. All analytical scripts are available in the GitHub repository (https://github.com/duanshumeng/Q40_DNA_RNA/). 1) Raw sequencing reads were processed using fastp (v0.19.6) [29] to eliminate primer sequences and purge low-quality reads from the raw data; 2) The resultant datasets were then refined through a quality control process employing fastqc (v0.11.5); 3) Alignment against the GRCh38.d1.vd1.fa reference genome, was efficiently executed with sentieon (v202112.05) [58], which incorporates an optimized BWA aligner as detailed in the provided script (wes-fastq2bam.wdl); 4) qualimap (v 2.0.0) was subsequently applied to ensure the quality of the aligned sequences; 5) Somatic and germline SNV/InDel within WES library one were detected using GATK (v4.4.0, Mutect2 for somatic variants, HaplotypeCaller for germline variants) as detailed in the provided scripts (Mutect2_calling-somatic.sh and Haplotyper_calling-germline.sh); 6) ASCAT (v3.1.2) was engaged for the detection of somatic CNV [59] as detailed in the provided script (ascat-WES.r); and 7) All subsequent visualizations were crafted within R version 4.2.2. Data downsampling To eliminate the influence of varying data volumes on the comparative assessment of data quality between Q30 and Q40 platforms, we performed in silico downsampling on the raw sequencing data (FASTQ files). Using seqtk (v1.4-r122), random sampling was conducted at the read-pair level to preserve the paired-end structure, with 33,300,000 reads randomly selected for each of the read1 and read2 FASTQ files. An identical random seed was applied during the subsampling of both read1 and read2 files, guaranteeing that the selected reads remained perfectly paired. All downstream analyses were performed again on the subsampled data. To evaluate variant and expression quantification performance across different sequencing depths and data volumes, we performed systematic in silico downsampling on both WES and RNA-seq datasets. Starting from high-coverage parent BAM files (depth > 120 × for WES), generated by aligning the uniformly subsampled 10 Gb FASTQ data to the reference genome, we simulated a range of coverage levels. For WES data, target depths of 10 ×, 15 ×, 20 ×, 25 ×, 30 ×, 60 ×, 90 ×, and 120 × were simulated. The required number of read pairs for each depth was calculated based on the exome target region size (37.35 Mb) and a read length of 150 bp, using the formula: Number of pairs = (Target Depth × Target Region Size)/(Read Length). This yielded a subsampling fraction (seed_freq) for each BAM file. For RNA-seq data, target data volumes of 2, 4, 6, 8, and 10 Gb were simulated. The required number of reads was calculated proportionally from the 10 Gb baseline (approximately 66.7 million reads). For example, the target read count for 8 Gb was (8/10) × 66.7 million reads. For all datasets, random downsampling at the read-pair level was performed using samtools view -s seed_freq on the parent BAM files, producing new BAM files for each target depth or data volume. These files were then processed independently through the respective analysis pipelines. Theoretical cost-effectiveness analysis We modeled the economic impact of Q40 sequencing using a cost-comparison framework in R. For gene expression (RNA-seq) and variant detection applications (WGS/WES), total costs were calculated as: Total Cost = Library preparation cost + (Sequencing depth × Effective genome size × Cost per GB). Library preparation was fixed at 300 RMB per sample, with sequencing costs at 30 RMB/GB based on 2024 market rates. Required depths were set at 30 × (germline) and 90 × (somatic) for Q30, reduced to 20 × and 60 × for Q40 based on accuracy-equivalence data. Two metrics quantified Q40 benefits:i)Data compression ratio (%) = [(Q30 data volume—Q40 data volume)/Q30 data volume] × 100%;ii)Cost compression ratio (%) = [(Q30 total cost—Q40 total cost)/Q30 total cost] × 100%. i)Data compression ratio (%) = [(Q30 data volume—Q40 data volume)/Q30 data volume] × 100%; Data compression ratio (%) = [(Q30 data volume—Q40 data volume)/Q30 data volume] × 100%; ii)Cost compression ratio (%) = [(Q30 total cost—Q40 total cost)/Q30 total cost] × 100%. Cost compression ratio (%) = [(Q30 total cost—Q40 total cost)/Q30 total cost] × 100%. Data volume was calculated as depth × effective genome size (3.2 GB for WGS, 0.035 GB for WES). RNA-seq volumes were based on typical requirements (10 GB for Q30, 4 GB for Q40). Data quality metrics Phred Quality Score and Duplication Rates were calculated by fastqc (v0.11.5), while Mapping Quality, Aligned Reads, and 5'/3' bias were calculated by qualimap (v2.0.0). Percentage of optimization To quantify quality improvement, we calculated the optimization percentage as:\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\frac{|{\mu}_{Q40}-{\mu}_{Q30}|}{{\mu}_{Q30}} \times 100\%$$\end{document}|μQ40-μQ30|μQ30×100%, where μμ represents the group mean. Statistical significance was assessed using Student's t-test. Statistical analysis Data were presented as mean ± standard deviation (SD) from n independent biological or technical replicates. For each metric, we first determined whether the data followed a paired or independent design. For paired comparisons, we assessed the normality of the paired differences using the Shapiro–Wilk test. If normality was met (p > 0.05), a paired-sample t-test was applied; otherwise, the Wilcoxon signed-rank test was used. For independent comparisons, we tested each group for normality. If both groups were normally distributed, an independent two-sample t-test was used; if not, the Wilcoxon rank-sum test (Mann–Whitney U test) was applied. The paired t-test was conducted using stats::t.test (paired = TRUE/FALSE), and the Wilcoxon signed-rank test was conducted using stats::wilcox.test (paired = TRUE/FALSE). For statistically significant results, the effect size was quantified by reporting Cohen's d for paired samples along with its 95% confidence interval (CI), calculated via a custom R script. The formula for paired Cohen's d (dz) is:\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${d}_{z }=\frac{{M}_{diff}}{{SD}_{diff}}$$\end{document}dz=MdiffSDdiffwhere \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${M}_{diff}$$\end{document}Mdiff is the mean of the paired differences, and \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${SD}_{diff}$$\end{document}SDdiff is the standard deviation of those differences. The formula for unpaired Cohen's d (dz) is:\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${d}_{z}= \frac{{M}_{Q40}-{M}_{Q30}}{{SD}_{pooled}}$$\end{document}dz=MQ40-MQ30SDpooledwhere \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${M}_{Q40}$$\end{document}MQ40 and \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${M}_{Q30}$$\end{document}MQ30 are the mean values of Q40 and Q30, respectively, and \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${SD}_{pooled}$$\end{document}SDpooled denotes the pooled standard deviation of the two groups. The conventional criteria for effect size interpretation are as follows: a positive d value indicates that the mean of the first group (e.g., Q40) is greater than that of the second group (e.g., Q30), whereas a negative dz indicates the opposite. Regarding the magnitude: |dz|< 0.2 indicates a negligible effect, 0.2 ≤|dz|< 0.5 a small effect, 0.5 ≤|dz|< 0.8 a medium effect, and |dz|≥ 0.8 a large effect.All statistical analyses were performed using R version 4.2.2. Performance metrics All computational scripts for performance metric analysis are available in the GitHub repository (https://github.com/duanshumeng/Q40_DNA_RNA/). SNR (Signal-to-Noise Ratio) We quantified biological signal distinguishability using the Quartet project's SNR metric [13]. After principal component analysis (PCA), a weighted Euclidean distance was computed between samples in the first two principal component dimensions. Signal was defined as the root mean square (RMS) of distances between different biological groups, while noise was the RMS of distances between technical replicates within the same group. SNR was calculated as \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$10{log}_{10}{\left(\frac{\textrm{Signal}}{\mathrm{Noise}}\right)}^{2}$$\end{document}10log10SignalNoise2.The analysis was performed using a custom R script (Fig. 5.2.RNA_Quartet.R). PCC (Pearson correlation coefficient) The PCC was calculated to assess performance across the four domains: 1) Absolute gene expression accuracy was assessed by correlating log₂(FPKM) values with ERCC spike-in concentrations and TaqMan quantification results [21]; 2) Relative expression accuracy was evaluated by correlating sample-pair expression ratios with TaqMan and Quartet reference fold-changes; 3) Somatic variant calling precision was measured via correlation between test sample VAF and HCC1395/BL reference VAFs; and 4) CNV reproducibility was determined by assessing consistency in copy number estimates across replicate pairs. Analyses were performed in R (v4.2.2) using the cor() function and in Python (v3.8) using the corr() function, as implemented in the provided scripts (Fig. 5.2.RNA_Quartet.R and VAF-PCC.py). F1 score F1 score was used to evaluate the accuracy of variant detection [16] in high-confidence genomic regions. True Positives (TP): Variants detected in the sample that were also in the high-confidence variant set. False Positives (FP): Variants detected in the sample but not in the high-confidence set. False Negatives (FN): Variants in the high-confidence set that were not detected. True Negatives (TN): Variants neither detected nor in the high-confidence set. Precision was the ratio of TP to the total detected variants (TP + FP), indicating the reliability of detected variants. Recall (or Sensitivity) was the ratio of TP to the total high-confidence variants (TP + FN), indicating the detection rate of true variants. While recall was not affected by FP, it can be artificially inflated by increasing the number of detected variants. Since the number of variants was low compared to the genome size, F1 score combines precision and recall into a single metric, balancing the trade-off between them. It was calculated as shown in Additional file 5: Fig. S8a. We employed hap.py and som.py (https://github.com/Illumina/hap.py) to calculate the precision and recall for germline and somatic variants, respectively. The F1 score was then computed using custom Python scripts (HCC1395_F1_and_Reproducibility.py and Quartet_F1_and_Reproducibility.py). MCC (Matthews Correlation Coefficient) In this study, we employed Matthews Correlation Coefficient (MCC) [21], a robust bioinformatics metric that encapsulates both sensitivity and specificity, to assess the concordance between differentially expressed genes (DEGs) detected from our experimental datasets and a curated set of reference DEGs. This approach, termed "MCC of DEGs", amalgamates true-positive and true-negative calls through a consensus voting mechanism. By discerning DEGs and non-DEGs, we determined the respective counts of true positives, true negatives, false positives, and false negatives, enabling MCC calculation as delineated in Additional file 5: Fig. S8b. The analysis was implemented using the custom R script (Fig. 5.2.RNA_Quartet.R). LODR (Limit of Detection of Ratio) The LODR metric defines the minimum expression level required to detect differentially expressed transcripts at a specified confidence threshold, quantifying diagnostic power relative to the transcript abundance and fold change [20, 49]. It was modeled and calculated using the erccdashboard R package [20]. This empirically derived threshold was calculated using ERCC spike-in control ratios. We determined LODR by modeling the relationship between differential expression P-values and expression levels (counts), defining the minimum signal where the upper confidence bound of this model intersects the false discovery rate (FDR) threshold. AUC (Area Under the Curve) AUC quantifies differential expression detection accuracy using receiver operating characteristic (ROC) curves [20, 49]. It was modeled and calculated using the erccdashboard R package [20]. We computed AUC from ROC curves plotting true-positive rate against false-positive rate using ranked differential expression P-values. Values range from 0.5 (random discrimination) to 1.0 (perfect discrimination), with higher values indicating superior transcript-level detection capability. MCR (Mendelian Consistent Rate) MCR was calculated as the proportion of variants violating Mendelian inheritance across Quartet samples15. Specifically, all detected variant sites constituted the Total_Variants distribution. From these, variants violating inheritance patterns (Discordant_Variants) were identified and removed. The remaining set was defined as Mendelian_Concordant_Variants. The final MCR was computed as Mendelian_Concordant_Variants/Total_Variants. Discordant variants were defined as follows: (1) variants absent in both twins or inconsistent with parental haplotypes; (2) excluding large deletions misclassified by structural variant callers; and (3) restricting structural variant analysis to Quartet-D5 due to homozygous reference/no-call ambiguity. Genotype-agnostic discordance criteria were applied. The analysis was implemented using a custom script (Quartet-MCR.sh). Reproducibility In this study, we assessed four distinct aspects of technical reproducibility as follows [43]:\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\text{Reproducibility Rate}=\frac{1}{2}\left(\frac{{n}^{a}}{{N}^{a}}+\frac{{n}^{b}}{{N}^{b}}\right)$$\end{document}Reproducibility Rate=12naNa+nbNb Here, Na and Nb represent the number of variations detected in two sequencing datasets, denoted as 'a' and 'b', respectively. na indicates the total number of variations from set 'a' that were also detected in set 'b', and nb signifies the total number of variations from set 'b' that were detected in set 'a'. A 'same variation' refers to a mutation occurring at the same locus with identical pre- and post-mutation sequences, that was, the 'CHROM', 'POS', 'REF', and 'ALT' fields, as well as the genotypes, must match in the VCF file. Scores range 0–1, with 1 indicating perfect concordance. The analysis was performed using custom Python scripts (HCC1395_F1_and_Reproducibility.py and Quartet_F1_and_Reproducibility.py). Supplementary Information Additional file 1: Table S1. Summary of data quality control metrics for the sequencing datasetAdditional file 2: Table S2. Statistical test results for data quality assessment by test type. This table provides information on the results of the normality test, the statistical method employed, sample size, effect size, and the 95% confidence intervalAdditional file 3: Table S3. BamQC analysis results of downsampled WES dataAdditional file 4: Table S4. Evaluation results of WES data performance based on F1 score and Mendelian-concordant rateAdditional file 5: Supplementary figuresAdditional file 6: Table S5. Theoretical cost‑effectiveness analysis of sequencing strategiesAdditional file 7: Table S6. Statistical test types and their results applied to WES dataAdditional file 8: Table S7. Statistical test types and their results applied to RNA‑seq data Additional file 1: Table S1. Summary of data quality control metrics for the sequencing dataset Additional file 2: Table S2. Statistical test results for data quality assessment by test type. This table provides information on the results of the normality test, the statistical method employed, sample size, effect size, and the 95% confidence interval Additional file 3: Table S3. BamQC analysis results of downsampled WES data Additional file 4: Table S4. Evaluation results of WES data performance based on F1 score and Mendelian-concordant rate Additional file 5: Supplementary figures Additional file 6: Table S5. Theoretical cost‑effectiveness analysis of sequencing strategies Additional file 7: Table S6. Statistical test types and their results applied to WES data Additional file 8: Table S7. Statistical test types and their results applied to RNA‑seq data