Specimen quality shapes the actionable genomic landscape in comprehensive cancer genomic profiling

Abstract

Comprehensive genomic profiling (CGP) is widely used to identify actionable alterations and guide precision oncology, yet only a minority of tested patients receive genome-matched therapies, underscoring a gap between genomic findings and clinical benefit. We hypothesized that this gap may partly reflect variation in the reliability and interpretability of genomic information generated from specimens of different quality and by different assay modalities. To examine this possibility, we performed a retrospective multicenter analysis of 2002 CGP tests conducted between 2019 and 2025 across 13 institutions in Japan. Detection of short variants, copy number alterations (CNAs), structural variants, and genomic signatures, including microsatellite instability, tumor mutational burden, and homologous recombination deficiency signature, was compared among FoundationOne CDx specimens classified as pass (F1-pass) or qualified (F1-qual) and liquid-based CGP (liq-CGP). Short variant detection remained largely preserved in F1-qual specimens, whereas CNA and genomic signature detection were substantially reduced. In pancreatic adenocarcinoma, KRAS variants were detected in 93% of F1-pass, 88% of F1-qual, and 57% of liq-CGP cases. These differences affected the proportion of patients offered genome-matched therapies. Machine learning models predicted QC status with area-under-the-curve values of 0.72–0.78. Our findings support QC-aware CGP selection in routine precision oncology.

Introduction

Comprehensive genomic profiling (CGP) represents the clinical implementation of next-generation sequencing, enabling the simultaneous interrogation of hundreds of cancer-related genes to identify genomic alterations relevant to cancer biology and therapy¹. CGP has been rapidly integrated into routine oncology practice, with more than 100,000 patients tested to date in Japan^2–4. CGP generates comprehensive genomic information that forms the basis of precision oncology. Multiple studies have demonstrated improved clinical outcomes among patients who receive genome-matched therapies based on CGP results, underscoring the clinical relevance of this genomic information^5,6.

Despite this promise, real-world studies consistently indicate that only ~10% of patients undergoing CGP ultimately receive genome-matched treatments^5,6. In addition to factors such as the availability and accessibility of genome-matched therapies and patient clinical eligibility, heterogeneity in the content and interpretability of genomic information generated by CGP may further constrain the clinical impact of precision oncology.

In parallel with its clinical adoption, extensive efforts have been devoted to the development and validation of multiple CGP platforms. Under well-controlled, ideal analytical conditions, these assays have demonstrated high concordance and comparable sensitivity across a wide range of genomic alterations, and in some cases have shown complementary or additive value^7,8. However, how such established performance characteristics translate into routine clinical practice, where specimen quality and testing strategies vary widely, remains insufficiently understood.

In real-world oncology care, CGP performance is shaped not only by assay design but also by upstream pre-analytical factors, particularly specimen quality. For tissue-based CGP (tissue-CGP), such factors are partially captured through internal quality control (QC) processes. We previously demonstrated that tumor cellularity is a major determinant of QC status in tissue-CGP, suggesting that specimen quality can substantially influence the genomic information obtained from CGP in routine oncology settings⁹. Importantly, different classes of genomic alterations, such as short variants, copy number alterations, and composite genomic signatures, may exhibit distinct sensitivities to compromised specimen conditions.

Despite the central importance of specimen quality in genomic medicine, the extent to which quality-stratified CGP influences the reliability and interpretability of clinically relevant genomic alterations has not been systematically evaluated using large-scale real-world data. In this study, we propose a real-world, quality-stratified framework for CGP interpretation. By analyzing 2,002 CGP tests performed in routine clinical practice, we examined how pre-analytical specimen quality and assay modality jointly shape the genomic landscape across major cancer types, and discuss the downstream implications for the application of genomic medicine in clinical oncology.

Results

Genomic landscape of FoundationOne CDx with respect to pass and qualified status

Between September 2019 and July 2025, a total of 1623 cases were submitted for FoundationOne CDx (F1CDx), a tumor-only tissue-based CGP assay performed using formalin-fixed paraffin-embedded (FFPE) specimens, at Hiroshima University and its affiliated hospitals. Among these, 1393 cases were classified as F1CDx with pass (F1-pass) and 230 cases as qualified (F1-qual) based on the QC status. The age, sex, and cancer distribution for each group are shown in Table 1. There were significant differences in the distribution of cancer types within each group (p = 0.0006). The genomic landscapes for each QC category are summarized in Fig. 1a. Although the distribution of cancer types differed between F1-pass and F1-qual, the detection rate of CNAs, including deletion of CDKN2A, CDKN2B, and MTAP, as well as amplification of MYC, was significantly lower in F1-qual (Fig. 1b). This finding is consistent with the known limitation that the sensitivity of F1CDx for detecting CNAs decreases under qualified conditions. Conversely, the detection frequency of KRAS alterations was significantly higher in F1-qual, which is likely attributable to the greater proportion of pancreatic cancer within this cohort, for which KRAS variants are more prevalent. In addition, alterations in TP53 and RB1 were more frequently observed in F1-pass; however, it remains unclear whether this difference is primarily driven by the lower sensitivity for detecting CNAs in F1-qual or by differences in cancer type distribution.

Table 1.

Patient demographics stratified by QC status in the FoundationOne CDx cohort

Variable	All (n = 1623)	F1-pass (n = 1393)	F1-qual (n = 230)	P value (F1-pass vs F1-qual)
Age, median (IQR)	66 (55-73)	65 (54-73)	68 (56-74)	0.15
Sex, n (%)				0.25
Female	770 (47.4)	669 (48.0)	101 (43.9)
Male	853 (52.6)	724 (52.0)	129 (56.1)
Cancer type, n (%)				0.0006
Bowel	204 (12.6)	185 (13.3)	19 (8.3)
Pancreas	199 (12.3)	146 (10.5)	53 (23)
Biliary Tract	179 (11)	145 (10.4)	34 (14.8)
Lung	154 (9.5)	134 (9.6)	20 (8.7)
Prostate	94 (5.8)	80 (5.7)	14 (6.1)
Breast	93 (5.7)	84 (6)	9 (3.9)
Ovary/Fallopian Tube	76 (4.7)	69 (5)	7 (3)
Head and Neck	64 (3.9)	58 (4.2)	6 (2.6)
CNS/Brain	59 (3.6)	54 (3.9)	5 (2.2)
Esophagus/Stomach	59 (3.6)	51 (3.7)	8 (3.5)
Uterus	57 (3.5)	55 (3.9)	2 (0.9)
Unknown Primary	54 (3.3)	42 (3)	12 (5.2)
Bladder/Urinary Tract	51 (3.1)	45 (3.2)	6 (2.6)
Soft Tissue	51 (3.1)	44 (3.2)	7 (3)
Liver	42 (2.6)	37 (2.7)	5 (2.2)
Cervix	28 (1.7)	25 (1.8)	3 (1.3)
Skin	26 (1.6)	24 (1.7)	2 (0.9)
Kidney	26 (1.6)	22 (1.6)	4 (1.7)
Thyroid	23 (1.4)	19 (1.4)	4 (1.7)
Thymus	18 (1.1)	15 (1.1)	3 (1.3)
Bone	15 (0.9)	14 (1)	1 (0.4)
Peritoneum	15 (0.9)	13 (0.9)	2 (0.9)
Ampulla of Vater	11 (0.7)	10 (0.7)	1 (0.4)
Peripheral Nervous System	9 (0.6)	8 (0.6)	1 (0.4)
Adrenal Gland	5 (0.3)	5 (0.4)	0 (0)
Testis	4 (0.2)	4 (0.3)	0 (0)
Pleura	3 (0.2)	1 (0.1)	2 (0.9)
Penis	3 (0.2)	3 (0.2)	0 (0)
Eye	1 (0.1)	1 (0.1)	0 (0)

QC Quality control, IQR Interquartile Range, F1-pass FoundationOne CDx-pass, F1-qual FoundationOne CDx-qualified, CNS Central Nervous System. Statistical analysis was performed using the Wilcoxon rank-sum test or the Fisher exact test.

Fig. 1. Comparison of pass (F1-pass) and qualified (F1-qual) specimens for all cases tested with FoundationOne CDx.

a Genomic landscape. b Comparison of variant frequency. Red lines indicate a higher frequency in F1-pass, while blue lines indicate a higher frequency in F1-qual. OR Odds ratio, FDR False discovery rate, CI Confidence interval.

Landscape comparison of pass and qualified status for pancreatic adenocarcinoma

To account for bias in variant frequencies attributable to cancer type and provide a more accurate benchmark of detection rates between F1-pass and F1-qual, pancreatic adenocarcinoma (PDAC) was used as a model cohort (Fig. 2a). Consistent with the overall analysis, the detection rates of CNAs such as CDKN2A, CDKN2B, and MTAP were significantly lower in F1-qual (Fig. 2b). Furthermore, since TP53 and ATM alterations have been reported to be mutually exclusive¹⁰, the slightly higher frequency of ATM alterations observed in F1-qual may explain the correspondingly lower detection rate of TP53 alterations in this cohort. In addition, the detection frequency of RNF43 alterations was also significantly reduced in F1-qual, even though most of these alterations were single-nucleotide variants (SNVs). While differences in tumor purity may account for this reduction, it is also noteworthy that the frequency of RNF43 alterations has been reported to vary by tumor subtype¹¹; however, whether such subtype differences are reflected in our cohort remains unclear.

Fig. 2. Comparison of pass (F1-pass) and qualified (F1-qual) specimens for pancreatic adenocarcinoma tested with FoundationOne CDx.

a Genomic landscape. b Comparison of variant frequency. Red lines indicate a higher frequency in F1-pass. OR Odds ratio, FDR False discovery rate, CI Confidence interval.

Comparison of landscapes for qualified tissue-based CGP versus liquid biopsy

When tissue-CGP cannot be performed due to insufficient tissue volume or limited specimen quality, liquid-based CGP (liq-CGP) is employed as an alternative approach. In this study, liq-CGP included FoundationOne Liquid CDx (F1LCDx) and Guardant360 CDx (G360CDx), both of which analyze plasma-derived circulating tumor DNA (ctDNA). Obtaining sufficient tissue specimens is often difficult¹², particularly in the case of pancreatic cancer. This makes liq-CGP an attractive alternative. To evaluate whether liq-CGP can serve as a reliable substitute, the genomic landscapes of F1-qual and liq-CGP were compared (Fig. 3a). Alterations in CDKN2A and CDKN2B tended to be less frequent in liq-CGP, but the difference was not significant (False discovery rate [FDR] of 0.056 and 0.063, respectively). On the other hand, the detection rate for KRAS variants, most of which are SNVs, was significantly reduced in liq-CGP (Fig. 3b). However, it has been well documented that detection rates in liq-CGP vary by cancer type¹³, with pancreatic cancer showing particularly low sensitivity. Therefore, F1-pass versus F1-qual, as well as F1-qual versus liq-CGP, were also compared across five additional cancer types with relatively large case numbers at our institution (colorectal adenocarcinoma [COAD], biliary tract cancer [BTC], invasive breast carcinoma [BC], non-small cell lung cancer [NSCLC], and prostate adenocarcinoma [PAC]) (Supplementary Figs. 1–5). Except for KRAS in NSCLC (Supplementary Fig. 2b), no significant differences in detection rates were observed between F1-pass and F1-qual in these cohorts. However, this may also be due to the small sample size of qualified cases. Masago et al. have shown that the KRAS variant is present in 70% of cases of invasive mucinous adenocarcinoma, a subtype of lung adenocarcinoma containing mucinous components¹⁴. In our cohort, mucinous components were observed pathologically in one pass case and four qualified cases. This difference in proportion is thought to be related to the difference in KRAS variant frequency. On the other hand, a significant difference in detection rates of CDKN2A/B was observed when comparing F1-qual and liq-CGP for BTC and NSCLC (Supplementary Figs. 2d and 4c), where CNAs are more frequently detected. Interestingly, no such differences were observed for several amplifications that are commonly detected in BC and PAC, although the limited sample sizes warrant cautious interpretation. These findings indicate that the suitability of liq-CGP as an alternative is dependent on the specific cancer type when tissue quantity is insufficient.

Fig. 3. Comparison of FoundationOne CDx-qualified (F1-qual) and liquid-based CGP for pancreatic adenocarcinoma.

a Genomic landscape. b Comparison of variant frequency. Red lines indicate a higher frequency in F1-qual. OR Odds ratio, FDR False discovery rate, CI Confidence interval.

Impact of differences in the genomic landscape on treatment options

To clarify how differences in genomic landscapes influence the presentation of treatment options, we examined the treatments presented to patients based on genomic alterations, including high tumor mutation burden (TMB-high), high microsatellite instability (MSI-High), and alterations in EGFR, BRAF, MET, KRAS, ERBB2, BRCA1/2, PIK3CA/PTEN/AKT1, FGFR2, NTRK1/2/3, and CDKN2A/MTAP. These genomic alterations were selected because they are clinically actionable and commonly considered when presenting treatment options to patients across the six cancer types examined in this study within current precision oncology practice. The results are summarized in Fig. 4. For TMB-high, a significant difference was observed in the ratio of patients for whom treatment options were presented between F1-pass and liq-CGP in NSCLC (FDR = 0.02). In addition, significant differences were identified for KRAS in PDAC and NSCLC, which are likely attributable to differences in the underlying genomic landscapes across these contexts (Figs. 2 and 3, Supplementary Fig. 2). Similarly, for CDKN2A/MTAP variants, which are more frequently detected as CNAs and are highly prevalent in PDAC, NSCLC, and BTC, the observed differences in treatment options also reflect variation in the genomic landscape (Supplementary Figs. 1–5).

Fig. 4. Comparison of genome-matched therapies in pass (F1-pass), qualified (F1-qual), and liquid-based CGP (liq-CGP) for six cancers.

Each bar indicates the proportion of cases in which genomic-matched therapies were offered. * indicates FDR < 0.05. PDAC Pancreatic adenocarcinoma, COAD Colorectal adenocarcinoma, BC Invasive breast carcinoma, PAC Prostate adenocarcinoma, NSCLC Non-small cell lung cancer, BTC Biliary tract cancer.

Prediction of QC status using machine learning

Although liq-CGP may serve as an alternative, depending on cancer type, tissue-CGP remains the preferred approach whenever feasible. However, in cases classified as qualified, the detection rate of certain alterations decreases, highlighting the need for a pre-submission assessment of QC status to minimize the risk of reduced detection sensitivity. To address this issue, machine learning (ML) models were developed to predict QC status. The predictive performance of four models, Gradient Boosting Machine (GBM), Random Forest (RF), Support Vector Machine (SVM), and a Neural Network (NN), was evaluated, and their accuracy and area under the curve (AUC) are summarized in Fig. 5a. Among these, GBM and RF achieved AUC (95% confidence intervals [95%CI]) of 0.777 (0.696–0.843) and 0.757 (0.670–0.829), respectively, demonstrating superior predictive performance compared to the logistic regression model we previously reported⁹. Additionally, the NN model showed a slightly lower AUC (95% CI) of 0.717 (0.620–0.806) but achieved the highest performance in terms of accuracy (95% CI) at 0.871 (0.830–0.903). Receiver operating characteristic (ROC) curves and confusion matrices are shown in Fig. 5b and c.

Fig. 5. Performance evaluation of four machine learning models for predicting pass versus qualified quality control status of a sample.

a Performance summary table. b Receiver operating characteristic curve. c Confusion matrix. 95%CI 95% Confidence Interval, GBM Gradient Boosting Machine, RF Random Forest, SVM Support Vector Machine, NN Neural Network.

Discussion

Our quality-stratified analysis demonstrates that comprehensive genomic profiling does not generate uniform genomic information in routine clinical practice. Rather, the reliability and interpretability of distinct classes of genomic alterations are differentially shaped by pre-analytical specimen quality and assay modality. This alteration-class–specific vulnerability provides a structural explanation for the heterogeneous clinical impact of CGP observed in real-world settings.

Large-scale cohort studies and pooled analyses have compared tissue- and liq-CGP^12,15,16, demonstrating complementary strengths and limitations of each approach. Tissue-CGP generally provides superior detection of CNAs and genomic signatures, while liq-CGP offers higher testing success rates and faster turnaround but lower sensitivity for CNAs and TMB, partly due to tumor-fraction limitations and clonal hematopoiesis^15,16. However, these studies have primarily focused on differences between assay platforms. Less attention is paid to pre-analytical factors that shape assay performance in routine clinical practice, in which tissue quality, particularly tumor cellularity and specimen size, varies substantially and is expected to differentially affect the detection of distinct classes of genomic alterations. Previous analyses have shown that CGP reportability decreases in FFPE specimens with low tumor content or limited tumor area¹⁷. Consistent with these observations, analytical validation studies suggest that CGP sensitivity varies by alteration type and computational tumor purity, with copy number assessment being particularly vulnerable under low tumor cellularity conditions¹⁸. This limitation is also recognized at the assay development level; for example, analytical validation of F1CDx has documented reduced sensitivity for copy number detection under low tumor cellularity conditions.

Genomic signatures such as microsatellite status (MSI), tumor mutational burden (TMB), and homologous recombination deficiency signature (HRDsig), integrate information across multiple genomic loci and therefore depend on adequate tumor representation within the analyzed specimen. Through our quality-stratified analysis, we demonstrated that reliable assessment of these signatures was substantially compromised in F1-qual (Supplementary Fig. 6). This limitation is especially relevant for MSI and TMB, which guide the use of immune checkpoint inhibitors in clinical practice. In qualified specimens, MSI status could not be determined, and some liq-CGP providers report only the presence of MSI-High, making it difficult to distinguish true microsatellite stability from analytically indeterminate results. Interpretation of TMB availability across platforms should also consider assay-specific reporting differences, including the limited provision of TMB results in some liquid-based assays. Previous studies have likewise reported lower concordance or sensitivity for MSI and TMB in liquid-based assays compared with tissue-CGP^15,16 and have emphasized the influence of FFPE quality on TMB assessment¹⁷. When FFPE specimens are anticipated to be of limited quality, complementary approaches such as PCR-based MSI testing or immunohistochemical assessment of mismatch repair proteins may help mitigate these limitations.

We focused on PDAC as a benchmark because KRAS variants are implicated in nearly all cases¹⁹, providing a robust reference for evaluating the analytical sensitivity of short variant detection. Using this framework, we compared quality-stratified tissue-CGP and liq-CGP to assess how specimen quality and assay modality influence short variant detection in routine clinical practice. Across tissue-CGP, the detection frequency of KRAS variants remained largely preserved even in FFPE samples that did not meet standard quality thresholds (Fig. 2), indicating that short variant detection is relatively robust to compromised tissue quality. In contrast, liq-CGP showed a modestly reduced detection rate for KRAS variants (Fig. 3), consistent with a known limitation of liq-CGP in cancer types with low ctDNA¹³.

While PDAC represents a biologically informative scenario in which analytical differences in short variant sensitivity are readily detectable owing to the near-universal presence of KRAS variants, such differences were not associated with meaningful clinical consequences in most other cancer types. In NSCLC and COAD, oncogenic short variants including KRAS, EGFR, BRAF, and APC were reliably detected regardless of specimen quality or assay modality (Supplementary Figs. 1 and 2). Although minor differences in detection frequencies were observed across groups, including instances in which KRAS was detected more frequently in F1-qual specimens (Supplementary Fig. 2), these patterns likely reflect non-overlapping, clinically selected testing cohorts rather than true differences in analytical sensitivity. Taken together, these findings indicate that short variant detection for clinically actionable targets^19–25 is largely robust in routine clinical practice and that modest apparent differences across CGP modalities are unlikely to have substantial clinical consequences in most cancer types.

In contrast to short variants, CNAs showed a marked dependence on specimen quality and assay modality in our quality-stratified analysis. This revealed a quality-dependent asymmetry in the actionable genomic landscape, wherein large-scale deletions and amplifications are underrepresented in suboptimal specimens. Consistent with known limitations of copy number calling under low tumor fraction conditions²⁶, the detection frequency of CNAs was substantially lower in F1-qual specimens and showed a pattern comparable to that observed with liq-CGP within this cohort (Figs. 1–3, Supplementary Figs. 1–5), indicating that copy number assessment is particularly vulnerable under suboptimal specimen conditions and assay types.

Among clinically actionable CNAs, CDKN2A/B/MTAP copy number loss, observed across multiple cancer types including PDAC (Figs. 2 and 3), NSCLC (Supplementary Fig. 2), and BTC (Supplementary Fig. 4), provides a clear illustration of how specimen quality directly shapes clinical interpretability. In our analysis, CDKN2A/B/MTAP loss was detected more frequently in F1-pass than in F1-qual and liq-CGP. This finding suggests a relative underrepresentation of large-scale genomic deletions in quality-compromised specimens. MTAP deletion has emerged as a therapeutically actionable vulnerability and serves as a biomarker for emerging PRMT5 and MAT2A inhibitors²⁷; failure to detect MTAP loss in quality-compromised tissue-CGP or liq-CGP may influence the therapeutic options presented to patients in routine practice, even when oncogenic short variants are successfully identified (Fig. 4). This observation underscores a key distinction between analytical robustness and clinical impact, as loss of CNA information can materially alter the actionable genomic landscape.

Not all clinically actionable CNAs exhibited statistically significant quality-dependent differences in detection frequency in our cohort. In PDAC, BTC, and BC, BRCA1/2 and PTEN deletions and ERBB2 amplifications, both clinically actionable alterations^28–35, were detected at similarly low frequencies among F1-pass, F1-qual, and liq-CGP. This likely reflects the low baseline prevalence of these alterations, such that even under optimal specimen conditions, their detection is infrequent. As a result, specimen quality is a less prominent practical determinant of detectability for these variants in routine clinical practice. However, this should not be interpreted as evidence that copy-number alterations are insensitive to specimen quality, as such alterations remain analytically susceptible to reduced tumor cellularity.

Beyond alteration-specific limitations, liq-CGP yielded a non-negligible proportion of clinically non-informative results in our real-world cohort. Excluding variants attributable to clonal hematopoiesis, no reportable variants were detected in ~20–30% of cases across PDAC, NSCLC, COAD, and BTC, highlighting an important limitation of liq-CGP when performed as a single-opportunity test. This limitation was less pronounced in BC and PAC, underscoring that the clinical utility of liq-CGP varies by cancer type and testing context.

Building on these considerations, uncertainty remains in routine clinical practice regarding whether to proceed with tissue-based CGP or to submit liquid-based CGP, particularly when tissue availability or specimen quality is limited. To address this challenge, we developed ML models to predict whether FFPE specimens would meet quality criteria for tissue-based CGP. These models demonstrated moderate and clinically informative performance, particularly in identifying specimens likely to be classified as qualified (Fig. 5). Importantly, all predictive features are routinely available at the time of pathological evaluation, including tumor nuclei ratio, specimen type, cancer type, and FFPE storage duration. In clinical practice, a high predicted probability of a qualified QC status may prompt consideration of alternative or stratified CGP strategies, such as repeat biopsy, selection of another CGP panel when appropriate, or the use of complementary assays targeting specific genomic features.

This study has several limitations. First, the study was conducted within the Japanese national health insurance system, where indications, timing, and reimbursement of comprehensive genomic profiling differ from those in other countries. Second, CGP is typically performed after standard therapies have been exhausted, and tissue- and liq-CGP were not performed concurrently, precluding direct cross-platform concordance assessment. These system-level factors may influence test selection and patient populations and should be considered when interpreting the absolute frequencies of detected alterations and genome-matched therapies. Third, it remains unclear whether genome-matched therapies, including the clinical trials presented, were applied to patients and influenced outcomes. Fourth, although cancer type was included as a predictor variable in the machine-learning models, robust performance analyses within individual tumor types were limited due to the small number of F1-qual samples in most cancer categories.

In summary, our quality-stratified landscape analysis demonstrates that the interpretability of genomic information generated by CGP is quality-dependent and alteration-class specific. While short variant detection remained largely robust across quality strata, CNA detection and genomic signature assessment were disproportionately vulnerable to compromised specimen conditions, with implications for therapeutic interpretability of genome-matched therapies. By integrating quality-stratified performance analysis with ML-based QC prediction, this study proposes a practical framework to support rational CGP selection in real-world clinical practice.

Methods

Study design

Patients who underwent CGP testing between September 2019 and July 2025 at the following institutions were enrolled: Hiroshima University Hospital, Hiroshima Prefectural Hospital, NHO Kure Medical Center and Chugoku Cancer Center, Hiroshima City North Medical Center Asa Citizens Hospital, JA Onomichi General Hospital, NHO Higashihiroshima Medical Center, JA Hiroshima General Hospital, Hiroshima Red Cross Hospital & Atomic-bomb Survivors Hospital, Miyoshi Central Hospital, Kagawa University Hospital, Takamatsu Red Cross Hospital, Yamaguchi University Hospital, and Shimane University Hospital. All CGP tests approved by the national health insurance system for patients with advanced solid tumors in Japan were included, including F1CDx, a tumor-only tissue-based assay using FFPE tumor specimens, and liquid-based assays, F1LCDx and G360CDx, which analyze plasma-derived ctDNA. For F1CDx, samples were classified as “pass” when predefined analytical quality metrics were met. Samples not meeting one or more thresholds were returned by Foundation Medicine Inc. as “qualified” when the results were still considered reportable. Tumor purity was estimated from sequencing-derived copy-number and SNP allele-frequency profiles relative to a process-matched normal reference, although the exact proprietary method has not been publicly disclosed, as summarized previously⁹. Detection of short variants (including substitutions, insertions, and deletions), CNAs, and structural variants, as well as genomic signatures including MSI, TMB, and HRDsig, was compared with respect to F1-pass, F1-qual, and liq-CGP (including 230 cases of F1LCDx and 149 cases of G360CDx). TMB values were obtained from the original clinical CGP reports, and no independent recalculation was performed in this study. Population allele frequencies were examined using the gnomAD whole-genome sequencing and jMorp/ToMMo databases^36,37. The pathogenicity of variants detected in the CGP tests was assessed using the annotations provided by commercial laboratories and public databases and informatics tools such as ClinVar³⁸, OncoKB³⁹, InterVar⁴⁰, and VarSome⁴¹, as we described previously⁴².

Genomic landscape visualization

The genomic alterations identified through CGP testing were first converted into the ANNOVAR⁴³ input format using an in-house script. Variant annotation was then performed using ANNOVAR, and the annotated output was processed using the maftools v2.8.5 R package to generate visualizations of the genomic landscape, retaining the evaluations described above. The landscape only included variants that had been evaluated as pathogenic by the expert panel at our hospital. TMB-high was defined as ≥10 mut/Mb for tissue-CGP and ≥12 mut/Mb for liq-CGP.

Search of genome-matched therapies

In this study, a genome-matched therapy was defined as any approved drug or investigational therapy supported by a genomic alteration detected by CGP in Japan, based on current clinical evidence and trial eligibility criteria. Based on the detected genomic alterations, potentially applicable clinical trials were identified from the Japan Registry of Clinical Trials, the University Hospital Medical Information Network, and ClinicalTrials.gov, in addition to genome-matched therapies provided by the testing companies, and were presented to patients in routine clinical practice. The presence or absence of a genome-matched therapy was tabulated using Python. Group comparisons were performed using Fisher’s exact test implemented in SciPy v1.15.2, and the results were visualized using Matplotlib.

Quality control status prediction using machine learning models

To predict the quality control status of F1CDx tissue samples, four ML models were constructed: GBM, RF, SVM, and NN. The predictive features included the following: (i) the tumor nuclei ratio in FFPE specimens as assessed by pathologists at Hiroshima University, (ii) cancer type, allowing the models to incorporate tumor-specific differences across the cohort, (iii) specimen type (surgical vs. biopsy), and (iv) FFPE storage duration, which we previously reported to have a strong impact on QC⁹. The GBM, RF, and SVM models were implemented using scikit-learn v1.7.2, while the NN model was constructed with TensorFlow v2.19.0. The dataset was divided into training and test sets using a stratified random sampling approach to preserve the proportion of QC categories. Specifically, 80% of the data were used for training and 20% for testing. The data splitting was performed using the train_test_split function from scikit-learn with stratification on QC labels, and the random state was fixed to ensure reproducibility. The distribution of cancer types in the study cohort is summarized in Table 1. Hyperparameter tuning was performed using Optuna v4.5.0 with K-fold cross-validation for the GBM, RF, and SVM models, and KerasTuner v1.4.7 with the Hyperband algorithm for the NN. The optimal NN configuration consisted of one hidden layer with 96 neurons, a dropout rate of 0.2, and a learning rate of 0.0005. ReLU activation was used in the hidden layer and sigmoid activation in the output layer for binary classification. The model was compiled with the Adam optimizer using binary cross-entropy as the loss function and trained for up to 50 epochs with early stopping (patience = 10). Model performance metrics and their 95% CI were estimated using the bootstrap method. ROC curves and confusion matrices were plotted using Matplotlib v3.10.6.

Statistics

All statistical analyses were performed using R software (version 4.1.0; R Foundation for Statistical Computing, Vienna, Austria). FDR correction for multiple testing was applied using the Benjamini-Hochberg method.

Ethics statement

This study was approved by the Ethical Committee for Epidemiology of Hiroshima University (Approval number: E2025-0135-00). Written informed consent was obtained from all patients. All procedures were conducted in accordance with the Declaration of Helsinki and relevant guidelines and regulations.

Author contributions

Conceptualization: H. Nakahara, H. Niitsu; Data curation: H. Nakahara, H. Niitsu, C.N.H.; Formal analysis: H. Nakahara; Methodology: H. Nakahara, C.N.H.; Investigation: H. Nakahara, H. Niitsu, Resources: M.M., H.M., Visualization: H. Nakahara, H. Niitsu; Project administration: H. Niitsu; Funding acquisition: H. Nakahara, H. Niitsu, T.H.; Writing-original draft: H. Nakahara, H. Niitsu; Writing-review & editing: M.M., H.M.,K.A., C.N.H., T.H.; Supervision: H.M., K.A., C.N.H., T.H.

Data availability

The data of the current study are not openly available due to ethical considerations but are available from the corresponding author upon reasonable request.

Code availability

The underlying prediction model for this study is available in hnakahara/QCpredictor and can be accessed via this link https://github.com/hnakahara/QCpredictor.

Competing interests

The authors declare no competing interests.

Supplementary information

The online version contains supplementary material available at https://doi.org/10.1038/s41525-026-00595-8.