Validation and clinical evaluation of a comprehensive circulating tumor DNA assay for genomic profiling in solid tumors

Abstract

Background

Since 2023, the Department of Pathology at Vienna General Hospital has implemented comprehensive genomic profiling (CGP) of circulating tumor DNA (ctDNA) using the AmoyDx^® Comprehensive Assay. Initially intended for cases where tissue biopsy was unfeasible, this study summarizes our experience including analytical validation, biomarker yield, and retrospective clinical utility of this assay.

Methods

Analytical validation was performed using commercial reference standards with variant allele frequencies between 0% and 5%. Cell-free DNA (cfDNA) was extracted from routine plasma samples and analyzed according to the manufacturer’s protocol and optimized filtering thresholds. A total of 559 samples from 501 patients were included in the biomarker yield analysis. Clinical utility was assessed in a subset of 126 patients. Biomarker annotation was retrospectively performed using the Cancer Genome Interpreter. Clinical data were retrospectively obtained via electronic medical record review.

Results

The assay demonstrated 98% analytical sensitivity for small variants at 0.5% variant allele frequency (VAF), with a limit of detection (95%) at 0.37% VAF and a limit of blank of 0.06% VAF, while demonstrating lower sensitivity for fusions (66% at 0.5% VAF). Mutations corresponding to actionable biomarkers (ASCO Tier 1/A) were identified in 28.6% of samples. CtDNA findings contributed to a documented change in therapy in 11.7% of patients, either by establishing first-line therapy in 6.2% or change in management based on detected actionable alterations in 5.5%. Only 15% of results were explicitly acknowledged in tumor board documentation. Clinical requisitions lacked legible diagnoses in 16% of cases, and an explicit clinical question in 75%.

Conclusion

The assay demonstrates high analytical validity for small variants at VAF > 0.5% and yields actionable biomarkers in a substantial proportion of cases. However, inconsistent clinical documentation and poor integration into care workflows limit the assessment of its clinical impact. We recommend that the implementation of ctDNA-based CGP in public health settings be accompanied by standardized requisition protocols and mandatory clinical data capture.

Supplementary Information

The online version contains supplementary material available at 10.1186/s12885-025-15354-8.

Introduction

Circulating tumor DNA (ctDNA) testing is now a central component of precision oncology, enabling non-invasive tumor genotyping, detection of resistance mutations, and monitoring of treatment response or minimal residual disease [1, 2]. While the clinical utility of ctDNA is well established in select indications—such as EGFR mutation testing in non-small cell lung cancer [3] and ESR1 testing in breast cancer [4] —broader application across tumor types has been driven not only by established indications but also by the advent of large, multi-gene panels [5], such as the FDA-approved FoundationOne Liquid [6] and Guardant360 [7] assays. Such comprehensive assays, often marketed for pan-cancer use, are designed to capture a wide spectrum of genomic alterations and biomarkers, but their widespread adoption has not been met with equivalent body of data regarding their use in real-world environments.

To support clinical implementation, laboratories are required under accreditation standards (e.g., CLIA, CAP, ISO 15189) to demonstrate not only the analytical validity of their assays, but also their clinical validity and utility [8, 9]. In publicly funded institutions, demonstration of clinical utility is also essential for cost-justification. However, establishing clinical utility is inherently challenging in oncology, where the impact of genomic results on management decisions depends on a wide range of contextual clinical factors. The importance of clinical context is also reflected by the practice guidelines for molecular test reporting issued by the Association for Molecular Pathology (AMP), American Society of Clinical Oncology (ASCO), and College of American Pathologists (CAP) [10, 11], and the European Liquid Biopsy Society (ELBS) [12].

While novel approaches such as cfDNA methylation profiling [13] and ultra-sensitive enzymatic mutation detection methods [14] continue to expand the analytical scope of liquid biopsy, next-generation sequencing (NGS) panels and digital PCR assays remain the most widely implemented tools for routine genomic profiling in oncology. Although cfDNA-based genomic profiling is an established technology, most published validations have been conducted in controlled research settings or as part of manufacturer-led analytical studies. The present work addresses a different aspect — the translation of a commercial, broad-panel ctDNA assay (AmoyDx Comprehensive) into a high-throughput clinical pathology laboratory. Our objectives were to (i) verify analytical performance under routine laboratory conditions, (ii) evaluate real-world biomarker detection rates across tumor types, and (iii) assess the practical integration of ctDNA testing into multidisciplinary cancer care.

The present assay demonstrated strong analytical validity and was implemented at our institution in 2023, with consistent detection of actionable genomic alterations across multiple cancer types. However, efforts to assess clinical impact were limited by inconsistent and incomplete clinical documentation, limiting the ability to determine whether and how results influenced clinical decision-making. These findings highlight the need for improved clinical data capture and integration as part of laboratory and hospital workflows to support regulatory requirements, optimize test utility, and ensure adherence to best practice standards in molecular diagnostics.

Materials and methods

Analytical validation design

Analytical validation was performed using reference standards with known allele frequencies to verify sensitivity, precision, trueness, limit of detection (LOD) and limit of blank (LOB) under routine workflow conditions. We used commercially available, orthogonally validated reference standards specifically designed for ctDNA assays with variant allele frequencies between 0.125% and 5% (0710 − 0141, 0710 − 0143, 0710 − 0144, SeraCare, covering a range of hotspot and non-hotspot mutations in 27 genes, as well as TPR::ALK and NCOA4::RET fusions; SID-000144, SensID, covering ESR1 hotspot mutations; HD786, Horizon, covering hotspots in GNA11, AKT1, PIK3CA, EGFR, as well as ROS1::SLC34A2 and CCDC6 fusions). We also used the AmoyDx Comprehensive assay reference control, covering mutations at with frequencies between 5 and 80% in APC, EGFR, KRAS, MET and TP53 (provided with the assay reagents). The SeraSeq standards were tested in three independent replicates for the 0.5% and 0% (wildtype) conditions and in four independent replicates for the 0.125% condition. The Horizon standard was tested in two independent replicates. The SensID ESR1 standards are provided as 5 separate vials, of which vial 1 (L536H and Y537C) and 4 (E380Q, S463P and Y537N) were tested twice, and vial 2 (L536P and Y537S), vial 3(L536R and D538G) and the negative control were tested once, with replicate in an independent run. The reference standard provided with the AmoyDx Comprehensive assay was tested in 10 independent runs. Concordance with tissue genotyping was assessed only in isolated cases where data were available, but a formal paired analysis was beyond the study’s scope.

Clinical testing design

We performed retrospective analysis of clinical ctDNA testing using the AmoyDx comprehensive assay at our institution by reviewing all data regarding processed samples between 01/2023 and 10/2025. CtDNA testing of patient samples was performed on physician request in the course of clinical care. At our institution, test requisition does not require a tumor board recommendation. Each sample was analyzed without replicate sequencing. If a test failed defined QC criteria (see below) and sufficient cfDNA isolate was available, testing was repeated from library preparation. Concordance with tissue genotyping was assessed only in isolated cases where data were available. A subset of cases with validation via matched tissue-based NGS is presented in the results section regarding fusions in clinical samples.

Clinical sample processing

Blood samples submitted to our laboratory were drawn into Roche Cell-Free DNA collection tubes (Roche, 07785666001). Upon arrival at our laboratory, blood was further processed either same-day or in rare cases on the next work day, and at most within 3 days of drawing. Plasma was prepared by two sequential centrifugations (1 × 2000 g for 20 min, 1 × 3200 g for 30 min), and stored at −80 °C until cfDNA isolation. CfDNA was isolated from plasma using EZ1&2 ccfDNA Kit (Qiagen, 9027011) according to manufacturer’s instructions, and stored at −80 °C until library preparation, typically within one week.

Library preparation and sequencing

Up to 30ng cfDNA or 50ul of cfDNA per sample were used as input for library preparation (see Table S2). Library preparation was performed using the AmoyDx Comprehensive Panel (AmoyDx), designed to detect alterations in 128 genes (comprising 110 genes for somatic alterations, and 18 genes for germline pharmacogenomics single nucleotide polymorphisms, SNP) according to the manufacturer’s instructions. In brief, the protocol comprises end-repair, adapter ligation, amplification (14 cycles), and several purification and quality control steps prior to target enrichment via hybrid capture, followed by additional amplification, purification and quality control (see below). Libraries were sequenced on a NextSeq550 or NextSeq500 using a mid-output flow cell and 2 × 150 bp read configuration at a final library concentration of 1.3 pM.

Data processing, run and sample quality control

Data processing from basecall (.bcl) files, including demultiplexing, was performed using the proprietary AmoyDx ANDAS Server (pipeline ADXPAN116 versions 0.2.0 and 0.2.1).

Assessment of run and sample performance involved control of sequencing quality according to thresholds defined by the manufacturer: Phred Q ≥ 30 in > 75% of bases, and > 95% coverage at 1,550× unique (unique molecular identifier, UMI–based) depth. The targeted sequencing depth corresponds to a theoretical limit of detection of 0.19% VAF for hotspot SNV/InDel and 0.52% VAF for non-hotspot SNV/InDel (see variant filtering criteria below). Tests that did not achieve this level of performance were repeated if sufficient input material was available, reported without results (i.e. explicitly as failed), or reported with a comment explaining poor test performance and highlighting the increased risk of false negative results. Tests with poor performance were reported after consideration of risk and benefit of disclosing information, e.g., tests with low uniformity of coverage and detected relevant alterations in a region of high coverage.

Variant quality control and filtering

According to the manufacturer of the assay and analysis pipeline (AmoyDx), variant quality control, flagging and filtering were based on the following criteria: Hotspot mutations: VAF ≥ 0.17%, deduplicated read count (UMI family size) after single-strand base recalibration ≥ 3. Non-hotspot mutations: VAF ≥ 0.30%, deduplicated read count after single-strand base recalibration ≥ 8 and after double-strand base recalibration ≥ 4. Hotspot annotation is performed after variant calling using a proprietary dataset. In the AmoyDx pipeline, the “deduplicated read count” is interpreted as the number of unique molecular identifier (UMI) families supporting a given variant after read collapsing, based on the vendor’s description of the UMI-based workflow and standard practice with UMI-tagged libraries.

During validation, we observed that a relevant fraction of bona fide variants was flagged by the default AmoyDx filters solely due to low allele frequency despite adequate supporting reads. Therefore, we introduced a single modification to the default filtering strategy: variants flagged only for sub-threshold VAF above the limit of blank were retained if all other read-quality criteria were met. This adjustment did not alter minimum read-count requirements, and all subsequently reported oncogenic or likely oncogenic variants fulfilled the above thresholds.

Fusion calls in reference control samples were filtered according to the default AmoyDx criteria (Known fusion driver genes involved: deduplicated fusion read count after single-strand base recalibration ≥ 14. No known fusion driver genes involved: deduplicated fusion read count after single-strand base recalibration ≥ 16). In addition, we performed NCBI BLAST alignment analysis for all detected fusions (Core nucleotide database, Homo sapiens (taxid:9606), megablast for highly similar sequences, accessed between May and November 2025).

The proprietary AmoyDx pipeline does not expose defined strand-bias thresholds, per-base or per-read quality thresholds, nor the parameters used for background error modeling. Therefore, direct assessment of read quality distributions or background noise profiles was not possible. However, the manufacturer’s documentation explicitly notes implemented QC filtering procedures for base, cycle, mapping, and read quality, as well as metrics for sequence, read length uniformity, and strand uniformity bias. In our hands, the observed variant allele frequencies in reference standards closely matched expected values, and only one false positive event in the 0% reference was detected (above the empirically determined limit of blank, 0.06% VAF, above the reference standard’s specified limit of 0.1% VAF, and excluding TP53 p.R175H as specified by the standard’s manufacturer), supporting adequate intrinsic error suppression by the manufacturer’s model.

Potential germline and clonal hematopoiesis-related alterations

Clonal hematopoiesis and germline variants could not be reliably identified because our workflow did not include germline or leucocyte sequencing. For the purpose of the retrospective analysis, we flagged putative CH-related mutations if they occurred in TP53, JAK2, SF3B1, ATM, KRAS, NRAS, GNAS, IDH1, IDH2, JAK1, JAK3, or KIT and were detected with < 10% variant allele frequency, and putative germline-mutations if they occurred in BRCA1, BRCA2, PALB2, MLH1, APC, PTEN, TP53, RET, RB1, TSC2, KIT, or NF1 and were detected with 40–60% or > 90% variant allele frequency.

Biomarker analysis

Retrospective analysis of biomarker yield was performed using only reported test results, alterations that were reported as interpreted as oncogenic or likely oncogenic, and high-confidence fusions (with BLAST alignment and > 10 double-strand recalibrated counts, or orthogonally validated in matched data from tissue-based NGS). Inferred negative biomarker results (“pertinent negatives”, e.g. absence of detectable oncogenic KRAS mutations) were only counted if the median sequencing depth of the respective gene’s target region allowed for a theoretical sensitivity of at least 0.5% variant allele frequency (corresponding to a unique molecular depth of > 800 reads).

Diseases were coded as entities according to MSKCC OncoTree [15] as implemented in Cancer Genome Interpreter [16]. NOTE: Cancer Genome Interpreter is a tool for scientific research, and was only used for retrospective analysis. We do not endorse the use of Cancer Genome Interpreter for medical purposes. Automated assignment of clinical actionability was performed for all clinically reported oncogenic and likely oncogenic variants using Cancer Genome Interpreter (accessed via API on 27/10/2025-4/11/2025), which provides clinical evidence tiers according to the VICC harmonized meta-knowledgebase [17] (based on ASCO/CAP/AMP recommendations [10]). The match between the detected alteration and the corresponding biomarker (e.g., “KRAS G12C” vs. “KRAS oncogenic mutation”) in the CGI database was manually confirmed, but no additional manual curation of biomarker evidence was performed.

Estimation of tumor fraction

Tumor fraction (TFx) is not calculated or exposed by the proprietary AmoyDx pipeline. TFx was estimated retrospectively from somatic variant allele frequencies (VAFs) derived from oncogenic and likely oncogenic small variant calls. Variants were first filtered to exclude those flagged as potential clonal hematopoiesis (CH)-related and potential germline variants, as defined above. For each remaining sample, the estimated tumor fraction was calculated as twice the median observed VAF of retained somatic variants, under the general assumption of heterozygous somatic events in diploid loci. Implausible TFx values (> 1.0) were truncated to 1.0. Samples lacking detectable oncogenic or likely oncogenic variants after filtering were assigned no TFx value.

Clinical reporting

Oncogenicity interpretation and reporting of small variants was performed in accordance with Variant Interpretation for Cancer Consortium Standard Operating Procedure [18] as part of a routine diagnostic workflow. After variant filtering as defined above, all oncogenic/likely oncogenic and variants of unknown significance were reported, either without or with qualifying comments regarding their biological or clinical significance. Variants below the LOD and above the LOB were qualified as low-confidence with a comment recommending repeated testing if clinically relevant. Negative tests achieving the targeted sequencing depth were reported with a cautionary comment explaining that the presence of circulating tumor DNA in the sample could not be verified, and therefore a false negative result could not be excluded. The pharmacogenomics-related germline SNP in 18 genes of the AmoyDx Comprehensive panel were not interpreted or reported.

Handling of potential germline variants and CH-related mutations differed slightly between reporting pathologists – in general, detections with increased likelihood of representing germline variants (e.g., variant allele frequency between 40 and 60% and above 90%) were reported with a comment explaining that germline status cannot be excluded without matched sequencing of normal cells or tissue. Similarly, oncogenic mutations in genes commonly associated with clonal hematopoiesis (e.g., JAK2, TP53, SF3B1) were generally reported with a comment explaining the potential origin in CH and that the tumor origin of these mutations could not be confirmed without matched sequencing of leukocyte DNA.

Fusion calls with known potential therapeutic relevance (e.g., known driver fusion) or > 10 double-strand base recalibrated read counts were always reported, with a comment recommending orthogonal validation. With the exclusion of one RET::KIF5B fusion (which was orthogonally verified), all of these fusions were verified via BLAST alignment. Fusion calls with lower evidence or unknown clinical relevance (abnormal fusion orientation, low double-strand base recalibrated read support, no verification via BLAST) were reported with slight variation among pathologists, reflecting differences in individual reporting practices rather than predefined rules. These results were accompanied by a comment explaining the technical uncertainty of the detection and recommending repeated sampling or orthogonal validation, if clinically relevant.

Ethics and data protection

Approval of the ethical board of the Medical University of Vienna for the retrospective study was granted under 1989/2024 (informed consent was waived for retrospective use of data). All data were handled in accordance with institutional data protection guidelines.

Clinical sample counts

Retrospective evaluation was performed for 607 tests based on 572 samples, processed from 12/2022 (after test implementation) to 10/2025. Analysis of diagnostic yield was restricted to 559 tests with signed-out test results, representing 559 samples of 501 unique patients.

Assessment of clinical utility

Retrospective review of test requisition and electronic medical records in order to assess clinical utility was performed for the first 126 consecutive patients (without further exclusion criteria, from 12/2022 to 12/2023) by manual search of all entries from 1 months prior to sample submission until 2 months after reporting test results. In summary, the following data were collected: presence/legibility of diagnosis and explicit question on test requisition form, reasoning for preferring liquid-based over tissue-based testing, documented acknowledgment of results in any form, in particular in the context of molecular or multidisciplinary tumor board documentation. A change in clinical management was attributed to ctDNA testing if it was documented in the medical record, was plausible based on the reported biomarker(s), and not previously identified by tissue-based testing. In contrast, in cases where tissue-based test results reporting the relevant biomarker(s) were issued prior to the ctDNA results (i.e., in case of parallel testing), the change in management was defined to have occurred due to tissue testing and not due to ctDNA testing.

Statistics

Graphs were generated using R Statistics [19] (4.5.1) and RStudio (2025.09.1) in combination with ggplot2 [20] and complexheatmaps [21] packages. Following R packages and versions were employed (excluding base R packages): viridis_0.6.5, viridisLite_0.4.2, ComplexHeatmap_2.26.0, randomcoloR_1.1.0.1, cowplot_1.2.0, ggrepel_0.9.6, ggplot2_4.0.0, httr_1.4.7 (for accessing Cancer Genome Interpreter via API), data.table_1.17.8, dplyr_1.1.4, stringr_1.5.2, readr_2.1.5.

Sensitivity was calculated by dividing the number of detected mutations per group (reference standard and Hotspot status) by the number of mutations available for detection in the respective group.

The limit of detection for variant calling was estimated by logistic regression modeling of variant detection probability as a function of the reference standard manufacturer’s specified expected variant allele frequency tested across multiple runs. For LOD calculation for small variants, we used the SeraSeq reference control materials (0.5%, 0.125%, 0% i.e. WT). For LOD calculation for fusions, we used the SeraSeq reference and Horizon control materials. Each variant was coded as “detected” or “not detected,” and a binomial logistic regression model was fitted to estimate the VAF corresponding to a 95% or 99% probability of detection (LOD95, LOD99).

The Limit of Blank was calculated from measured variant allele frequencies of specified “absent” mutations in the negative (0% VAF) reference samples (“blank”): mean(blank) + 1.645*SD(blank) [22]. Bias was calculated as the difference between the mean measured VAF and the expected VAF. Precision was calculated as the mean deviation of individual allele frequencies from the mean allele frequency, divided by mean allele frequency.

Formal statistical testing of significance was only performed in the course of Spearman’s test for correlation between concentration of cfDNA isolate, or amount of cfDNA input, and UMI-deduplicated depth after sequencing.

Editorial assistance

Language refinement and structural editing were supported by OpenAI’s ChatGPT (GPT-5) under the supervision of the corresponding author. The tool was used exclusively to improve readability and coherence; all analyses, interpretations, and conclusions were produced and verified by the authors.

Results

Assay performance characterization

Analytical validation of the assay, including the proprietary pipeline for variant calling and filtering, was performed using commercially available reference standards with known alterations and allele frequencies (VAFs), hereby characterizing sensitivity, precision, trueness, and limit of blank (LOB) under routine workflow conditions and across independent preparation and sequencing runs (Fig. 1A, Tables S1A-S1B).

Fig. 1.

Analytical validation of the AmoyDx ctDNA assay using commercial reference standards. A SeraSeq reference standards with known variant allele frequencies (VAF) at 0.5%, 0.125% and 0% (negative control) were tested in replicates across independent sequencing runs. Mean detected VAF is displayed as a solid line, while limit of Blank (LOB) is shown as a dashed line. The default variant QC filtering settings of the proprietary pipeline discarded a relevant portion of bona fide variants exclusively due to low VAF – despite achieving VAF and read counts above the manufacturer’s specified thresholds. Variants that passed the default filtering criteria are shown in red, those that were excluded by the default filter due to low VAF are shown in green, and those that outright failed QC due to other factors than low VAF (e.g., low read counts, bad base quality, strand bias, etc.) are shown in blue. We introduced a single modification to the default filtering strategy and retained variants above the LOB that were discarded by the default filter solely due to low VAF. Variants that were retained after our customization are thus shown as red and green points, and translate to a LOD95 of 0.37% VAF. B The AmoyDx kit positive control confirms high precision and trueness across a broad range of VAFs (5%–80%), but this reference standard does not represent a clinically relevant range for ctDNA testing. C For analytical validation of fusion detection, we assessed test performance in 5 different reference standards. All expected and unexpected detections across all standards are shown. Sensitivity in the 0.5% SeraSeq reference was 66%, whereas no events were detected at 0.125% (filtered by QC due to low read counts). Fusions at higher allele frequencies in the Horizon reference and the AmoyDx kit reference were detected with 100% sensitivity, translating to an LOD95 of 3.97% VAF. Recurrent unexpected fusions (in particular, FGFR3::LTBP1) lacking confirmation of sequence identify via BLAST suggested potential artifacts from the bioinformatics pipeline

We first analyzed SeraSeq^® cfDNA reference standards provided at of 0.5%, 0.125%, and 0% VAF (negative control). Unexpectedly, the default AmoyDx variant filtering pipeline discarded many true variants by flagging them for low-frequency artefacts according to preset thresholds (Fig. 1A). This initially reduced analytical sensitivity (80% at 0.5% VAF), highlighting a critical limitation in the default pipeline. We therefore optimized filtering by including variants that were flagged by the AmoyDx pipeline exclusively due to low allele frequency. Differences in analytical performance between the original AmoyDx filtering and our customised implementation are summarized in Table S1C. After optimization of the filtering criteria as above, at 0.5% VAF, the assay demonstrated analytical sensitivity of 98.3% for hotspot mutations and 100% for non-hotspot mutations. At 0.125% VAF, sensitivity decreased to 73.8% and 70.1%, respectively, corresponding to an LOD95 at 0.37% VAF and LOD99 at 0.53% VAF.

Variants were detected with mean allele frequencies of 0.53% and 0.15%, indicating high trueness, with a bias of ± 0.03% VAF and absolute precision of 26.1% and 44.0% (corresponding to absolute VAF of 0.06 and 0.14%, respectively). Using negative controls (0% VAF), we established the limit of blank (LOB) at a VAF of 0.06%. We observed rare variant calls above the LOB in the negative control, but only one of them outside of the reference standard’s range of expected frequencies (> 0.1% VAF, and excluding TP53 p.R175H) (Fig. 1A).

Further validation using reference standards designed to test ESR1 mutations (SensiCare, 1% VAF) and a separate 5% VAF standard (Horizon) demonstrated 100% sensitivity and specificity (Fig. S1A-B). Notably, the AmoyDx kit’s positive control includes variants and fusions at VAFs ranging from 5% to 80%, which, although confirming high assay sensitivity (100%) and precision (9.5%–1.8%, corresponding to absolute VAF of 0.51% and 1.41%) in this range (Fig. 1B), is inadequate for validating performance at lower, clinically relevant frequencies.

We next evaluated fusion detection (Fig. 1C). Of the two fusions present in the 0.5% VAF cfDNA reference, one was consistently detected across all three runs, while the other was detected in only one, yielding 66% sensitivity. No fusions were detected at 0.125% VAF. In contrast, all fusions in the 5% reference standard and kit positive control were detected at expected frequencies (3.3%–20%), corresponding to an LOD95 of 3.97%.

We also observed recurrent detections of unexpected fusions involving FGFR3, LTBP1, NTRK1, and NTRK3 across all standards (12 detections across 9 independent runs and 5 samples), suggesting potential false positives (Fig. 1C). Median sequencing depth of the unexpected fusion sequences was 1401, with allele frequencies between 0.21% and 5.9%, comparable to the expected fusions. However, only expected fusions were detected with ≥ 10 double-strand base recalibrated reads (Fig. S1C). NCBI BLAST analysis revealed that expected fusions from commercial controls aligned to both gene partners on standard RefSeq chromosomes. In contrast, potential false positive detections, but also fusions specified in the AmoyDx kit positive control, did not align, implicating the AmoyDx pipeline in the generation of spurious fusion calls and mandating caution when interpreting findings.

Technical performance of the assay in clinical samples

A total of 572 cfDNA samples were submitted for analysis. 35 samples were re-tested from available cfDNA isolates due to suboptimal sequencing depth (target threshold: unique molecular depth exceeding 1550× in > 95% of target regions), resulting in a total of 607 tests.

The median cfDNA isolate concentration per test was 1.34 ng/µL, with a range of 0.12 to 56.4 ng/uL (Fig. 2A). Consequently, the required minimum and maximum input cfDNA amount specified by the manufacturer (10ng and 30ng) were thus achieved in 99,8% and 86% of tests, respectively. While sequencing quality and overall target region coverage thresholds were met in virtually all tests (100% and 99.7%, respectively), the targeted sequencing depth was met in only 414 of 607 tests (68.2%). UMI-corrected sequencing depth was directly correlated to the amount of input cfDNA and the cfDNA extract concentration (Spearman’s ρ = 0.36, and 0.35, respectively, both p < 0.0001; Fig. 2B) The targeted depth was achieved in 74.8% of samples above 25 ng input (corresponding to isolate concentrations of 0.5ng/µl), but only 9.8% of those below this limit. Notably, no sample with less than 12.5ng input (0.25ng/µl) achieved the targeted sequencing depth (Fig. 2C).

Fig. 2.

cfDNA input and test performance characteristics across clinical samples. A Distribution of cfDNA isolate concentrations among 572 clinical samples. Note that the manufacturer’s recommendation for minimum (10ng) and maximum (30ng) cfDNA input amount was achieved in 99,8% and 86 of samples, respectively. B Correlation between cfDNA extract concentration and mean unique (UMI-corrected) sequencing depth. The cfDNA extract concentration is shown instead of the actual input amount in nanogram due to the continuous distribution of measurements and better viewing of data points. C Relationship between cfDNA isolate concentration and test performance; the targeted sequencing depth as specified by the manufacturer is 1550× across > 95% of target bases. The percentage contribution of each bin (concentration vs. test performance) to the total number of tests is shown. Note the considerable reduction in test success rate with less than 25ng cfDNA input (< 0.5ng/ul extract concentration)

Of the 572 clinical samples tested, 559 final reports were issued (97.7%). Among these, 71.9% met the required sequencing depth of > 1550x in > 95% of target bases, while the rest were reported with a comment regarding poor test performance – e.g., in light of confident and pertinent positive findings. The driving factor behind poor test performance in issued tests was non-uniform coverage: 97% of issued tests had sufficient median depth to achieve a theoretical LOD of 0.5% VAF in mutational hotspots.

We retrospectively estimated the tumor fraction (TFx) of samples from detected oncogenic and likely oncogenic alterations, discounting mutations flagged as putative CH-related or germline variants. Within the 559 samples with issued reports, TFx could be estimated for 316 (56.4%), with a median of 8.5% VAF. A table detailing all test metrics is provided as Table S2.

Mutation landscape

A total of 559 reports were finalized, corresponding to 501 unique patients, of whom 51 contributed multiple samples at different timepoints (Table S3A). The patient cohort ranged in age from 25 to 94 years (median 67), with female patients accounting for 43.1% of the population. The most common tumor types were prostatic adenocarcinoma, breast cancer, non-small cell lung cancer (NSCLC), and colorectal adenocarcinoma.

Oncogenic mutations were identified in 407 of 559 samples (72.8%), with a mean of 1.7 mutations per sample (median 1; maximum 20), and were distributed across 63 of the 110 targeted genes, most frequently involving TP53, PIK3CA, KRAS, ESR1, and AR (Fig. 3A–B, Table S3B). Variants of unknown significance (VUS) were detected in 374 samples (66.9%), with a mean of 1.6 variants per sample (median 1; maximum 21).

Fig. 3.

Mutation landscape of tested patient cohort. A OncoPrint showing the frequency of detected oncogenic mutations per gene, and distribution across the entire sample cohort. Note the greater apparent tumor mutational burden in cases with TP53 mutations (total number of mutations per sample is shown in the topmost row, above the disease annotation bar). No apparent clustering of diseases is seen, but mutual exclusivity of certain driver genes can be appreciated. B Distribution of oncogenic mutations across samples. Only the top 20 most frequent genes with oncogenic mutations are shown. Note that these data include mutations that may be related to clonal hematopoiesis, or represent germline variants

To estimate the potential contribution of mutations arising in clonal hematopoiesis (CH), we annotated detected variants in recurrent CHIP genes (TP53, JAK2, SF3B1, ATM, KRAS, NRAS, GNAS, IDH1, IDH2, JAK1, JAK3, KIT) (Fig. S3A). Variants in these genes detected at a VAF < 10% were flagged as ‘CH-suspect’, accounting for 28% of all detected oncogenic or likely oncogenic mutations, and present in 34% of issued reports. 16.2% of cases with oncogenic/likely oncogenic mutations showed exclusively CH-flagged mutations. Because matched leukocyte DNA was not investigated, this represents a conservative annotation rather than definitive exclusion. Similarly, to estimate potential germline variants, we annotated variants with detected allele frequences between 40 and 60% and > 90% in genes associated with hereditary cancer syndromes (Fig. S3B), accounting for 6.4% of all detected oncogenic or likely oncogenic mutations. Relative to the total number of mutations per gene, putative germline mutations were most frequently detected in PALB2(3/3), BRCA1 (9/13), BRCA2 (11/22), RET (2/5) and MLH1 (1/3).

Overall, 245 putative fusion sequences passing the default AmoyDx filtering criteria were detected in 194 samples (Fig. S3C-D, Table S3C). A highly recurrent detection (FGFR3::LTBP1) was clearly identified as an artefact due to lack of verification via BLAST, and often derived from a sequence involving homopolymers (or even without reporting a consensus fusion sequence). Further 47 detections involved sequences supporting abnormal fusion orientation (e.g. 5’ to 5’ fusions). Of the remaining 55 putative fusion events, only 8 were supported by ≥10 double-strand base recalibrated reads, of which all but one verified with BLAST alignment. Matched tissue-based NGS data was available for 4 of these 8 high-confidence fusions; three could be verified, while in one case tissue analysis yielded a fusion involving only one of the partners reported from cfDNA.

Biomarker yield

To assess the potential clinical relevance of test results, we performed retrospective disease-specific biomarker annotation using the Cancer Genome Interpreter (Fig. 4A, Table S4A-C). Regarding only positive findings, biomarkers with recognized clinical evidence (any tier) were identified in 373 samples (66.7%), while high-evidence, Tier A biomarkers were found in 160 samples (28.6%). When discounting detections previously flagged as putative CH-related mutations or germline variants, we counted overall biomarker detection in 48.1% of samples, and Tier A biomarkers in 20.4%, delineating a considerable portion of mutations with uncertain actionability due to unknown mutational origin.

Fig. 4.

Biomarker yield and clinical relevance. A Frequency of clinically actionable biomarkers, displayed as highest tier per sample. The bars with an asterisk (*) indicate the distribution of biomarker tiers when inferred negative biomarkers (e.g., absence of mutations in KRAS, EGFR, NRAS) were discounted from the analysis. The same color legend from this graph (4A) is used for all graphs in this figure. Only positive findings (i.e. excluding inferred negative biomarkers) were used for the following graphs, while the negative findings are specifically detailed in figure 4E. B Highest biomarker evidence tier per sample, grouped by the respective cancer diagnosis. Note the considerable number of samples from patient with prostate adenocarcinoma due to local practice of clinicians at our institution. C Same data as in figure 4B, but with each cancer type scaled to the same size in order to better appreciate differences in relative abundance of biomarkers tiers. D Distribution of biomarker assertations across samples. In case of multiple assertations per gene, only the highest evidence for each gene was counted per sample. E Distribution of inferred pertinent negative findings. The analysis was restricted to patients who had no other (positive) Tier A findings, thus presenting patients where negative findings are likely to have an impact. Only samples with a theoretical LOD of <0.5% VAF for the respective genes are included. F 51 patients in our cohort contributed more than one sample within the time of retrospective analysis. The time interval in days between serial samples is shown (in case of more than two samples, the maximum time difference). G Changes in biomarker status (more biomarkers of the same Tier, or changes in biomarker Tier) were observed in 26 patients, of which 20 showed an increase in biomarker yield at least once in a sequential test

We verified that the inclusion of patients with multiple samples did not significantly skew overall biomarker yield. When restricting analysis to the first sample from each patient, the findings remained consistent: 71.9% of samples contained oncogenic mutations and 29.7% harbored Tier A biomarkers.

When stratified by tumor type, breast cancer, lung cancer, and colorectal cancer represented the largest disease groups with high rates of Tier A biomarker assertions (Fig. 4B–C). This distribution reflects both the relative frequency of these cancers in the cohort and their established molecular targets. In contrast, although prostate cancer was the most common single tumor type, only 15% of cases demonstrated Tier A biomarkers. By comparison, approximately 50% of cases with NSCLC, colorectal adenocarcinoma, thymic carcinoma, and breast cancer showed high-evidence biomarkers.

In 113 samples (20.2% of the total cohort), Tier A biomarker status could be inferred based only on pertinent negative findings (i.e., wildtype status for KRAS, NRAS, or EGFR). These cases showed sufficient technical performance to allow a theoretical lower detection limit of < = 0.5% VAF. However, after discounting putative CH-related and germline variants, only 46 samples (40%) had an estimated tumor fraction (TF) > 1%, underlining the caveat of overstating negative calls in lieu of low tumor ctDNA content. Among 186 samples that yielded neither positive nor inferred negative biomarkers, only 23 (12.4%) had an estimated TFx > 1%.

Regarding only positive findings (i.e., excluding inferred negatives), biomarker assertions were most frequently associated with detected mutations in PIK3CA, KRAS, ESR1, PTEN, NF1 (Tier A biomarkers: ESR1, KRAS, PIK3CA, BRCA2, EGFR) (Fig. 4D). For comparison, the absolute number of inferred negative findings is shown in Fig. 4E. Across the panel, clinically actionable biomarkers—including both positive findings and pertinent negatives—were observed in only 43 of the 110 targeted genes (Table S4B), and only 16 genes were associated with Tier A biomarkers.

Among the 51 patients with multiple tests, 45 submitted two samples, five submitted three, and one submitted four. The time interval between samples ranged from 4 to 1016 days (median 294; mean 383) (Fig. 4F). Differences in positive (excluding inferred negative) biomarker findings were noted in 26 patients (51%, Fig. 4G), of whom 20 had a greater number of biomarkers and/or a higher tier of biomarker level in a subsequent test.

Prostatic adenocarcinoma

Prostatic adenocarcinoma represented the most frequently tested cancer type in our cohort (Fig. 4D), accounting for 195 of the 559 samples (34.9%), which prompted us to characterize this group of patients separately. At our center, these patients predominantly represent cases of metastatic castration-resistant prostate cancer (mCRPC), typically following multiple lines of systemic therapy and often with limited tissue availability due to bone metastases.

Genomic alterations were most frequently identified in TP53, AR, PIK3CA, PTEN and CDK12 (Fig. 5A). Tier A biomarkers were identified in 15% of prostate cancer cases, the majority of which involved genes associated with homologous recombination repair (HRR), including CDK12, BRCA2, and BRCA1 (Fig. 5B). Notably, all observed CDK12 alterations were frameshift mutations, with more than half located in exon 1 (transcript: NM_016507.4), suggesting a potential regional mutational hotspot that may warrant further investigation.

Fig. 5.

Molecular landscape of prostate cancer cases. A Distribution of oncogenic mutations across 195 samples. Note that these data include mutations that may be related to clonal hematopoiesis, or represent germline variants. B Distribution of biomarker assertations across samples. In case of multiple assertations per gene, only the highest evidence for each gene was counted per sample

Retrospective assessment of clinical utility

To evaluate the clinical implementation and utility of broad-panel ctDNA testing in our institutional setting where test requisition does not yet formally require a tumor board recommendation, we conducted a retrospective review of electronic health records for the first 126 consecutively tested patients. We assumed that a documented change in management (e.g., initiation of tyrosine kinase therapy in EGFR-mutated lung cancer, or PARP inhibitor therapy in BRCA-mutated prostate cancer) might be a readily available surrogate for clinical utility. Additionally, we considered documented discussion in tumor boards or documented acknowledgement as additional surrogates in case we could not identify patients with confirmed management change. Since the clinical utility of a ctDNA test is partly dependent on correctly interpreting the clinical context at time of testing, we also assessed the quality of clinical information provided with the test requisition.

A cancer diagnosis was clearly stated on 84% of test requisitions; in the remaining 16%, it was either absent or difficult to interpret (Fig. 6A). Only 24% of cases included a specific clinical question or suspected resistance mechanism to guide molecular analysis (Fig. 6B). Upon reviewing records from two months prior to and following sample submission, the rationale for ctDNA testing could not be confidently determined in 60% of cases. When present, the most frequent documented indications were previous insufficient tissue biopsy or anatomical challenges precluding re-biopsy (Fig. 6C). In 3% of cases, a formal tumor board recommendation was documented, albeit without an explicit rationale.

Fig. 6.

Retrospective assessment of clinical utility. 126 consecutive cases we evaluated according to surrogate endpoints for clinical impact, from one month prior to two months after reporting results. Whereas (A) a cancer diagnosis was generally present on requisition forms, (B) specific diagnostic questions were generally absent. C Retrospectively, a documented reason for choosing ctDNA over tissue testing could be identified only in a small portion of cases (Re-bx: in order to avoid a repeat biopsy after insufficient tissue; Anat: Patient anatomy/tumor location did not permit biopsy; Pref: Patient chose not to undergo tissue biopsy; TB: documented recommendation in tumor board discussion, but no explicit reason; Med: poor medical condition of patient precluding a tissue biopsy that would be otherwise anatomically feasible). D Proportion of cases with documented acknowledgment of results – any form of written documentation was counted. E Frequency of test result discussion in the context of a tumor board. One potential confounder for the absence of documented test discussion could be non-documented discussion of negative test reports (which may have lower perceived importance). F Proportion of cases with ctDNA-influenced therapy change. In a number of patients, first-line therapy was initiated after ctDNA testing, which qualified as a clinical decision enabled by the test. In a similar number of patients, a previously established therapy was modified according to the test results

Despite the assay’s potential to guide management decisions, structured acknowledgment of test results was identified in only 40% of cases (Fig. 6D), and discussion during a multidisciplinary tumor board in 15% (Fig. 6E). Institutional variability was notable, with traceable indications ranging from 30% to 80% across departments (data not shown).

Clinical outcome data were also inconsistently recorded. ECOG performance status within one week of sampling was documented in only 40 patients (range 0–2, median 0). A change in patient management plausibly attributable to ctDNA results was identified in 11.7% of patients, with 5.5% receiving therapy targeting an identified actionable alteration and 6.2% undergoing guideline-concordant initiation of first-line treatment based on the ctDNA molecular profile (Fig. 6F).

Discussion

Circulating tumor DNA analysis has emerged as a powerful diagnostic modality in oncology, offering minimally invasive, repeatable access to tumor genomic information. Its applications span from mutation detection and therapy selection to real-time monitoring of resistance evolution and minimal residual disease. In the context of solid tumors—particularly those with inaccessible or insufficient tissue—ctDNA profiling enables genomic insights that might otherwise be unavailable. However, the integration of this technology into clinical decision-making remains contingent not only on assay sensitivity and specificity, but on its thoughtful use within a clinically informed framework.

Although the AmoyDx ctDNA panel is manufacturer-validated, institutional verification was required to confirm assay robustness, reproducibility, and reportable range within our laboratory setting, including assessment of sensitivity across multiple reference materials and evaluation of variant filtering criteria for low allele frequencies. While several commercially available ctDNA assays such as FoundationOne Liquid CDx [6] and Guardant360 [7] have demonstrated broad analytical validity and regulatory approval, our study differs primarily in its real-world scope and methodological transparency rather than in the underlying chemistry.

In contrast to vendor-led validations, we report performance metrics derived from routine diagnostic testing in an unselected hospital cohort, encompassing diverse tumor entities and pre-analytical variability typical of clinical care. We further estimated the contribution of clonal hematopoiesis and germline variants to apparent biomarker yield, providing an explicit measure of potential confounding not typically quantified in commercial studies. Fusion detection criteria were refined beyond default pipeline thresholds by integrating double-strand recalibrated read support and manual BLAST confirmation to both partner genes, thereby reducing false-positive fusion calls observed in reference standards. Finally, we complemented analytical validation with a pragmatic assessment of clinical implementation, quantifying requisition quality, tumor-board acknowledgment, and management changes attributable to ctDNA results. Together, these elements provide a comprehensive and transparent view of assay performance and integration that complements prior large-scale validations of commercial ctDNA platforms.

The AmoyDx Comprehensive assay used in this study demonstrated high analytical reliability for variants at 0.5% VAF (LOD95 0.34%), but suboptimal sensitivity for variants below these limits and fusion events below 5% VAF. These findings are aligned with a growing body of evidence supporting the sensitivity and specificity of ctDNA assays, especially in advanced or metastatic disease. Previous studies detailing the validation of commercial assays have described comparable analytical limits of small variant detection (0.37–0.4% VAF, FoundationOne Liquid CDx [6]; 0.3–0.4% VAF, Guardant360 [7]). Notably, the abovementioned assays estimate the tumor fraction and copy number alterations, which gives them an advantage over the assay used in this study. Since we did not calculate the tumor fraction at time of reporting, negative results were reported while uncertain of the quantity or presence of tumor cfDNA in the sample, severely impacting the confidence of negative findings and inferred biomarkers.

In our study, the target threshold for sequencing depth (>95% coverage at >1550x unique depth) was only achieved in 68.2% of performed tests, most often due to low uniformity of coverage. This may be due to insufficient recovery of input molecules during library preparation, as demonstrated by frequently poor test performance despite adequate cfDNA input. Indeed, efficient recovery of rare DNA fragments from the initial sample has been argued to be one of the most critical steps in ctDNA analysis sensitivity [23]. On this note, our findings suggest that the lower recommended cfDNA input for this assay may be higher than specified by the manufacturer (10ng); in our experience, less than 25ng were consistently associated with poor performance.

Previous prospective and retrospective studies have come to heterogeneous results when assessing biomarker yield and clinical utility of ctDNA testing. The prospective PRISM study using FoundationOne Liquid CDx (324-gene panel, >1000 patients) showed 94% assay success, 64% actionable alterations, and 21% receiving matched therapy [5]. In regards to retrospective data, a real-world analysis of over 3,000 patients with advanced NSCLC tested with Guardant360 showed actionable alterations in 41.9% of patients, and significantly improved survival with guided therapy (OS: 36.1 vs. 16.6 months; p< 0.001) [24]. Previous studies with smaller cohort sizes showed actionable alterations in only 20% (>900 patients) [25] or even merely 8% (199 patients) [26]. Much of these differences may be attributable to the inherent bias of “real-world evidence” studies [27], breadth and sensitivity of specific ctDNA assays, diversity in patient cohorts and even definition of target actionability.

In our study, the lack of matched blood leukocyte sequencing implies likely confounding of results by mutations arising in clonal hematopoiesis (CH) or germline variants, which may inflate our results with apparent biomarkers that are not truly actionable. The genes that are most commonly affected in CH and tested in our assay are TP53 and SF3B1. Although a considerable portion of mutations in these genes in our study is most likely related to CH, they have a limited impact on biomarker analysis regarding Tier A biomarkers. Germline variants have substantial medical and ethical implications for the patient and their relatives, but their correct identification is also particularly relevant for their approved actionability in specific scenarios, such as BRCA1/2 mutations in the setting of breast cancer [28]. Together with the lack of reported TFx in the present assay, these limitations clearly highlight the advantages of combined blood leukocyte DNA and cfDNA sequencing.

Compared to small variants, the assay demonstrated relatively low sensitivity for fusions, implying that some actionable fusions may not have been identified. In 118 cases of non-small cell lung cancer, we detected three tissue-validated fusions and one with mismatch of one fusion partner between plasma and tissue - out of four cases where matched data was interrogated. While this does not represent systematic orthogonal validation, the data support confidence in fusion events detected with ≥10 double-strand recalibrated read counts. In contrast, frequent and recurrent fusion calls (likely representing false positive detections) observed with this assay may stem from mapping artifacts, or the use of non-standard reference sequences in the proprietary pipeline, and were derived primarily from lower-evidence (single-strand recalibrated) reads.

In our setting, we detected known driver mutations and other variants with established therapeutic, prognostic, or investigational relevance in a significant subset of patients. However, much of the panel content remained underutilized, with oncogenic mutations detected in only 63 of 110 genes covered by the panel, and actionable alterations (as defined via CAP/ASCO/AMP Tier A-D) detected in only 16 (tier A) or 43 (tiers A-D) genes, respectively. Furthermore, it is plausible that many of these matching biomarkers were clinically irrelevant at time of testing due to advanced line of therapy. Although the unique composition of our cohort – with a high number of prostate cancer cases, which are expected to harbor intermediate or low tumor mutational burden [29] – may contribute to our apparent underutilisation, the present data suggest that optimization of panel content based on specific use-cases may improve cost-effectiveness relationship.

In the context of metastatic castration-resistant prostate cancer (mCRPC), where tissue biopsies are often infeasible or low-yield, and tumor evolution occurs rapidly under therapeutic pressure, ctDNA provides a non-invasive and dynamic window into tumor genomics [30]. We identified CDK12 mutations, particularly in Exon 1, as the most frequent HRR-associated alteration in prostate cancer patients. However, the predictive value of this finding in the context of PARP-inhibitor therapy remains ambiguous; while the FDA has approved PARP inhibitors for CDK12-mutated prostate cancer, the EMA has not. A recent meta-analysis suggests a benefit of PARP inhibition in patients with CDK12 mutations, albeit less clear than in those with mutations in BRCA1/2. Moreover, different CDK12 mutations likely have variable impact and need to be studied in greater detail [31].

Our data emphasizes the role of repeated sequential testing of cancer patients using liquid biopsy in the routine setting. In approximately half of patients (51%), we identified a difference in biomarker profile between the first and a subsequent test. In the context of liquid biopsy, technical and pre-analytical considerations (e.g., variable concentrations of circulating tumor DNA) need to be taken into consideration alongside potential changes in tumor biology. Nevertheless, in approximately three quarters of patients with differences between first and subsequent testing, we identified more or a higher tier of biomarkers in a subsequent test, indicating the potential clinical benefit of repeated testing.

In contrast to the analytical strengths of ctDNA testing, our findings highlight a central challenge in the clinical implementation of ctDNA profiling: the limited availability of structured clinical information to accompany molecular diagnostics. We attempted to characterize the clinical utility of a broad-panel ctDNA assay outside of a controlled study setting - since acquisition of high-quality end-points such as clinical outcome data not feasible, we hypothesized that a documented change in management, discussion in tumor boards, or acknowledgement of results may serve as surrogates. In addition, the quality of information provided at test requisition is not only immediately relevant to the laboratory, but may reflect the perceived importance or complexity of the test to clinicians. Although we retrospectively identified cases where testing impacted clinical management, the larger picture highlights inconsistent documentation as an important limiting factor in our study. This may be expected outside the strictly-controlled environment of a prospective study, yet the present data beckons to reconsider established practice at our institution in light of potential long-term losses and disadvantages.

In many of our cases, our ctDNA request forms lacked central clinical details—such as the cancer diagnosis, disease stage, prior systemic therapies, or the clinical indication for testing. From the perspective of the diagnostic laboratory, contextual details such as current disease status, prior lines of therapy, suspected resistance mechanisms, and treatment goals are essential for meaningful interpretation of ctDNA results [12]. Without this information, even robust molecular findings may be misinterpreted, undervalued, or overlooked. Furthermore, accredited molecular diagnostics laboratories must demonstrate not only analytical validity but also appropriate clinical utility of tests within a defined use-case. When ctDNA results are returned without adequate clinical context, laboratories face challenges in meeting the interpretive and documentation requirements of regulatory and accrediting bodies (e.g., CLIA, CAP, ISO 15189).

Beyond immediate clinical use, ctDNA results hold significant value for retrospective biomarker analysis and real-world evidence generation. Yet in our study, the poor quality of accompanying clinical data severely limited these opportunities, greatly reducing the research utility of both current and archived samples, and decreasing the return on investment for institutions and funders alike. These findings should not be interpreted as a lack of clinical interest; rather, they reflect the reality of limited time, documentation burden, and the absence of structured fields in electronic health records to capture this information.

At our institution, the present assay has been provided at clinician’s request since 2023. Initially, this examination was intended to provide diagnostic support to select patients where a biopsy was not feasible, with select guideline-based indications (e.g. ESR1) accumulating over time – however, this practice has become essentially decoupled from structured MTB workflows. Even if a mandatory recommendation by a tumor board were formally instated, our established hospital requisition workflows could not prevent clinicians from requesting tests without prior approval, and our institute would not be able to confirm or refute requests without impractical effort. One possible explanation for not awaiting tumor board recommendation before testing is that the clinical workflow would require an additional patient visit after the tumor board’s approval, thus extending time to results. If this were a driving issue, samples might be collected and stored temporarily (either frozen plasma or cfDNA) pending tumor board recommendation.

Whether a patient’s test result is discussed in a molecular tumor board varies at our hospital; well-established therapeutic indications (e.g. ESR1 mutations in breast cancer, or BRCA1/2 mutations in prostate cancer) in our experience are often not discussed in a tumor board, at least not with any formal documentation. More complex cases (e.g. later-line therapy, compassionate use off-label) are routinely discussed in an tumor board – however, the registration of a patient to the tumor board relies on the initiative of the managing clinician, who is also required to be present during the discussion – imposing a pivotal role on individuals who often manage several critical tasks at a time.

Given the high cost of liquid biopsy assays, rigorous justification of their clinical value is essential—particularly in publicly funded health systems. In our setting, the lack of standardized clinical data fields and tumor board documentation made it difficult to determine whether and how ctDNA findings were acted upon or influenced patient care, limiting our ability to justify the cost of ctDNA testing to institutional stakeholders. This appears particularly unfortunate in light of our conclusion that repeated testing may be necessary to utilize the full diagnostic potential of ctDNA analysis, where testing costs may increase exponentially.

Although our study cannot provide a formal conclusion on cost-effectiveness, it clearly highlights that high-quality databases, such as national or international prospective registries, are necessary in order to address this question in the future and provide prospective validation of our data. Ultimately, any large-scale decisions or trends will require prospective multicentric studies in order to fully understand the relationship between investment and patient benefit. Nonetheless, in our current clinical implementation, a more rigorous pre-test selection (i.e. via tumor board recommendation) may help to compensate the additional cost incurred by matched blood leukocyte sequencing, which would raise the confidence of negative results or putative germline variants.

Besides providing analytical validation and a retrospective assessment of clinical utility, in the present study we identified potential areas of improvement for our institution: harmonized and standardized tumor board templates for structured reporting, collaborative redesign of hospital information workflows with the support of data science experts, education on the importance of clinical annotation, and potential industrial ventures to support the so-called digital transformation of the working environment. These steps may improve diagnostic interpretation and patient outcomes, but will certainly facilitate evidence generation, regulatory compliance, and cost-effectiveness assessments for emerging technologies such as ctDNA.

Limitations

This study presents with several limitations; firstly, arising from the retrospective design, and retrospective annotation of biomarkers using Cancer Genome Interpreter. Secondly, specific limitations arise from assay design and performance; due to lack of blood leukocyte analysis, we were unable to discriminate between tumor mutations, clonal hematopoiesis, and germline variants. Since we did not calculate tumor fraction at time of reporting, the negative predictive value of our tests is unknown. A portion of tests performed with variable sequencing depth, which was sometimes inadequate to confidently exclude low frequency alterations. We highlighted the high frequency of low-evidence fusion events which we consider likely to be false positive. Regarding quality control, per-strand (forward/reverse) read counts were unavailable, preventing formal strand-bias testing. Read-level quality scores and background error modeling parameters were not accessible within the proprietary pipeline, precluding independent verification of these aspects of variant calling.

Conclusion

This study reaffirms that ctDNA analysis using the AmoyDx Comprehensive Assay is analytically robust for small variants, with an LOD95 of 0.37% VAF. In addition, it has the potential to detect fusions, albeit limited by lower sensitivity and specificity. The assay identified potentially actionable genomic alterations in a significant proportion of our patients; however, our assessment of utility remains limited due to lack of known tumor fraction, discrimination of clonal hematopoiesis or germline variants, and incomplete clinical context. Besides technical assay refinement, we must strive for structured and harmonized clinical documentation and increased clinician awareness in order to improve patient treatment and prospective generation of datasets. The field of liquid biopsy is rapidly evolving; large-scale prospective studies, as well as ongoing evolution of assay chemistry and panel design, will shape future landscape of clinical ctDNA testing.

Abbreviations

CGP

Comprehensive Genomic Profiling

ctDNA

Circulating tumor DNA

cfDNA

Cell-free DNA

VAF

Variant allele frequency

ASCO

American Society of Clinical Oncology

CAP

College of American Pathologists

AMP

Association for Molecular Pathology

CLIA

Clinical Laboratory Improvement Amendments

ISO 15189

International Organization for Standardization standard 15189

MTB

Molecular Tumor Board

LOD

Limit of detection

LOB

Limit of blank

NSCLC

Non-small cell lung cancer

ELBS

European Liquid Biopsy Society

HRR

Homologous recombination repair

mCRPC

Metastatic castration-resistant prostate cancer

VUS

Variants of unknown significance

UMI

Unique molecular identifier

SOP

Standard operating procedure

VICC

Variant Interpretation for Cancer Consortium

MSKCC

Memorial Sloan Kettering Cancer Center

API

Application Programming Interface

Q30

Phred quality score ≥ 30

BLAST

Basic Local Alignment Search Tool

RefSeq

Reference Sequence (NCBI)

MDT

Multidisciplinary Team

FDA

U.S. Food and Drug Administration

EMA

European Medicines Agency

NGS

Next-generation sequencing

bp

Base pairs

pM

Picomolar

ECOG

Eastern Cooperative Oncology Group (performance status)

Authors’ contributions

AO and LS designed the study. AO wrote the main manuscript text and prepared all figures. LS, AO, FO, AS, EM and LM participated in data collection. LZ performed considerable manuscript and figure editing. All authors reviewed the manuscript.

Funding

Not applicable.

Data availability

The datasets used and/or analysed during the current study are included in the supplementary data, alongside code to reproduce the main figures.

Declarations

Ethics approval and consent to participate

This study was performed in accordance with the Declaration of Helsinki. Approval of the ethical board of the Medical University of Vienna for the retrospective study was granted under 1989/2024 (informed consent was waived for retrospective use of data).

Consent for publication

Not applicable.

Competing interests

A.O. has received honoraria from Illumina, StemLine, AstraZeneca. The remaining authors have no conflicts of interest to declare.