The role of circulating tumor DNA (ctDNA) to detect minimal residual disease in locally advanced gastroesophageal carcinoma: the BUTTERFLY study

Abstract

Background

Despite advances in perioperative and neoadjuvant strategies, patients with locally advanced gastroesophageal cancers remain at high risk of recurrence after curative intent treatment. No validated biomarkers are available to detect minimal residual disease (MRD) or to guide post-operative risk-adapted management. Circulating tumor DNA (ctDNA) has emerged as a noninvasive tool for disease monitoring; single-parameter or tumor-informed assays, however, may lack sensitivity in low-tumor burden settings. Multimodal, tumor-agnostic approaches may overcome these limitations.

Methods

The BUTTERFLY study is a prospective, multicenter observational study enrolling patients with stage II-III gastric, gastroesophageal junction, or esophageal cancer treated with perioperative chemotherapy or neoadjuvant chemoradiotherapy followed by surgery. It evaluates the diagnostic performance and prognostic value of an academic, tumor-agnostic, multimodal ctDNA assay for MRD detection and prognostic stratification. Serial plasma samples are collected from baseline through post-operative follow-up and at relapse. Cell-free DNA is analyzed using the Agnostic Liquid Biopsy Multimodal Advancement (ALMA) platform, integrating tumor fraction estimation, somatic copy number alterations, fragmentomic features, single-nucleotide variants, and whole-genome methylation profiling. Multimodal features are combined with clinical variables using machine learning–based models to enhance MRD detection and relapse risk stratification. The primary endpoint includes sensitivity and specificity of ALMA-defined ctDNA/MRD status at the 4-8 weeks after surgery landmark, whereas secondary endpoints assess diagnostic performance at other time points and associations between ctDNA status and dynamics with disease-free survival, overall survival, treatment response, and lead time to recurrence.

Future perspectives

If validated, this tumor-agnostic, multimodal ctDNA approach may enable earlier molecular relapse detection and support personalized post-operative management strategies.

Highlights

• •Prospective multicenter study in locally advanced gastroesophageal cancer.
• •Evaluation of the tumor-agnostic multimodal ALMA assay for MRD detection.
• •Integration of fragmentomics, methylation, SCNAs, SNVs, and TF.
• •Longitudinal plasma sampling from baseline through treatment and relapse.
• •Potential for early relapse detection and personalized post-operative care.

Introduction

Gastric cancer (GC) ranks seventh in global incidence and fifth in mortality, whereas esophageal cancer (EC) and gastroesophageal junction cancer (GEJC) rank eighth in incidence and sixth in mortality, together accounting for >1.7 million cases and 1.3 million deaths in 2020.^1,2 For locally advanced disease, multimodal treatment with perioperative therapy and surgery remains the only potentially curative strategy and is critical for extending survival in GC/GEJC. Despite advances with neoadjuvant and perioperative regimens, however, recurrence rates after perioperative chemotherapy or neoadjuvant chemoradiotherapy (CRT) followed by surgery remain high, particularly in patients who did not achieve a pathological complete response.3, 4, 5 With perioperative fluorouracil, leucovorin, oxaliplatin and docetaxel (FLOT), 5-year survival remains ∼50%. The addition of durvalumab to FLOT in the phase III MATTERHORN study, while having improved event-free survival outcomes relative to FLOT alone, is still associated with a >30% recurrence rate within 2 years. Histopathological biomarkers (e.g. nodal status and tumor regression grade) support prognostication but do not enable risk-adapted treatment decisions and are limited by their postsurgical availability. The advancement of the MATTERHORN study further highlights the need for earlier, surgery-independent biomarkers to refine therapeutic stratification.⁶

Liquid biopsy (LB) and the analysis of cell-free DNA (cfDNA) isolated from plasma enable minimally invasive longitudinal monitoring of tumor-associated genomic and epigenomic alterations. The detection and analysis of circulating tumor DNA (ctDNA)—the fraction of cfDNA released into the bloodstream by tumor cells—have been revolutionizing clinical practice across several cancer types.7, 8, 9 ctDNA can be measured using either tumor-informed assays, based on the genomic alterations identified in the primary tumor, or tumor-agnostic assays, such as the approach proposed in this study. Although both have been extensively researched, there is not yet evidence that one or the other should be preferred to monitor minimal residual disease (MRD).⁷

Beyond monoparameter assays, recent studies show that integrating multiple cfDNA-derived features substantially enhances the sensitivity of ctDNA detection. These features include tumor fraction (TF) estimation, somatic copy number alterations (SCNAs) profile, fragmentomic patterns—which reflect tumor-specific chromatin structure and nuclease activity—methylation signals, and mutational profiles.^10,11 The combined analysis of these complementary dimensions enables more accurate longitudinal disease monitoring and supports timely, tailored therapeutic interventions.¹²

A multiomic cfDNA approach is particularly critical in clinical scenarios characterized by very low tumor burden, such as MRD after curative intent surgery or at the completion of systemic chemotherapy. In these settings, ctDNA concentrations often fall below the detection threshold of conventional single-parameter assays, resulting in high false-negative rates and limited clinical utility.¹² The simultaneous interrogation of orthogonal biological signals can compensate for the limited amount of ctDNA molecules. By leveraging the complementary nature of these features, multiomic strategies markedly improve analytical sensitivity and specificity, enabling the detection of residual tumor clones that would otherwise remain occult.

Critically, such enhanced resolution has direct clinical implications: earlier identification of molecular relapse, more accurate stratification of relapse risk, and improved capacity to guide treatment decisions or surveillance strategies. Although challenges remain—including standardization, cost, and the need for rigorous validation—multiomic integration currently represents one of the most promising avenues for advancing MRD detection and for extending the actionable utility of LB.

Within the spectrum of gastrointestinal (GI) malignancies, data on localized colon cancer have shown that ctDNA concentrations can be used to predict treatment outcomes and support individualized strategies.^13,14 Retrospective studies and prospective case series have provided initial evidence that detection of ctDNA may be associated with survival and treatment response in gastroesophageal cancers.^15,16

This study aims to assess the diagnostic performance and clinical utility of a tumor-agnostic, multimodal ctDNA assay for MRD assessment, using subsequent clinically documented recurrence as the clinical reference standard, and to evaluate its prognostic and predictive value in patients with stage II-III gastroesophageal adenocarcinoma or locally advanced esophageal squamous-cell carcinoma after preoperative chemotherapy or CRT at two international high-volume centers. The results of this study may inform future larger studies investigating ctDNA-based tailored treatments for patients with operable disease that could be implemented in clinical practice.

Patients and methods

Overall study design and cohort selection

This study is an international, prospective, observational, multicenter study designed to assess the accuracy of a tumor-agnostic ctDNA assay for MRD detection and its prognostic and predictive value in patients with stage II-III gastroesophageal adenocarcinoma or locally advanced esophageal squamous-cell carcinoma after preoperative chemotherapy or CRT. This study includes patients treated at two different high-volume centers: Istituti di Ricovero e Cura a Carattere Scientifico Humanitas Research Hospital (Milan, Italy) and Memorial Sloan Kettering Cancer Center (New York, United States). The target enrollment is ∼120 patients, over an anticipated accrual period of 36 months. Our multidisciplinary project involves medical oncologists, surgeons, gastroenterologists, biologists, and pathologists. Patients with newly diagnosed stage II-III gastroesophageal adenocarcinoma or squamous-cell carcinoma attending the participating centers and satisfying the inclusion criteria of the study are proposed to participate in this study.

Patients enrolled in this trial are treated according to the best clinical practice and the current local standard-of-care recommendations adopted across participating centers. These reflect the current major national and international guidelines in Western countries, which recommend the perioperative FLOT regimen for locally advanced adenocarcinomas for four preoperative and four post-operative cycles in patients fit enough to receive this triplet regimen or an alternative fluoropyrimidine-based chemotherapy regimen for patients who cannot be treated with FLOT. Starting from 15 November 2025, durvalumab became available in Italy; thus, patients will be treated with durvalumab plus FLOT (D-FLOT), as per the MATTERHORN study schema. On the contrary, patients with locally advanced squamous-cell carcinoma are treated with the CROSS regimen [five weekly cycles of carboplatin and paclitaxel with concurrent radiotherapy (41.4 Gy in 23 fractions, 5 days per week)] followed by surgery. Surgery and post-operative management are conducted according to the current standard guidelines and clinical practice at each participating center. Tumors located in the lower esophagus or at the esophagogastric junction are treated with modified hybrid two-stage esophagectomy, whereas tumors located in the upper or middle thoracic esophagus are treated with a three-stage, fully minimally invasive esophagectomy. GCs are treated with minimally invasive subtotal or total gastrectomy, depending on tumor location.

To address treatment heterogeneity, the specific perioperative or neoadjuvant regimen, use of durvalumab, completion of planned treatment, radiotherapy exposure, surgical procedure, pathologically documented response (pathological response), and post-operative management will be prospectively recorded. The treatment regimen will be included as an adjustment variable in exploratory analyses. When the number of events allows, patients with adenocarcinoma treated with chemotherapy alone will be compared with those treated with D-FLOT; otherwise, these analyses will be interpreted descriptively. Patients with squamous-cell carcinoma will be analyzed separately.

In the framework of this study, longitudinal plasma samples will be collected at different time points, as graphically detailed in Figure 1.

Figure 1.

Study design. In patients who undergo perioperative CT or preoperative CRT, ^§time points for ctDNA assessment during adjuvant therapy and follow-up are measured from the date of surgery.

CRT, chemoradiotherapy; CT, chemotherapy; ctDNA, circulating tumor DNA; GE, gastroesophageal; LB, liquid biopsy.

In particular, for all patients enrolled in the study, plasma for ctDNA analysis will be collected at six mandatory time points: at baseline before initiation of preoperative therapy (LB1), at completion of preoperative therapy before surgery (LB3), 4-8 weeks after surgery (LB4), and during follow-up at 6 and 12 months (LB5 and LB6). Plasma will also be collected at the time of disease relapse (LB9), which represents an additional mandatory sampling point. Furthermore, three optional collections will be carried out during preoperative treatment when feasible (LB2), at 18 and 24 months of follow-up (LB7 and LB8). LB2 samples are not mandatory, but if they are assessed, they are collected after two cycles of FLOT or during the third week of the CROSS regimen.

The 4-8 week post-operative window for LB4 is selected as the key postsurgical time point for MRD assessment because it allows clearance of immediate surgery-related cfDNA and inflammatory signals while capturing early residual disease before the initiation of post-operative management.

Sample collection

For longitudinal cfDNA analysis, a minimum of 15 ml of whole blood will be collected at each specified time point in K2-EDTA tubes, stored, and processed following the requirements for cfDNA analysis¹⁷ as previously described.¹⁸ cfDNA will be extracted and purified from plasma samples using the Maxwell RSC ccfDNA Plasma Kit (Promega) on an automatic purification System (Maxwell RSC, Promega). cfDNA concentration will be measured by Qubit dsDNA HS Assay Kit (Thermo Fisher Scientific, Waltham, MA), and its size distribution will be evaluated using the TapeStation System Analyzer (Agilent Technologies, Santa Clara, CA).

Additionally, peripheral blood mononuclear cells (PBMCs) will also be purified from the blood collected in K2-EDTA tubes and processed for subsequent analyses to account for clonal hematopoiesis of indeterminate potential (CHIP) whenever feasible.

Agnostic Liquid Biopsy Multimodal Advancement platform for longitudinal ctDNA analysis

To pursue the study’s objectives, cfDNA will be analyzed using a tumor-agnostic and multimodal approach developed and standardized in IRCCS Humanitas Research Hospital, which serves as the central analytical laboratory for the study. As reported in Figure 2, our experimental protocol Agnostic Liquid Biopsy Multimodal Advancement (ALMA) allows, with a single multistep experiment, to infer approximately five different biological features characterizing cfDNA in patients with cancer:

• ▪Percentage of ctDNA on the total amount of cfDNA (TF)
• ▪SCNA profile (SCNAs >1 Mb)
• ▪Fragmentomic profile
• ▪Single-nucleotide variants (SNV)
• ▪Whole-genome methylation pattern [cell-free Methyl-CpG-binding Domain sequencing (cf-MDB-seq)]

Figure 2.

ALMA workflow. Schematic representation of the experimental workflow of the ALMA approach.

ALMA, Agnostic Liquid Biopsy Multimodal Advancement; cf-MDB-seq, cell-free Methyl-CpG-binding Domain sequencing; sWGS, shallow whole-genome sequencing; SNV, single-nucleotide variant; SCNA, somatic copy number variation.

cfDNA will be processed following a standardized procedure (KAPA HyperPlus, Roche, Basel, Switzerland) and barcoded to create a pool (eight plasma samples per pool, 80-200 ng).

As shown in Figure 2 (purple square), 10% of the pool undergoes whole-genome amplification and low-coverage sequencing (1X)—shallow whole-genome sequencing (sWGS)—using the MGI sequencing platform (Figure 2, purple square). To investigate genome-wide cfDNA fragmentation patterns, aligned Binary Alignment Map files will be processed using the end selection fragmentomics pipeline developed by Ju et al.¹⁹ Briefly, cfDNA fragment ends will be compared with a reference nucleosome track, retaining only fragments with at least one end located within the nucleosome-protected regions. This end selection process is used to derive three fragmentomic features: (i) N-index, defined as the proportion of cfDNA fragments with ends located in nucleosome-protected regions; (ii) Delta-S150, representing the change in the percentage of short fragments (≤150 bp) before and after end selection; and (iii) Delta-M, measuring the percentage change of fragments carrying the DNASE1L3-preferred cleavage motif. As reference controls, cfDNA samples from healthy donors are processed using the same analytical pipeline. Nucleosome-protected regions are also used to subset the aligned reads for downstream analyses. TF, ploidy, and somatic copy number aberrations (SCNAs) will be inferred from subsetted aligned reads using the SAMURAI pipeline.²⁰ A size selection of cfDNA fragments between 90 and 150 bp will be applied, and ichorCNA²¹ will be used within the pipeline to estimate TF and identify large-scale SCNAs. We will use the libraries previously generated for sWGS analysis for a targeted sequencing experiment using the KAPA HyperCap Target Enrichment Probes Kit (Roche Diagnostics, Indianapolis, IN), with a sequencing coverage of 10.000X (Figure 2, blue square). This targeted enrichment will be carried out using a customized panel of 182 cancer-related genes to characterize the evolving molecular landscape of the tumor, thereby potentially guiding the selection of the most appropriate therapeutic option.

The remaining part (90%) of the original pool is used for the whole-genome methylation analysis (Figure 2, green square), and the analysis will be carried out using a bisulfite-independent experimental protocol (cf-MDB-seq).²² This protocol is based on an enrichment step followed by sequencing of the methylated region captured (50 million reads per sample). Raw sequencing data are aligned to the human reference genome (hg38) and processed using an in-house pipeline built on the mesa R package (https://github.com/cruk-mi/mesa). Methylation profiles are normalized using PBMC reference samples, and tumor-derived informative regions are selected through in silico peripheral blood leukocyte depletion, as described by Burgener et al.²³

Machine learning model development

The five distinct biological features analyzed by ALMA will first be classified using predefined analytical thresholds. Specifically, TF positivity will be defined as TF ≥2%; SNV positivity will be defined as the detection of at least one high-confidence plasma-derived somatic variant with a variant allele frequency ≥0.15%, after filtering variants likely attributable to clonal hematopoiesis or technical artifacts; fragmentomic positivity will be defined as a fragment length ratio exceeding the 90th percentile of the healthy control distribution; and methylation positivity will be defined as the detection of tumor-associated differentially methylated regions (DMRs) relative to the healthy control reference cohort. Predefined analytical thresholds will be used to derive binary feature-level positivity indicators, whereas continuous or quantitative ALMA-derived metrics may also be retained for machine learning model development.

These molecular variables, together with selected clinical parameters, will subsequently be integrated using a random survival forest (RSF) model to generate an ALMA risk score. The survival outcome for model development will be defined as the time from the corresponding LB time point to clinically documented recurrence or censoring.

Before model development, the study population will be randomly divided at the patient level into training and internal held-out validation subsets, using an ∼2 : 1 ratio and stratification by recurrence status when feasible. Feature selection will be carried out exclusively within the training dataset using predefined statistical and machine learning approaches, retaining biologically relevant and nonredundant variables. Model optimization, including hyperparameter tuning and determination of the risk score threshold defining ALMA positivity, will be carried out within the training set using cross-validation procedures to minimize overfitting. The final locked RSF model and positivity threshold will then be evaluated in the held-out validation subset without further modification.

Given the exploratory nature of the study and the expected number of recurrence events, model development and validation analyses will be considered hypothesis-generating.

All study procedures are listed in Table 1.

Table 1.

Study procedures flowchart

Assessment/events	Screening (28 days)	During neoadjuvant treatmenta	Before surgery	After surgery	Follow-up
Day/months	D0		≤4 weeks to surgery	4-8 weeks after surgery	6 months (±6 weeks)	12 months (±6 weeks)	18 months (±6 weeks)	24 months (±6 weeks)	At relapse
Signed ICF	X
Demographic data	X
Medical history	X
Prior and concomitant therapies	X	X	X	X	X	X	X	X
Vital signs	X	X	X	X	X	X	X	X
Physical examination	X	X	X	X	X	X	X	X
Hematology	X	X	X	X	X	X	X	X
Blood chemistry	X	X	X	X	X	X	X	X
Liquid biopsy (ctDNA)	X	Xa	X	X	X	X	Xa	Xa	X
Tumor sample and biomarkers analysis	X		X	X (surgical)
CEA, CA 19-9	X	X	X	X	X	X	X	X
Adverse events		X	X	X	X	X	X	X
Radiological restaging procedures including TC TB and PET-FDG (if indicated)	X		X	X	X	X	X	X
Surgical assessment	X	X	X	X	X	X	X	X
Oncological assessment	X	X	X	X	X	X	X	X

This table details procedures for the screening, treatment, and follow-up periods.

CA, carbohydrate antigen; CEA, carcinoembryonic antigen; ctDNA, circulating tumor DNA; FDG–PET, [¹⁸F]2-fluoro-2-deoxy-D-glucose positron emission tomography; ICF, informed consent form; TC TB, CT SCAN Total Body.

^aNot mandatory.

Study objectives

This project is conceived to assess the following objectives.

Primary objective

The primary objective was to explore the diagnostic performance of the ALMA tumor-agnostic, multimodal ctDNA assay for MRD assessment at the post-operative LB4 landmark, collected 4-8 weeks after surgery, for the detection of subsequent clinically documented recurrence within 24 months from the LB4 assessment in patients with stage II and III gastroesophageal cancer undergoing curative intent standard-of-care treatments.

Secondary objectives

The secondary objectives were as follows:

• •To evaluate the diagnostic performance of the ALMA tumor-agnostic, multimodal ctDNA assay at other prespecified LB time points.
• •To correlate ctDNA status at different time points with response to treatment and survival outcomes.
• •To correlate dynamic changes in serial ctDNA measurements during treatment with response and survival outcomes.
• •To explore the lead time between ctDNA/MRD positivity and clinically documented recurrence.

Translational analysis

Translational analysis aimed for the following:

• •To assess the predictive and prognostic value of integrating ctDNA information with genomic, molecular, and immune-related signatures obtained through multimodal profiling of formalin-fixed, paraffin-embedded tumor samples [whole-genome sequencing (WGS), whole-exome sequencing (WES), loss of heterozygosity, copy number variations, immunohistochemistry (IHC) for clinically validated biomarkers, and transcriptomics (RNA) analyses].
• •To assess the predictive and prognostic roles of traditional biomarkers (such as programmed death-ligand 1 combined positive score, tumor proportion score and tumor area positivity score, (human epidermal growth factor receptor 2 status, DNA mismatch repair status, claudin18.2, and fibroblast growth factor receptors 2b) on patients’ clinical outcomes.
• •To assess the mechanisms of acquired resistance to treatment through exploratory biomarker analyses on tumor samples and ctDNA.
• •To investigate the tumor microenvironment composition, its spatial organization, and its influence on the clinical outcomes (spatially resolved transcriptomic).

Endpoints

This study will evaluate the following.

Primary endpoint

• •Sensitivity, specificity, positive predictive value, and negative predictive value of ALMA-defined ctDNA/MRD status at the post-operative LB4 landmark for the detection of subsequent clinical recurrence, defined as radiological and/or pathological evidence of disease relapse within 24 months from the LB4 LB assessment, in patients with stage II and III gastroesophageal cancer undergoing curative intent standard-of-care treatments.
• -Disease-free survival (DFS), overall survival (OS), and clinical and pathological response by ctDNA status.
• -DFS, OS, and clinical and pathological response according to dynamic changes in serial ctDNA measurements.
• •Lead time between ctDNA/MRD positivity and clinically documented recurrence.

Translational endpoint

The translational endpoint was correlation between ctDNA-derived features and other molecular markers, and their integrative predictive and prognostic value.

Eligibility

Inclusion criteria

Subjects fulfilling all the following inclusion criteria are eligible for the study:

• •Age ≥18 years.
• •Newly histologically documented stage II or III gastric, gastroesophageal junction, and esophageal adenocarcinoma or squamous-cell carcinoma, referred to one of the participating centers.
• •Absence of distant metastases as defined by baseline radiological tumor staging with contrast-enhanced CT scan of the thorax and abdomen-pelvis. Alternative staging procedures, such as non–contrast-enhanced CT and magnetic resonance imaging, may be considered for patients with contraindications to contrast-enhanced CT. Use of [¹⁸F]2-fluoro-2-deoxy-D-glucose positron emission tomography and/or laparoscopy to complete tumor staging will be at investigators’ discretion and according to the institutional procedures of each participating center.
• •Deemed eligible for standard neoadjuvant treatment with perioperative chemotherapy or preoperative CRT and surgery as per investigators’ assessment following multidisciplinary board discussion.
• •Ability to provide informed consent according to International Council for Harmonisation (ICH) Guideline/European Union for Good Clinical Practice (GCP) and Italian regulations.
• •Eastern Cooperative Oncology Group performance status ≤2.

Exclusion criteria

The presence of any one of the following exclusion criteria will lead to the exclusion of the subject:

• •Stage I, recurrent, or metastatic gastroesophageal cancer at baseline.
• •Histology other than adenocarcinoma and/or squamous-cell carcinoma.
• •Prior anticancer systemic or radiotherapy for gastroesophageal cancer.
• •History of another neoplastic disease, unless in remission for ≥5 years. Participants with basal cell carcinoma of the skin, squamous-cell carcinoma of the skin, or carcinoma in situ (e.g. breast carcinoma and cervical cancer in situ) who have undergone curative intent therapy with no evidence of residual or recurrent disease can be considered per investigators’ assessment.
• •Contraindications to surgery other than tumor stage per investigators’ assessment.
• •Patient unable to comply with the study protocol owing to psychological, social, or geographical reasons.

Statistical methods

Given the study’s descriptive intent, no formal statistical hypothesis was identified for testing, and no formal hypothesis-testing sample size calculation was carried out.

The planned enrollment is ∼120 patients over an anticipated accrual period of 36 months. This estimate is based on the expected number of eligible patients with stage II-III GC, GEJC, or EC treated with curative intent multimodal therapy at the participating centers. Based on historical recurrence rates in this disease setting, ∼40 recurrence events are expected during the first 24 months of follow-up. These numbers are considered adequate for descriptive analyses of ctDNA status and dynamics, as well as for exploratory, hypothesis-generating multimodal modeling.

At each prespecified LB time point, patients will be classified as ctDNA/MRD-positive or ctDNA/MRD-negative, on the basis of the ALMA assay output.

In the held-out internal validation subset of patients, sensitivity and specificity of the final locked ALMA assay output will be calculated by comparing ctDNA/MRD results for a specific time point with the subsequent occurrence of documented clinical recurrence, defined as radiological and/or pathological evidence of disease relapse within 24 months from that specific time point.

The primary diagnostic accuracy analysis will be carried out at the post-operative LB4 landmark, corresponding to the LB samples collected 4-8 weeks after surgery. Patients will be included in this landmark analysis if they are alive, are recurrence-free, and have an available ALMA result at LB4.

True-positive results will be defined as ctDNA/MRD-positive findings at LB4 followed by clinical recurrence within 24 months from LB4. False-positive results will be defined as ctDNA/MRD-positive findings at LB4 without documented clinical recurrence after adequate follow-up within the same 24-month follow-up window. True-negative results will be defined as ctDNA/MRD-negative findings at LB4 without subsequent recurrence within the same 24-month follow-up window, whereas false-negative results will be defined as ctDNA/MRD-negative findings at LB4 followed by clinical recurrence within the same 24-month follow-up window.

Secondary diagnostic accuracy analyses will evaluate sensitivity, specificity, positive predictive value, and negative predictive value of ALMA-defined ctDNA/MRD status at the other prespecified LB time points. These analyses will be carried out separately for each time point, using clinically documented recurrence within 24 months from the corresponding LB assessment as the operational clinical reference standard, provided that adequate follow-up is available.

LB9 samples, collected at the time of clinically documented relapse, will not be included in diagnostic performance analyses of ALMA-defined ctDNA/MRD predicting recurrence. Instead, LB9 will be analyzed descriptively by reporting the proportion of ALMA-positive samples among patients with confirmed recurrence, defined as the ctDNA detection rate at relapse, with corresponding 95% confidence intervals.

Patients without adequate follow-up will be considered nonassessable for binary diagnostic performance analysis and censored as appropriate in time-to-event analysis.

Data will be summarized as frequencies and proportions or as median and range. Sensitivity, specificity, positive predictive value, and negative predictive value will be reported with corresponding 95% confidence intervals. Chi-square/Fisher’s exact tests and the Wilcoxon test will be used to evaluate differences between groups in case of categorical data or continuous variables, respectively. Survival curves will be estimated using the Kaplan–Meier method, and differences among subgroups will be evaluated using the log-rank test. The Cox proportional hazards model will be used to estimate hazard ratios and corresponding 95% confidence intervals. Statistical significance will be set at 0.05 (two-sided).

Prespecified exploratory subgroup analyses will evaluate assay performance and prognostic associations by anatomical site (gastric, gastroesophageal junction, and esophageal), histology (adenocarcinoma versus squamous-cell carcinoma), treatment strategy (FLOT or alternative chemotherapy alone, D-FLOT, and CROSS), pathological stage, and pathological response. Because of the expected limited number of events in some subgroups, these analyses will be interpreted descriptively and will be used mainly to inform future confirmatory studies.

Ethical considerations

The study investigators ensure that this study is conducted in agreement with this protocol, the Good Clinical Practice, the current version of the Declaration of Helsinki, and the applicable regulations. The protocol was reviewed and approved by the competent Independent Ethics Committee of our center (authorization no. 4418 to conduct the clinical study code ONC/OSS-26/2024 on 3 June 2025). Before conducting any of the procedures not carried out routinely in this study, the investigator explains the study to the patient by means of the informed consent process and collects a signed copy of written informed consent. The investigators also explain that during this study, patients’ biological samples are collected and analyzed by the IRCCS Humanitas Research Hospital laboratory by means of the informed consent process. Each patient enrolled in this study receives a unique subject ID.

Discussion

GC and gastroesophageal cancer remain major clinical challenges because of their high recurrence rates, poor prognosis, and frequent presentation at advanced stages. In clinical practice, there is a strong need for novel biomarkers to optimize treatment strategies and improve patient outcomes, particularly in the locally advanced setting. In recent years, ctDNA has emerged as a highly sensitive LB modality for the real-time detection of genetic and epigenetic tumor alterations. This approach offers several advantages over conventional tissue biopsy because it not only provides insight into intratumoral heterogeneity but also enables longitudinal monitoring through serial sampling, allowing assessment of dynamic treatment response and early detection of disease recurrence.²⁴ Notably, ctDNA has been extensively evaluated for MRD assessment, supported by its relatively short half-life of ∼2 h, which allows its concentrations to mirror tumor burden and kinetics in real time.²⁵

The methodologies used for ctDNA detection span from digital polymerase chain reaction techniques to next-generation sequencing platforms. These analyses can be carried out using either a tumor-informed approach, in which the primary tumor is sequenced to identify somatic mutations that are subsequently traced in plasma, or a tumor-agnostic approach that relies on standardized gene panels to detect recurrent mutations. Although the tumor-informed strategy generally provides higher sensitivity, its implementation remains less accessible.²⁶ More recently, methylation-based assays have been introduced, enabling combined genomic and epigenomic profiling to further improve both sensitivity and specificity.²⁴ In this context, emerging evidence suggests that global cfDNA methylation patterns may provide complementary information for MRD assessment in EC and esophagogastric junction cancers, supporting the clinical relevance of epigenetic signals beyond purely sequence-based alterations.²⁷

Several systematic reviews and meta-analyses collected data from different retrospective and observational trials that had evaluated ctDNA before and after radical treatment (CRT or surgery) in patients with locally advanced gastroesophageal cancers. These analyses indicate that ctDNA dynamics may complement conventional imaging and serum markers in evaluating tumor regression and detecting MRD, often anticipating radiological recurrence by several months.²⁷

The post-operative detection of ctDNA is strongly associated with a higher risk of recurrence and inferior survival outcomes, whereas ctDNA-negative patients may remain recurrence-free for a significantly longer period. These results underscore the potential role of ctDNA as not only a diagnostic but also a prognostic biomarker in these patients. The PLAGAST trial is a recent observational study that collected blood samples from patients with locally advanced resectable gastric/gastroesophageal junction adenocarcinomas and retrospectively analyzed ctDNA, using a personalized, tumor-informed 16-plex mPCR-NGS assay (Signatera; Natera, Inc., Austin, TX), at pre-neoadjuvant therapy (NAT), during-NAT, post-NAT, and postsurgical MRD window. ctDNA positivity across multiple disease time points was associated with inferior relapse free survival and OS, whereas patients achieving clearance during NAT showed superior outcomes compared with those clearing ctDNA only after NAT. The evaluation of ctDNA dynamics over time, also in the presurgical setting, may represent an additional layer of information for future risk stratification strategies, potentially informing multidisciplinary discussions on surgical management. Notably, ctDNA status emerged as the strongest independent predictor of recurrence, outperforming even pathological complete response.¹⁵

Ongoing clinical trials are now exploring ctDNA-guided adjuvant decision making in stage II-III GC, testing therapy intensification or de-intensification on the basis of post-operative ctDNA status. ctDNA is also being leveraged to track resistance mechanisms, identify emerging mutations, and guide early intervention or combination treatment strategies in response to clonal evolution.²⁸

Despite these promising data, some technical challenges continue to limit the widespread clinical use of ctDNA MRD detection. Preanalytical variables (handling, processing times, and cfDNA extraction methodology) can affect reproducibility, and it may be difficult to distinguish tumor-derived variants from those arising from clonal hematopoiesis.²⁹ In the present study, this last issue will be addressed by applying a conservative filtering strategy to identify and exclude variants with CHIP-like features, including recurrent variants showing stable variant allele frequencies across longitudinal plasma samples and variants occurring in genes commonly associated with clonal hematopoiesis. Furthermore, we will collect PBMCs/white blood cells as a matched hematopoietic reference whenever feasible. Plasma variants also detected in the matched hematopoietic compartment will be flagged and excluded from ctDNA-based MRD calls, unless supported by tumor tissue sequencing, when available, providing an additional level of validation of the tumor origin of selected variants.

Building upon this evidence, our study proposes an integrated and innovative approach to assess ctDNA and MRD through an academic, tumor-agnostic, multimodal ctDNA assay (ALMA workflow), specifically developed and standardized at IRCCS Humanitas Research Hospital, serving as the coordinating analytical center of the multicenter study.³⁰ Unlike previous studies relying exclusively on single-parameter genomic assays, the ALMA platform combines multiple layers of biological information, including TF estimation, SCNAs, fragmentomics, SNVs, and whole-genome methylation patterns, thereby enabling a comprehensive molecular profiling of residual disease. This integrated multilayer approach has been designed to improve sensitivity and specificity and may enhance MRD detection accuracy, ultimately refining risk stratification in stage II-III gastroesophageal cancers.

A major strength of our work is its prospective nature, including longitudinal sampling at predefined clinical milestones—from baseline to 24-month follow-up and recurrence—allowing real-time tracking of ctDNA kinetics relative to treatment response and outcomes. Early identification of molecular relapse before radiographic evidence may offer an opportunity for preventive therapeutic intervention or personalized surveillance, aligning with evolving precision oncology paradigms.

Furthermore, integrating ctDNA findings with tissue-based molecular data (WGS, WES, transcriptomics, IHC, and spatial transcriptomics) will offer a unique opportunity to explore the biological underpinnings of treatment resistance, clonal evolution, and immune–tumor microenvironment interactions.

We acknowledge the biological and clinical heterogeneity of the study population, which includes gastric, gastroesophageal junction, and esophageal tumors, both adenocarcinoma and squamous-cell carcinoma histologies, and chemotherapy- and radiotherapy-based treatment strategies. Nevertheless, this heterogeneity reflects the real-world spectrum of curative intent management for upper GI cancers and may therefore provide a valuable benchmark for future MRD studies in these diseases. To account for the potential influence of primary tumor site, histology, and treatment modality on ctDNA shedding, recurrence patterns, and clinical outcomes, these factors will be incorporated, whenever feasible, into prespecified subgroup, stratified, and multivariable-adjusted exploratory analyses. This will include clinically relevant treatment subgroups, such as patients receiving D-FLOT versus chemotherapy alone and those with squamous-cell carcinoma treated according to the CROSS regimen.

If validated, the ALMA platform could serve as a clinically actionable tool for MRD monitoring in gastroesophageal cancers, helping bridge the gap between molecular biology and real-world clinical decision making. The implementation of such LB-based technologies may ultimately improve patient outcomes, optimize therapeutic resources, and reduce recurrence-related morbidity in this high-risk population.

The evaluation of ctDNA dynamics over time, rather than absolute values at a single time point, may also represent an additional layer of information for future risk stratification strategies. Whether such longitudinal changes could support treatment personalization, including surgical decision making, warrants dedicated investigation in specifically designed trials.

In conclusion, our prospective study represents a significant step toward establishing ctDNA as a robust prognostic and predictive biomarker in gastroesophageal cancer, offering a framework to redefine post-operative management and to lay the foundation for future MRD-driven therapeutic trials.