The role of dynamic monitoring of plasma cell-free DNA methylation in predicting pathological response in resectable stage IIB-IIIB non-small cell lung cancer: biomarker analyses from a prospective phase II trial.

Background
Surgery remains the paramount therapeutic strategy for resectable non-small cell lung cancer (NSCLC) [1]. However, the high rate of recurrence after surgery (30–60%) necessitates multidisciplinary treatment for locally advanced NSCLC [2]. Compared with the limited benefit of classical cytotoxic chemotherapy, immune checkpoint inhibitors (ICIs), whether used alone or combined with chemotherapy, have revolutionized the treatment paradigm of NSCLC patients, achieving significantly better pathological response and superior survival [3]. Notably, the KEYNOTE-671 and CheckMate 816 trials demonstrated superior efficacy of chemoimmunotherapy over chemotherapy in neoadjuvant setting for patients with resectable NSCLC with similar safety profiles [4, 5]. However, neoadjuvant chemoimmunotherapy does not confer benefit to all patients, as shown in completed or ongoing clinical trials. Several clinical parameters, including RECIST-criteria-based image evaluation and some potential biomarkers, such as PD-L1, tumor-infiltrating lymphocyte, or CD8+ T-cell infiltration, have demonstrated limited effectiveness in predicting treatment response [6–8]. Hence, optimal approaches are urgently needed for clinical decision-making.
Emerging liquid biopsy techniques have demonstrated a substantial efficacy in early cancer detection and treatment response prediction [9, 10]. For instance, circulating tumor DNA (ctDNA), a component of cell-free DNA (cfDNA), has offered a minimally invasive approach for predicting treatment response in the neoadjuvant setting and early detection of relapse after surgery in patients with NSCLC [11, 12]. Although ctDNA abundance, typically quantified by variant allele frequencies (VAFs) or haploid genome equivalents (hGE) per milliliter, provides some indication of tumor burden, its accuracy is constrained by detection thresholds and biological noise such as clonal hematopoiesis [13–15]. While advancements in personalized tumor-informed assays have improved the limit of detection (LOD), they come with a considerable cost [16].

Recent advancements in epigenetic modifications, particularly DNA methylation, have brought new insights into the field of liquid biopsy, addressing the drawbacks of ctDNA while maintains detection sensitivity to some degree [17, 18]. DNA methylation plays a crucial role in gene expression and the modification of chromatin conformation during carcinogenesis [19]. Notably, methylation biomarkers in cfDNA have been widely identified in the early stage of various cancers [20]. Epigenomic features of cfDNA, alone, or in combination with ctDNA, have demonstrated higher detection sensitivity compared to ctDNA alone [21]. Integrated lung cancer screening models that incorporate cfDNA methylation features and chromosomal alterations, exhibited an accuracy of over 90%, surpassing that of current clinical methods [22]. Alterations in cfDNA methylation can also serve as predictive biomarkers for targeted therapy and immunotherapy [23, 24]. A recent study established a pan-cancer methylation signature to quantify cancer-specific methylation (CSM) and fragment length score (FLS) in plasma cfDNA, and early changes of CSM and FLS could both predict outcomes in multiple advanced solid tumors treated with pembrolizumab [25]. In addition to mutations, chromosomal arm-level structural alterations are also prominent features of cfDNA from cancer patients and have shown potential as predictive biomarkers for treatment response in small cell lung cancer [26]. Given its exceptional performance in both lung cancer screening and predicting treatment response in late-stage lung cancer, we hypothesized that cfDNA features, including methylation patterns and chromosome variation, could serve as potential predictive biomarkers for neoadjuvant chemoimmunotherapy in NSCLC.
In the present study, we conducted a prospective open-label, single-arm, phase II trial (RENAISSANCE study) to investigate the efficacy and safety of neoadjuvant toripalimab (PD-1 inhibitor) plus chemotherapy in resectable stage IIB-IIIB NSCLC. Moreover, we innovatively explored if cfDNA features were associated with pathological response to neoadjuvant chemoimmunotherapy in NSCLC. By conducting targeted methylation panel sequencing with a median depth of 3000X, we analyzed the methylation profiling for baseline plasma samples and further calculated methylation fragment ratio scores (MFR) for all involved plasma samples. Meanwhile, we performed whole methylome sequencing (WMS) using enzymatic conversion of unmethylated cysteines (enzymatic methyl-seq) and assessed chromosomal aneuploidy of featured fragments (CAFF) score. The associations of CAFF and MFR before, during and after neoadjuvant therapy with pathological response were assessed. Additionally, an integrated model incorporating cfDNA features and immune parameters was developed to predict pathological response.

Methods

Study design and participants
This prospective single-arm, phase II clinical trial was conducted at Peking University Cancer Hospital & Institute (registration number: NCT04606303). It was conducted in accordance with the Declaration of Helsinki (revised in 2013) and Good Clinical Practices and approved by the Ethics Committee of Peking University Cancer Hospital & Institute (Institutional Review Board No. 2020YJZ58). All patients provided written informed consent prior to their enrollment in this study.
From December 18, 2020, to January 9, 2023, we enrolled a total of 100 patients according to the following main inclusion criteria: (1) age over 18 years; (2) clinical stage IIB to IIIB NSCLC (stage IIIB, T3-4N2M0 only, according to the 8th edition of the AJCC staging system for lung cancer); (3) Eastern Cooperative Oncology Group (ECOG) performance status of 0 or 1; (4) at least one measurable target lesion according to Response Evaluation Criteria in Solid Tumor version 1.1 (RECIST v.1.1); (5) adequate organ function; and (6) treatment-naïve or non-response to conventional chemotherapy (Fig. 1A). The key exclusion criteria were as follows: (1) presence of distant metastases or unresectable disease; (2) known EGFR mutations or ALK fusions; (3) uncontrolled autoimmune disease; (4) previous radiotherapy; and (5) previous use of antibiotics within 14 days.

Therapeutic procedures and sample collection
All patients received the following drugs intravenously in each 21-day treatment cycle: toripalimab (240 mg, day 1; TopAlliance Co. Ltd, Shanghai, China), cis-platinum (75 mg/m2, days 1 and 2), and pemetrexed (500 mg/m2, day 1) for lung adenocarcinoma (LUAD) or albumin-bound paclitaxel (260 mg/m2, day 1) for lung squamous cell carcinoma (LUSC) or NSCLC-not identified subtype (NSCLC-other). After two cycles, radiographical re-evaluation was conducted. If surgery was feasible, as evaluated by a multidisciplinary clinical team (MDT), it was scheduled for 4–6 weeks after the first day of the second cycle; otherwise, additional one or two cycles were considered, and the feasibility of surgery was re-evaluated after the additional treatment cycles. If surgery still could not be realized after 4 treatment cycles, the therapeutic schedule would be reformulated (Fig. 1A, Additional file 1: Study protocol).
Generally, tissue samples were collected before treatment via percutaneous needle or transbronchial biopsy, and PD-L1 expression was evaluated via immunohistochemistry and determined by the tumor proportion score (TPS) if possible. PD-L1 negative was defined as a TPS < 1%, whereas PD-L1 positive was defined as a TPS ≥ 1%. Within 1 week before cycle 1, blood samples were collected from some patients enrolled in this study for lymphocyte subpopulation analysis if possible. During treatment, plasma samples were collected within 3 days before every treatment cycle (C0: before cycle 1; C1: before cycle 2; C2: before cycle 3; C3: before cycle 4) and 1 day before surgery (BS) for cfDNA analysis if possible (Fig. 1C).

Safety and response assessments
Adverse events (AEs) and abnormal laboratory findings were monitored every week and graded according to the National Cancer Institute Common Terminology Criteria for Adverse Events (CTCAE) version 4.0. Treatment was considered to be delayed or ceased if treatment-related adverse events (TRAEs) occurred and resumed if certain criteria were met. Dose reduction was permitted for chemotherapy drugs if necessary.
Radiographic response was assessed according to the Response Evaluation Criteria in Solid Tumors (RECIST) criteria 1.1 [27]. The pathological response was determined according to the protocol of Travis et al., and assessment of pathological response was conducted by experienced thoracic pathologists at the Department of Pathology, Peking University Cancer Hospital & Institute. Major pathological response (MPR) was defined as the presence of 10% or less residual viable tumor (%RVT) in the resected primary tumor, and complete pathological response (pCR) was defined as the absence of viable tumor cells in both the resected primary tumor and lymph nodes. Non-MPR was defined as the presence of more than 10% RVT in the resected primary tumor. MPR was chosen as the primary endpoint, and the exploratory endpoint was the efficacy of cfDNA features in predicting response to neoadjuvant toripalimab and chemotherapy.

cfDNA extraction
Approximately 4 mL of each plasma sample was subjected to cfDNA extraction by using the MagMAX Cell-Free DNA Isolation Kit from Thermo Fisher Scientific, following the manufacturer’s instructions. The quantity and quality of the extracted cfDNA were assessed using a Bioanalyzer 2100 from Agilent.

WMS and targeted methylation sequencing (TMS)
Subsequently, a range of 5 to 30 ng of cfDNA was utilized to generate WMS libraries. This process was carried out with the NEBNext Enzymatic Methyl-seq Kit provided by New England Biolabs, following the manufacturer’s instructions. Half of the generated libraries were sequenced on a NovaSeq 6000 system from Illumina, featuring paired-end read lengths of 100 bp. The remaining half of the libraries were subjected to a TMS assay, covering a total of 1.16 Mbp across 766 genes using hybridization probes from Roche Diagnostics. These libraries were amplified with 9 cycles of PCR, followed by quantification using the Qubit dsDNA HS Assay Kit from Thermo Fisher Scientific. Finally, they were sequenced on a NovaSeq 6000 system from Illumina.

TMS-based methylation fragment ratio score
The construction of methylation correlation blocks (MCBs) served as the basic unit for methylation markers, grouping adjacent CpGs with highly similar methylation status. The criteria including proximity, Pearson’s correlation coefficient, and consecutive CpG sites were employed to define MCBs. The MCBs were then ranked based on cancer-specific mutual information (MI), and those with top MI scores were selected as hypermethylated markers for lung cancer detection [28].
Finally, a methylation score model, namely methylated fragment ratio (MFR) utilizing the methylation level of MCB was calculated, which was defined as the percentage of fully methylated fragments. It was constructed to assess the abnormality of methylation patterns in tested samples compared to healthy controls in the training cohort. Detailed MFR calculation method was described in our previous study [28].
Patients were categorized into high and low status using a pre-defined MFR cutoff value of 3.83 which was set at the specificity of 97% in the training control samples (n = 350) [28]. Early MFR clearance was defined as a shift from high to low status between C0 and C1 (before cycle 2), sustained until before surgery.

WMS-based chromosomal aneuploidy of featured fragments (CAFF) score
Cancer cells frequently exhibit chromosomal instability, characterized by partial or complete amplification or loss of chromosomal arms. To address this, we calculated a plasma aneuploidy (PA) score for each sample as previously reported [29]. This score is determined based on the levels of copy number alterations observed in five specific chromosome arms, which have shown the most significant copy number alterations compared to baseline samples. With a pre-defined cut-off value of 2.245 (set at the specificity of 98% in training control samples, n = 352) [29], patients were categorized into CAFF low and CAFF high groups. The definition for CAFF clearance followed the same criteria as MFR.

Statistical analysis
For continuous data, the median (range) will be presented. Categorical and ordinal data will be presented as frequencies (proportions). Fisher’s exact test was used to compare categorical variables, whereas the non-parametric Wilcoxon rank sum test was performed to compare continuous variables. The pathological response between the two groups was compared using either the chi-square test or Fisher’s exact test. Logistic regression was performed to identify independent factors associated with pCR or MPR. Survival curves were estimated using the Kaplan–Meier method. Hazard ratios (HRs) for event-free survival (EFS) were calculated using Cox-proportional hazard models. P-value was using the log-rank test. Receiver operation curves (ROC) and area under curve (AUC) were utilized to evaluate the predictive accuracy of the established model. Bootstrapping was employed to estimate the concordance index (C-index) for the models. Initially, the C-index is calculated for each model using observed event indicators and predicted values. Subsequently, bootstrapping is performed to resample the data 1000 times, and C-index values are calculated for each resampled dataset. Bootstrapping method was performed with the Hmisc package. All tests were two-sided, and a P value < 0.05 was considered to indicate statistical significance. All statistical analyses were conducted using R version 4.2.0 software.

Results

Patient characteristics and treatment safety
From December 18, 2020, to January 9, 2023, a total of 100 eligible patients at Peking University Cancer Hospital & Institute were enrolled in this study. The baseline clinical characteristics of the 100 patients are listed in Table 1. In total, 89 (89.0%) patients were male, and 86 (86.0%) patients had a smoking history (ever or current smoking). The median age was 63 years (range 45–76 years). Among the 100 patients, 19 (19.0%) had adenocarcinoma, and 77 (77.0%) had squamous cell carcinoma according to the pre-treatment biopsy. Notably, four patients were identified as having NSCLC, but the specific histological subtype could not be determined according to the pretreatment biopsy (termed NSCLC-other). Furthermore, 31 (31.0%), 46 (46.0%), and 23 (23.0%) patients had clinical stage IIB, IIIA, and IIIB disease, respectively. In addition, 85 (85%) patients underwent PD-L1 testing, and 65 (65.0%) patients had a PD-L1 TPS equal to or greater than 1% (PD-L1 positive), with 29 having a TPS greater than 49%.

During neoadjuvant treatment, 95 (95.0%) patients experienced TRAEs of any grade, and the incidence of individual TRAEs is shown in Additional file 2: Table S1. Among the 95 patients, 35% had grade 3 or 4 TRAEs. Hematological TRAEs were common and occurred in 78 (78.0%) patients. The most common TRAE was neutropenia, which occurred in 59 (59.0%) patients, 33 of whom were grades 3–4. Thirty-nine (39.0%) patients experienced non-hematological TRAEs of grades 1–2, of which nausea was the most common, occurring in 21 (21.0%) patients. Notably, 21 (21.0%) patients experienced immune-mediated TRAEs; grade 1–2 immune-mediated TRAEs occurred in 18 (18.0%) patients, and grade 3–4 immune-mediated TRAEs occurred in 3 (3.0%) patients, including one patient with enteritis and two patients with pneumonia.

Surgical outcomes and efficacy
Among the 100 patients, 83 (83.0%) patients ultimately underwent surgical resection (Fig. 1B). Of the 17 patients (17%) who did not proceed to surgery, 10 declined the procedure, 4 withdrew due to adverse events, and 3 experienced disease progression. Among them, 60 (72.3%) patients received 2 treatment cycles, 21 (25.3%) received 3 treatment cycles, and 2 (2.4%) received 4 treatment cycles. R0 resection was achieved in all patients, and they were included in the final efficacy analysis (Table 2). Surgical complications occurred in 21 (25.3%) patients, with the most common complications being pulmonary air leaks (9.6%) and pneumonia (7.2%). No postoperative deaths occurred within 30 to 90 days. According to the final pathological evaluation, 54 (65.1%) patients achieved MPR including 38 (45.8%) with pCR (Table 2, Fig. 2A). As of July 31, 2025, at a median follow-up time of 30.4 months (IQR 24.7–38.3), the median event-free survival (EFS) rate has not reached (Fig. 2B). The 1-year and 2-year EFS rates were 95.2% (95% CI 90.7–99.9%) and 86.4% (95%CI 79.2–94.2%) (Table 2). Notably, patients with MPR showed longer EFS than those with non-MPR (HR 0.19, 95% CI 0.06–0.54, P < 0.001; Fig. 2C), and this was also observed in pCR vs. non-PCR (HR 0.24, 95% CI 0.07–0.84, P = 0.015; Fig. 2D).

Clinical characteristics of patients included in the biomarker analysis
Among the 83 patients who underwent surgery, a total of 195 plasma samples were collected from 60 patients before each treatment cycle (C0, C1, C2, and C3) and prior to surgery (BS); paired WMS and TMS analysis were performed on all of these samples (Fig. 1C). These 60 patients constituted the biomarker analysis cohort, and the major clinical characteristics, including gender, smoking history, and clinical stage were not different from those of the whole study cohort of 100 patients (all P > 0.05, Table 1). Among these 60 patients, 40 (66.7%) patients received 2 treatment cycles, 18 (30.0%) received 3 treatment cycles, and 2 (3.3%) received 4 treatment cycles, and 53 underwent PD-L1 expression testing of biopsy tissue.
In the biomarker analysis cohort, 39 (65.0%) patients achieved MPR, and 29 (48.3%) patients achieved pCR (Additional file 3: Fig. S1A). The median %RVT of MPR group was 0.0% (range 0.0–5.0%), and that of non-MPR group was 70.0% (range 13.0–100.0%). The EFS of the biomarker analysis cohort is shown in Additional file 3: Fig. S1B.
We found that smokers were more likely to achieve MPR (nominal P = 0.006) and pCR (nominal P = 0.011, Additional file 2: TableS2). Patients with LUSC also had a higher MPR rate than those with LUAD (nominal P = 0.028), though the pCR difference was not significant (nominal P = 0.079). Additionally, patients with PD-L1 positive (TPS ≥ 1%) had a higher pCR rate (nominal P = 0.025) and a marginally significant higher MPR rate (nominal P = 0.054). A higher proportion of pretreatment blood CD8 + CD28- T lymphocytes was observed in MPR patients (nominal P = 0.045), suggesting a potential predictive biomarker, though this difference was not significant for pCR (nominal P = 0.120).

Baseline cfDNA feature analyses
For the 60 patients in the biomarker analysis cohort, paired TMS-based MFR score and WMS-based CAFF score were computed for each of the collected plasma samples (Fig. 1C). Notably, three patients had no blood collection at C1, and four patients who received 3 treatment cycles had no blood collection at C2. We investigated whether specific cfDNA features could predict efficacy of neoadjuvant chemoimmunotherapy in patients with NSCLC. We initially examined both MFR and CAFF scores at baseline (C0) to determine whether they were associated with pathological response. However, there was no significant difference in either the MFR score or CAFF score between the MPR and non-MPR groups at C0 (Fig. 3A), suggesting that the baseline MFR and CAFF scores did not have predictive value for pathological response. However, it is noteworthy that the MFR and CAFF scores were both significantly correlated with clinical stage (Fig. 3B), and they were both positively correlated with baseline tumor size (Fig. 3C). These results suggested that they could serve as staging-related biomarkers reflecting tumor progression and burden. Additionally, both MFR and CAFF scores retained limited associations with other clinical parameters (Additional file 2: Table S3).
Previous studies have shown that the methylation status of specific genes or regions is associated with the efficacy of immunotherapy in late-stage solid tumors, including NSCLC and melanoma [30–32]. Therefore, we sought to explore whether specific methylated regions of cfDNA were associated with pathological response. Through analysis, we identified a total of 18 differentially methylated regions (DMRs) between MPR and non-MPR groups. The methylation levels of these DMRs were significantly lower (median: 0 vs. 0.0016, P < 0.001) in MPR group than in non-MPR group, indicating that hypomethylation of these regions may be associated with better clinical outcomes (Fig. 4A), and these DMRs were predominantly located in the promoter regions of DNA sequences, followed by introns and 5′ UTRs (Fig. 4B). To further pinpoint the genes associated with these DMRs, we identified some hypomethylated genes, such as GATA2, HOXB3, and MYADM, and other corresponding hypermethylated genes, including LHX8, H4C4, CGAS, and TNFRSF10D (Fig. 4C).
We next investigated DMRs between pCR and non-pCR groups and identified 26 DMRs, six of which overlapped with those identified between MPR and non-MPR groups (Additional file 3: Fig. S2A). Similar to our observations between MPR and non-MPR groups, a significantly lower average methylation level was observed in pCR group than in non-pCR group (median 0 vs. 0.0022, P < 0.001, Additional file 3: Fig. S2B). Once again, the most commonly situated sites within the DNA sequence for these DMRs were in the promoter regions (Additional file 3: Fig. S2C). The six shared genes involved in these DMRs—C16orf86, GATA2, GDF6, LHX8, PAX1, and ZNF570—all exhibited hypomethylation (Additional file 3: Fig. S2D) and could serve as indicators for better pathological response.

cfDNA features during and after treatment in association with MPR
Next, we assessed whether cfDNA features could serve as dynamic monitoring biomarkers for therapeutic efficacy. Because the majority of patient samples were collected from C1 (N = 57) and BS (N = 60), with fewer samples from C2 (N = 16) and C3 (N = 2) (Fig. 1B), we selected C0, C1, and BS for subsequent analyses.
For MPR patients, both MFR and CAFF scores significantly decreased from C0 to C1 (P < 0.001) and remained low at BS (P < 0.001) (Fig. 5A, B). In contrast, non-MPR patients presented a different dynamic pattern of cfDNA features. Their MFR scores similarly decreased at C1 (P < 0.05) but slightly increased at BS, with no significant difference between BS and C0 time points (P = 0.132, Fig. 5A). A similar dynamic pattern was observed for the CAFF score in the non-MPR group, with CAFF score exhibiting no significant reduction at C1 compared to C0 (P = 0.111, Fig. 5B).
To facilitate the application of these two features in clinical practice, we classified patients into two categories (high status and low status) at each time point using pre-defined cut-off values for MFR and CAFF scores [28, 29] (Additional file 2:Table S4). Although MFR status at C0 and C1 was not associated with MPR, patients with low MFR status at BS were more likely to achieve MPR than those with high MFR status (74.5% vs. 11.1%, P < 0.001, Fig. 5C). For CAFF, the low status at C1 and BS both indicated a higher likelihood of MPR (P = 0.031 and 0.008, respectively, Fig. 5D).
For the 57 patients whose plasma samples were collected at all three time points, their MFR status (low or high) at C0, C1, and BS is illustrated by a Sankey diagram (Fig. 5E). Notably, according to their MFR status at C0 and C1, these patients were classified into three patterns: the “C0 low” pattern with low MFR status at C0 (N = 21), the “C0 high/C1 low” pattern where the MFR status was initially high at C0 but turned into low at C1 (N = 27), and the “C0 high/C1 high” pattern with high MFR status at both C0 and C1 (N = 9). Interestingly, the MPR rate gradually decreased from the “C0 low” pattern (71.4%, 15/21) to the “C0 high/C1 low” pattern (66.7%, 18/27) and to the “C0 high/C1 high” pattern (55.6%, 5/9). According to the MFR status at BS, these three patterns were further divided into different subgroups (BS low and BS high), and patients with BS low status all showed significantly higher MPR rates than those with BS high status in these three patterns (the “C0 low” pattern: BS low vs. high 78.9% vs. 0%; the “C0 high/C1 low” pattern: 73.9%vs 25%; the “C0 high/C1 high” pattern: 83.3% vs. 0%) as shown in Fig. 5F. Notably, patients with high MFR score at BS presented a greater %RVT in the final pathological evaluation compared to patients with low MFR status (Additional file 3: Fig. S3A).
For CAFF, a similar dynamic pattern was observed (Fig. 5G). Using a classification strategy similar to that used for MFR, 26 patients were classified into the “C0 low” pattern, 25 into the “C0 high/C1 low” pattern, and 6 into the “C0 high/C1 high” pattern. The MPR rates for these three patterns were 69.2% (18/26), 72.0% (18/25), and 33.3% (2/6), respectively. Among these three patterns, low CAFF status at BS was all associated with a higher MPR rate (Fig. 5H), and patients with high CAFF status at BS also presented a greater %RVT in the final pathological evaluation compared to patients with low CAFF status (Additional file 3: Fig. S3B).
Moreover, among patients with MFR clearance (C0 high/C1low/BS low), 73.9% (17/23) achieved MPR, which was significantly higher than the rate among those without clearance (14.3%, 1/7). Similarly, the MPR rate was significantly higher in patients with CAFF clearance (86.7%, 13/15) compared with those without clearance (35.7%, 5/14) (Additional file 3: Fig. S4A).

cfDNA features during and after treatment in association with pCR
Next, we analyzed whether the dynamic patterns of cfDNA could also predict pCR. Both the MFR and CAFF scores significantly decreased from C0 to C1 (P < 0.01) and remained low at BS (P < 0.05) in the pCR group (Additional file 3: Fig. S5A-B). In addition, low MFR and CAFF status at BS were both associated with higher pCR rate (P = 0.027 and 0.050, respectively, Additional file 3: Fig.S5C-D). The low MFR and CAFF status at BS were also associated with a higher pCR rate in all three patterns (Additional file 3: Fig. S5E-H). Furthermore, the percentages of patients with pCR among those with MFR clearance and CAFF clearance were 52.2% (12/23) and 53.3% (8/15), respectively (Additional file 3: Fig. S4B).

Combined longitudinal cfDNA metrics and immune parameters to predict MPR
In this study, we chose MPR as the primary endpoint. We next explored whether an integrated model involving cfDNA features at multiple time points could predict MPR more accurately. As shown above, the baseline MFR and CAFF demonstrated no predictive value for pathological response and were therefore not included in the model.
We first evaluated the performance of the models including the MFR and CAFF status at different time points. The model based on the cfDNA feature (MFR plus CAFF status) at both C1 and BS exhibited superior performance with an AUC of 0.79 (95% CI 0.72–0.87) (Fig. 6A and Additional file 3: Fig. S6A). In order to improve the accuracy of MPR prediction, two immune parameters (PD-L1 and CD8 + CD28- T lymphocytes) identified in this study as being associated with MPR were also integrated into the final model alongside cfDNA feature (MFR and CAFF status) at both C1 and BS. The final multivariate logistic regression model demonstrated a significantly greater AUC of 0.86 (95% CI 0.81–0.91) compared to the model with immune parameters alone (AUC = 0.67, 95% CI 0.59–0.76), with a P value = 0.013. (Fig. 6B and Additional file 3: Fig. S6B). MFR at BS time point and CD8 + CD28- T lymphocytes remained as significant biomarkers (P = 0.025 and P = 0.020, respectively) in the final model (Additional file 2: Table S5) and with contributions of 33.37% and 31.1%, respectively (Fig. 6C). Altogether, these findings indicated that the integrating immune parameters and cfDNA feature can achieve better MPR prediction.

Discussion
This is the first prospective study exploring the dynamic changes of cfDNA features and their effects on pathological response prediction in a phase II trial of neoadjuvant chemoimmunotherapy in a locally advanced NSCLC cohort. A total of 100 patients were enrolled, and 83 patients underwent surgery yielding a 100% R0 resection rate. Pathological outcomes were 65.1% (54/83) MPR rate and 45.8% (38/83) pCR rate. Within a median follow-up of 30.4 months, the 1-year and 2-year EFS rates were 95.2 and 86.4% respectively. Notably, we observed a significant correlation between dynamic changes of cfDNA methylation level (MFR) and chromosomal aneuploidy (CAFF) and the pathological response to neoadjuvant chemoimmunotherapy. An integrated prediction model leveraging both baseline immune parameters and dynamic cfDNA methylation features showed certain MPR prediction accuracy.
The MPR and pCR rates observed in this study are within the range of several phase II trials of neoadjuvant chemoimmunotherapy based on resected stage IB to III NSCLC patients (MPR rate 26.9–82.9%, pCR rate 18.2–63.4%) [33, 34]. Several international multi-center phase III trials, including CheckMate-816 and KEYNOTE-671, not only confirmed a significant improvement in pathological response rate but also demonstrated significantly improved EFS and overall survival compared with chemotherapy alone [4, 5]. Two multi-center phase III trials from China (RATIONALE-315 and Neotorch) also reported similar findings [35, 36]. The Neotorch trial, a phase III study of toripalimab combined with chemotherapy in stage III NSCLC, reported an MPR rate of 48.5%, a pCR rate of 24.8%, a 1-year EFS rate of 84.4%, and a 2-year EFS rate of 64.7%. Although the same PD-1 antibody was used in both studies, the higher pathological response and event-free survival (EFS) rates observed in our trial compared to the Neotorch trial may be explained by differences in patient staging. Neotorch’s interim analysis included only stage III patients [35], whereas our study population consisted of 31% stage II patients. Additionally, although our study is a phase II trial, compared to the currently reported results of the Neotorch trial, our study paid more attention to pathological response and have reported more detailed and in-depth biomarker exploration findings. In summary, all these studies confirm the efficacy of neoadjuvant chemoimmunotherapy in locally advanced NSCLC.
Although neoadjuvant chemoimmunotherapy has shown superior efficacy compared to chemotherapy alone, a significant proportion of patients could not benefit from this approach [37, 38], suggesting the urgency of identifying reliable risk factors or biomarkers for early response prediction and dynamic monitoring to optimize treatment. Tumor-informed ctDNA is limited in neoadjuvant settings due to the unavailability of primary tumor samples [39]. Beyond genetics, non-genetic information like DNA methylation offers additional value. Several studies have demonstrated that cfDNA methylation signatures exhibit superior efficiency in multi-cancer detection, localization, and subtyping, and are developed for MRD monitoring [17, 40, 41]. For instance, a pan-cancer screening study demonstrated that the integrated model (THEMIS) utilizing multimodal cfDNA features derived from WMS, including methylation level and fragmentation, exhibited higher accuracy than each indicator alone [29]. However, the application potential of cfDNA methylation in therapeutic response prediction and effect monitoring has not been studied thoroughly.
In this study, we simultaneously performed deep targeted methylation panel sequencing and WMS to calculate two-dimensional features of cfDNA, namely, MFR and CAFF, respectively. Our findings indicate that both MFR and CAFF could serve as potential indicators of pathological response to neoadjuvant chemoimmunotherapy and complement each other to some degree. MFR is the percentage of fully methylated fragments within the DNA fragment being analyzed and represents the methylation level of cfDNA [29]. CAFF is a common genetic alteration that has been used for cancer detection and to quantify copy number changes in chromosome arms. One previous study demonstrated that chromosome aneuploidy outperformed somatic mutation and focal copy number deletions/amplifications in predicting chemosensitivity in multiple cancers [42]. A significant advantage of these two parameters is that they can be expressed as numerical magnitudes, which may fully capture the essence of dynamic changes of plasma biomarkers during treatment. In this study, we found that baseline MFR and CAFF scores (at the C0 time point) were strongly correlated with clinical stage and tumor lesion size, suggesting their potential associations with tumor burden, and the changes of these biomarkers during treatment may predict therapeutic effects theoretically. During treatment, we observed that both the MFR and CAFF scores tended to decrease rapidly at the beginning of treatment and remained at low levels in the MPR group compared with those in the non-MPR group.
In studies focusing on perioperative ctDNA evaluation and its prognostic relevance, the preoperative ctDNA detection rate was approximately 50% in stage II-III NSCLC patients [12, 43]. Recent neoadjuvant immunotherapy trials also indicated that less than 50% of the enrolled patients presented with evaluable ctDNA levels [5, 44]. Therefore, dynamic evaluation of the treatment response of patients with undetectable or unevaluable ctDNA levels remains a major challenge. In the present study, 195 plasma samples were collected from 60 patients, and paired WMS and TMS analyses were successfully performed on all collected samples, resulting in a 100% detection rate for cfDNA methylation. Binary variables based on numerical scores were used to represent the longitudinal dynamic process throughout the entire neoadjuvant treatment period. According to the status of MFR and CAFF (high or low) at C0 and C1, we identified three dynamic change patterns: the “C0 low” pattern, the “C0 high/C1 low” pattern, and the “C0 high/C1 high” pattern. This classification might fully reflect the landscape information of the whole treatment group, particularly those patients in “C0 low” pattern. Patients with the “C0 low” pattern initially exhibited low MFR or CAFF, but some individuals turned into high at C1 or BS. For instance, two patients with low MFR at C0 turned into high MFR at BS; notably, both of them showed non-MPR at the final pathological evaluation. This result highlights that dynamic monitoring is still needed for those with initial low MFR or CAFF. Compared with the other two patterns, the MPR rate of the “C0 high/C1 high” pattern was lower, indicating potential compromised treatment efficacy in subsequent treatment (C1) and potential drug resistance following the initial treatment cycle in this pattern; alternative treatment strategies may need to be considered for these patients. Moreover, MPR rates were much higher in the patient group with MFR or CAFF clearance (C0 high/C1low/BS low) compared to those without clearance.
Furthermore, the status of the MFR or CAFF at BS was found to be correlated with the %RVT in this study, which provides a novel analysis scope for response assessment. Patients with high MFR or CAFF showed significantly higher %RVT compared to those with low MFR or CAFF. A recent report from the CheckMate 816 trial revealed that the %RVT might best approximate EFS, indicating that the dynamic changes of MFR and CAFF have the potential to predict not only MPR to neoadjuvant chemoimmunotherapy but also the degree of pathological regression. The role of these dynamic patterns in EFS prediction needs further prospective validation in locally advanced NSCLC.
When we evaluated the performance of two cfDNA features (MFR and CAFF) in predicting MPR at different time points, the combination of the C1 and BS time points (AUC = 0.79, 95%CI 0.72–0.87) was more accurate than C1 (AUC = 0.64, 95%CI 0.58–0.70) or BS (AUC = 0.72, 95%CI 0.66–0.78) alone. This finding highlights the importance of dynamic monitoring of cfDNA methylation during neoadjuvant chemoimmunotherapy again. To predict pathological response more accurately, we incorporated immune parameters (PD-L1 expression and CD8 + CD28 − T lymphocyte levels) with dynamic cfDNA features. The preliminary integrative model achieved a relatively high predictive accuracy compared to individual parameters, with an AUC of 0.86, indicating promising potential for practical application.
Several limitations of our study should be acknowledged. First, although this exploratory investigation was conducted with 60 patients, the small sample size and predominance of squamous cell carcinoma subtypes might cause potential bias in this cohort. After thorough comparison of the major clinical characteristics, biomarker analysis population was not different from the whole clinical trial cohort. In addition, the use of 195 longitudinal samples and two detecting dimensions for each sample (paired TMS-based MFR score and WMS-based CAFF score) may further improve research scope. However, the optimal predictive score or dynamic patterns of MFR or CAFF need to be validated in larger prospective cohorts. Second, the relatively short follow-up period limited our ability to conduct survival analysis, and longer follow-up periods are needed to assess the prognostic effect of MFR and CAFF.

Conclusions
Neoadjuvant chemoimmunotherapy could achieve an optimal response rate. Dynamic monitoring of cfDNA methylation has the potential to predict the pathological response of NSCLC patients to neoadjuvant chemoimmunotherapy.

Supplementary Information