Perioperative ctDNA-Based Molecular Residual Disease Detection for Non–Small Cell Lung Cancer: A Prospective Multicenter Cohort Study (LUNGCA-1)

Surgical resection remains the foremost treatment for localized non–small cell lung cancer (NSCLC), but a substantial number of patients still experience local recurrence or distant metastasis after surgery (1, 2). Presently, prognostic stratification of NSCLC is mainly based on clinicopathologic parameters, such as TNM stage, airway spread, and pathologic subtype, etc. (3). These population-level indicators are known to have limited effectiveness, as a significant proportion of low-risk patients based on these parameters still relapse, whereas some high-risk patients remain disease-free after a significant time period. These indicators also have limited efficacy in selecting patients for adjuvant therapies. According to the National Comprehensive Cancer Network guidelines, adjuvant chemotherapy is recommended for patients with stage II–IIIA and high-risk stage IB NSCLC patients after radical surgery, which however has shown only 5% improvement in the 5-year overall survival rate (4–7). Therefore, more effective approaches are in urgent need to identify patients at high risk of relapse and/or with potential to benefit from adjuvant therapies.

In recent years, circulating tumor DNA (ctDNA) sequencing has emerged as a noninvasive method for early diagnosis, prognostic stratification, disease surveillance, and treatment response evaluation of different cancer types (8–15). It has also shown promises in monitoring and tracking of personalized, disease-related markers at the molecular level, a novel application termed molecular residual disease (MRD) testing (11, 16, 17). In the TRACERx landmark study (11), the presence of ctDNA identified patients who were likely to experience recurrence after surgery prior to traditional radiologic methods. Chaudhuri and colleagues (17) subsequently showed that patients with posttreatment-detected ctDNA advanced to radiographically observable disease progression by a median of several months. These studies have shed light on the prognosis prediction value of postoperative ctDNA. However, limited studies exist on whether MRD could be effectively detected by ctDNA in the perioperative period (18). The perioperative period, commonly referred to as the timeframe from a few days before to several days–weeks after surgery (19), is precious for decision-making about adjuvant therapies and follow-up strategies, because adjuvant therapies are usually initiated within two months postoperatively (6, 20). Thus, accurate and early identification of high-risk patients suitable for adjuvant therapies during the perioperative period has great value, based on the fundamental principle that early administration of adjuvant treatments would be more curative when the volume of the metastatic disease is small.

We are conducting a large-scale multicenter, prospective observational cohort study named LUNGCA (ClinicalTrials.gov identifier: NCT03317080) for dynamic monitoring of ctDNA in surgical patients with lung cancer. The LUNGCA study systematically evaluates patients’ ctDNA during the perioperative period (before surgery during their standard checkups, postoperative 3 days before release from hospital, and 1 month during their first post-surgery follow-up visit) and then monitors ctDNA every 3–6 months until 3 years after surgery. From September 2017 to May 2020, 426 patients have been enrolled in the LUNGCA cohort. Although the monitoring study is ongoing, we retrospectively analyzed a part of the cohort consisting of 950 plasma samples of 330 stage I–III NSCLC patients collected at three perioperative time points (LUNGCA-1). Our analysis mainly focused on the efficacy of perioperative ctDNA testing for MRD detection and postoperative relapse prediction. Although we are not able to change the courses of a particular patient's care in our study due to the observational design, analytical results reported herein may provide guidance as to how clinical practice could be improved and personalized according to a patient's molecular profile during the perioperative period.

LUNGCA is a prospective observational study aimed to analyze the value of ctDNA monitoring in lung cancer patients for postoperative evaluation, therapy response assessment, relapse prediction, and molecular phenotype categorization. Patients with suspected lung cancer who underwent curative-intent surgery since September 2017 at West China Hospital, Chengdu Shangjinnanfu Hospital, and Sichuan Provincial People's Hospital were consecutively recruited into the LUNGCA cohort. Blood samples before and after surgery at different time points were collected for ctDNA analysis. Fresh frozen or paraffin-embedded tissues of primary cancer were also collected. The clinical follow-up strategy mainly involved radiologic surveillance (computed tomography/magnetic resonance imaging) every 3–6 months. The LUNGCA study was conducted in accordance with the Declaration of Helsinki, and was approved by the institutional review boards of the participating hospitals. Written informed consent was obtained from each patient.

The cohort evaluated in this article (LUNGCA-1) was enrolled from September 2017 to May 2020, as a part of the LUNGCA study. Patients between 18 and 80 years of age and with pathologic stage I–III NSCLC (AJCC 8th) were eligible for inclusion. Patients with multiple primary lung cancers, or pathologic stage IV disease, or non-NSCLC histology, or with a history of malignancy in the past 5 years were excluded from the study. The samples analyzed were intraoperative tumor tissues and blood samples obtained before surgery, as well as blood samples at 3 days (2–15 days in reality) and 1 month (3–6 weeks in reality) after surgery.

DNA isolation, library preparation, and next-generation sequencing (NGS) were conducted at Genecast Biotechnology Co., Ltd. An NGS panel spanning 769 cancer-related genes (Supplementary Table S2) was used for hybridization-based sequencing. Bioinformatics tools used for processing NGS data included Trimmomatic (v0.36; ref. 21) for quality trimming, VarDict (v1.5.1), FreeBayes (v1.2.0; ref. 22) for point mutation calling, CNVkit (v0.9.2) for copy-number variation (CNV) calling, as well as FACTERA v1.4.4 and FusionMap (23, 24) for gene fusion calling. A tumor-informed strategy was adopted for mutation analysis in plasma samples. For SNV/InDels, a statistical test was performed on each candidate ctDNA mutation against an internal background reference library to eliminate technical artifacts. A mutation in the plasma sample was defined as positive if it had a P value < 0.01. A sample-level combined P value was further calculated, and a threshold of P < 0.01 on the combined P value was applied to call a plasma sample positive. For CNVs, a plasma sample was considered CNV positive only if a same gene-level copy-number alteration (gain or loss) was also detected in the corresponding, originally resected tissue. Finally, for fusions, a plasma sample was defined as positive if at least three unique sequence reads spanned the breakpoint and exactly matched to the reference derived from the fusion genes in its paired tissue sample. Detailed laboratory and data analysis procedures are included in the Supplementary Methods.

The primary outcome was recurrence-free survival (RFS), defined as the time interval from the surgery to the first verified recurrence (local or distant) or death for any cause. Independency between the ctDNA-positive/negative subgroups and the clinical features were assessed using the Fisher exact test. The overall relapse rates between ctDNA-positive and -negative subgroups were compared with the Fisher exact test. R packages survival and survminer were used for prognostic analysis. Chi-square proportion calculated by R package rms was used to assess the relative contribution of each variable to survival risk. Statistical significance is defined with P < 0.05.

The raw data set has been deposited to the Genome Sequence Archive in the National Genomics Data Center of China, under project number PRJCA003692 and accession number HRA000430 that are accessible at https://ngdc.cncb.ac.cn/gsa-human/.

A total of 426 lung cancer patients treated with curative-intent surgery were initially recruited for the LUNGCA cohort from September 2017 to May 2020 (Fig. 1). Among them, 73 patients who had either history of malignancy in the past 5 years, multiple primary lung cancers, pathologic stage IV disease, or non-NSCLC histology were excluded, leaving 353 patients with stage I–III NSCLC (AJCC 8th) enrolled for sequencing analysis on tumor and blood samples. Additional 23 patients with no trackable mutation detected with the target NGS panel in their tumor tissues were excluded from downstream analysis. Finally, a cohort of 330 evaluable patients (LUNGCA-1) with at least one detected somatic mutation in their tumor tissues were included in subsequent ctDNA analysis, comprising of a total of 950 plasma samples with 330, 296, and 324 at preoperative baseline, 3 days and 1 month after surgery, respectively (Fig. 1).

We first summarized the demographic characteristics, mutations in tumor tissues, and ctDNA detection status of these 330 patients (Fig. 2; Supplementary Table S3–S5). The average age at diagnosis was 59 years (range, 28–80 years). Of all patients, 51.2% were female, 84.8% had lung adenocarcinoma (LUAD), 67.0% had stage I disease, and 39.1% received adjuvant therapies. The follow-up period of all patients ranged from 341 to 1,340 days with a median of 1,068 days. By their last follow-up, 70 patients (21.2%) relapsed, whereas the rest 260 (78.8%) remained disease-free.

Somatic mutations detected in tumor tissues were primarily point mutations, amounting to 1,867 (90.28%) of all variants detected, whereas the rest 181 CNVs and 20 gene fusions only contributed to a small proportion (9.72%; Supplementary Fig. S1). Frequently mutated driver genes in tumor tissues included EGFR (64%), TP53 (49%), RBM10 (17%), CDKN2A (9%), PIK3CA (9%), KMT2D (6%), RB1 (6%), ERBB2 (5%), ALK (5%), and KRAS (5%; Fig. 2). Based on the baseline variant sets identified in patient tissue samples and using the mutation tracking paradigm in the Patients and Methods section, ctDNA mutations were detected in 69 of 330 patients (20.9%) at preoperative baseline, 19 of 296 (6.4%) at 3 days after surgery, and 19 of 324 (5.9%) at 1 month after surgery, respectively (Fig. 2). As expected, TP53 and EGFR were the most frequently mutated genes in pre- and postoperative ctDNA analyses (Supplementary Fig. S2), and in particular, were detected in 41 (12.4%) and 25 (7.6%) preoperative plasma samples, respectively.

Several clinicopathologic variables were statistically associated with the presence of ctDNA in baseline plasma before surgery (Supplementary Table S6). Specifically, patients with lung squamous cell carcinoma (LUSC), tumor diameters larger than 3 cm, and at pathologic stages II and III were prone to be ctDNA positive before surgery. Strikingly, only the pathologic stage was significantly associated with ctDNA status across all three perioperative time points.

The prognostic value of ctDNA before treatment remains controversial in different cancers (17, 25–27); thus, we examined the association of preoperative ctDNA status with RFS in NSCLC. Our data showed that 46.4% (32 of 69) of patients with positive preoperative ctDNA experienced postoperative relapse, compared with 14.6% (38 of 261) ctDNA-negative patients (Fig. 3A, P < 0.001). Moreover, ctDNA-positive patients at the preoperative period had significantly worse RFS than ctDNA-negative ones (HR 4.2; 95% CI, 2.6–6.7; P < 0.001; Fig. 3B). After adjusting for clinicopathologic variables, preoperative ctDNA status remained an independent risk factor for RFS (HR 2.6; 95% CI, 1.3–5.1; P = 0.005; Table 1).

The higher positive rate of preoperative ctDNA in LUSC (74.4%) than in LUAD (11.8%) in our study is consistent with the earlier described “shedding hypothesis” (11, 28). We investigated whether the increased ctDNA leakage before surgery in LUSC would translate into a worse prognosis in the current cohort. Interestingly, the preoperative ctDNA status was a robust prognostic factor for RFS in LUAD (HR 6.0; 95% CI, 3.4–10.6; P < 0.001; Supplementary Fig. S3A), whereas its association with RFS in LUSC is insignificant (HR 2.4; 95% CI, 0.5–10.7; P = 0.241; Supplementary Fig. S3B).

We next assessed the prognostic value of postoperative ctDNA. The Kaplan–Meier estimates showed that the positivity of ctDNA at postoperative 3 days was of high predictive value for relapse (HR 8.6; 95% CI, 4.7–15.6; P < 0.001; Supplementary Fig. S4A). A similar result was found in the analysis of ctDNA at postoperative 1 month (HR 14.3; 95% CI, 7.9–25.9; P < 0.001; Supplementary Fig. S4B). In multivariate analysis including clinicopathologic risk factors, ctDNA status at postoperative 3 days, as well as ctDNA status at postoperative 1 month were still independent factors for RFS (both P < 0.001; Supplementary Table S7).

Based on these findings, we further evaluated the efficacy of postoperative ctDNA-based MRD detection for relapse prediction. “MRD positive” was defined for patients with detectable ctDNA at 3 days and/or 1 month postoperatively, whereas patients with no detectable ctDNA at any postoperative time point were defined as MRD negative. A total of 329 patients with at least one available plasma sample at these two time points were included in the subsequent analysis. Of those, 26 patients were MRD positive, whereas the remaining 303 patients were negative. During the follow-up period, MRD-positive patients had a significantly higher overall recurrence rate (80.8%, 21 of 26 patients) than those who were MRD negative (16.2%, 49 of 303 patients; Fig. 3C, P < 0.001). Additionally, compared with preoperative ctDNA status, MRD status exhibited stronger prediction power for RFS (HR 11.1; 95% CI, 6.5–19.0; P < 0.001; Fig. 3D).

Besides MRD status, clinicopathologic variables including age, tumor size, histologic subtype, TNM stage and adjuvant therapies were also significantly associated with RFS in univariate Cox analysis (Table 1). After adjusting for those clinicopathologic variables in a multivariate Cox proportional hazard regression model, MRD status remained as an independent risk factor for RFS (HR 8.6; 95% CI, 4.8–15.3; P < 0.001; Table 1). Based on each variable's relative contribution to RFS prediction, MRD status was the most important factor, even stronger than the combination of all clinicopathologic variables including the TNM stage (Fig. 3E).

Considering that preoperative ctDNA was only prognostic in LUAD but not in LUSC, we asked whether this difference would also exist in survival analysis of ctDNA-based MRD. Interestingly, we found that MRD status was a strong prognostic biomarker for both LUAD (HR 14.1; 95% CI, 7.7–25.9; P < 0.001) and LUSC (HR 4.9; 95% CI, 1.5–16.7; P = 0.005; Supplementary Fig. S5A and S5B), a different observation from that of preoperative ctDNA status.

EGFR is one of the most common driver genes in LUAD, and its mutations are associated with favorable prognosis in LUAD (29). In our cohort, up to 71.4% of LUAD patients harbored pathogenic EGFR mutations (see Supplementary Table S8 for the mutation list) in their primary tumor tissues. Therefore, the exploration of plasma prognostic biomarkers for EGFR-mutated LUAD would have great clinical significance. We performed subgroup analysis in EGFR-mutated LUAD patients and EGFR wild-type cases, and found that ctDNA-based MRD positivity was associated with poor RFS in both subgroups [EGFR-mutated (HR 11.4; 95% CI, 5.4–24.1; P < 0.001), EGFR wild-type (HR 20.2; 95% CI, 6.6–62.3; P < 0.001); Supplementary Fig. S5C and S5D).

The prognostic value of ctDNA-based MRD in different stages of NSCLC was further assessed. The Kaplan–Meier estimates showed that MRD positivity was significantly associated with worse RFS in both stage I (HR 18.0; 95% CI, 7.0–46.0; P < 0.001) and stage II–III NSCLC (HR 5.5; 95% CI, 2.9–10.7; P < 0.001; Supplementary Fig. S5E and S5F). Collectively, our results suggest that ctDNA-based MRD status is a strong predictor for NSCLC relapse independent of pathologic subtype, EGFR mutation status, and TNM stage.

In a standard clinical setting, stage II–III NSCLC patients and stage IB NSCLC patients with high risk based on clinicopathologic factors would commonly receive adjuvant therapies after radical surgery, which however yielded improved survival in only a small fraction (5, 30–32). In our study, we explored the possibility of ctDNA-based MRD detection in aiding the selection of eligible NSCLC patients for adjuvant therapies. Among the 26 MRD-positive patients, five of 17 patients who received adjuvant therapies remained relapse-free, whereas all nine patients not receiving adjuvant therapy experienced relapse (Fig. 4A). Kaplan–Meier estimates showed that the MRD-positive patients with adjuvant therapies had significantly better RFS than those not receiving adjuvant therapy (median RFS: 574 days vs. 315 days; HR 0.3; 95% CI, 0.1–0.8; P = 0.008; Fig. 4B). By contrast, patients receiving adjuvant therapies had significantly shorter RFS than those not receiving adjuvant therapy in the MRD-negative population (median RFS not reached for both groups; HR 3.1; 95% CI, 1.7–5.5; P < 0.001; Fig. 4B), probably due to the inclination of adjuvant therapy in more advanced stages of patients and patients with higher clinicopathologic risks. After adjusting for other clinicopathologic variables, adjuvant therapy remained to be an independent favorable factor for RFS in the MRD-positive population (HR 0.2; 95% CI, 0.1–0.6; P = 0.002) but not in the MRD-negative population (HR 1.6; 95% CI, 0.7–3.6; P = 0.283; Supplementary Table S9).

Since adjuvant therapy is generally not recommended for stage I patients, we further zoomed in to analyze the patients with stages II and III diseases. The result showed that adjuvant therapies could significantly improve RFS in MRD-positive patients in this subpopulation (HR 0.1; 95% CI, 0.03–0.4; P < 0.001) but not in their MRD-negative counterparts (HR, 1.0; 95% CI, 0.2–4.1; P = 0.961; Supplementary Fig. S6). Taken together, despite the nonrandomizing nature of this study, these findings indicate that adjuvant therapies may be more suitable for MRD-positive NSCLC patients.

Various approaches have been developed for MRD detection in multiple types of solid tumors, each with its own advantages and limitations. The “tumor-naïve approach” adopted by Parikh and colleagues (33) included comprehensive genomic and epigenetic analyses of plasma samples to maximize detection sensitivity and did not rely on tumor sequencing, but thus may inherit limited specificity and positive predictive value due to biological noises, whereas the remaining majority of the MRD analysis methods incorporated a personalized “tumor-informed approach” in the sense that mutation detection in plasma is guided by the baseline mutation profile through tumor sequencing. The latter would improve the detection specificity but may suffer from limited sensitivity on novel mutations in the plasma (11, 17, 34). Although Chaudhuri and colleagues (17) and Tie and colleagues (34) used prefixed panels for targeted sequencing in both tumor tissue and plasma samples, Abbosh and colleagues adopted a different approach by profiling tumor tissue variants across the whole-exome region followed by variant tracking in plasma with a focused patient-specific multiplexed PCR panel (11). The former fixed-panel strategy offers fast and robust MRD assays, but its sensitivity is greatly limited by the number of available mutations for tracking within the panel region. The latter approach ensures tracking of a fixed number of mutations for each patient with reduced cost for ultradeep sequencing, which may greatly benefit recurring tests in continuous MRD monitoring; however, its turnaround time for initial panel design and assay validation hinders its use in MRD analysis during the perioperative period during which the decision for adjuvant therapy needs to be made. Our study adopted the more widely used tumor-informed strategy to ensure specificity. Multiple types of somatic mutations including SNV/indel, CNV, and gene fusion detected in a patient's tumor tissues were tracked in respective plasma samples in a personalized manner. The MRD assay used in our study followed the fixed panel strategy with a 2.4 Mb panel design to cater most of NSCLC patients, but its sensitivity is limited by the depth of sequencing for plasma samples. Optimization of our panel regions and testing parameters remains a continuous future work.

Although ctDNA sequencing has been a mature technology for comprehensive cancer profiling, sensitive and accurate ctDNA mutation calling has always been a challenge. We have made several incremental improvements in our data analysis over existing studies. On top of the variant-level probability test based on position- and subtype-specific technical noise distribution for the confident calling of individual variants, a combined P-value approach was adopted for determining the sample-level ctDNA positivity to offset the potential loss of specificity resulting from increased numbers of traceable mutations in patients with higher mutation load. The statistical tests based on a large background reference library and combined P value eliminate the caller's dependency on a fixed set of cutoff thresholds and allows the caller to adopt to the different levels of background technical noises in different genomic regions and hence improves the mutation detection sensitivity. This strategy was shown effective by the 81% relapse rate in the MRD-positive group in the entire cohort, and the 100% relapse rate in the MRD-positive group in patients with no adjuvant therapy.

Based on the aforementioned incrementally improved analytical approach, we analyzed 950 perioperative plasma samples from 330 patients with stage I–III NSCLC and explored the efficacy of perioperative ctDNA for MRD detection. A majority of patients in our cohort were at stage I (67.0%), which differs from most previous studies which mainly recruited stage II and III patients (17, 18). Our cohort may more closely reflect the real-world stage distribution of NSCLC patients who received surgical treatment, and hence our study becomes incrementally more representative of current clinical practice (35, 36). Several key findings through our analysis of perioperative ctDNA samples include: (i) ctDNA could serve as a prognostic biomarker for NSCLC before surgery; (ii) relapse for NSCLC could be predicted through ctDNA-based MRD detection as early as in the perioperative period; (iii) ctDNA-based MRD outperforms clinicopathologic variables including TNM stage for RFS prediction; (iv) MRD-positive patients but not MRD-negative patients would benefit from adjuvant therapies. These findings highlight the important value of ctDNA-based MRD detection during the perioperative period for prognostic stratification, recurrence prediction, and postoperative management of NSCLC.

Based on preoperative plasma samples, we have confirmed that ctDNA status was significantly associated with RFS in NSCLC, which is consistent with the previous study by Chabon and colleagues (25). However, through subgroup analysis, our data are probably among the first to show that preoperative ctDNA positivity is a prognostic biomarker only for LUAD but not for LUSC. This probably reflects the biological discrepancy between the two NSCLC subtypes, although the underlying mechanism is unknown. LUSCs usually exhibit more tissue necrosis than LUADs, and hence are more prone to shedding ctDNA passively into the bloodstream (37, 38). This phenomenon was supported by the significantly higher detection rates of preoperative ctDNA in LUSCs than that of LUADs in our study, which was consistent with the TRACERx study and the work of Zhang and colleagues (11, 28). It is reasonable to suggest that the preoperative ctDNA in LUSCs could be mainly derived from necrotic tumor cells in the primary tumors rather than biologically active micrometastatic deposits (37), in which case the presence or absence of ctDNA before surgery would have limited direct connection with recurrence after surgery that aims to completely resect the primary tumor. This assumption could also help to explain our finding that ctDNA positivity after surgery was significantly associated with poorer RFS in patients with LUSCs, as by then it becomes more indicative of residual tumor cells.

Several studies have demonstrated the value of longitudinal ctDNA analysis as a promising biomarker for detecting molecular relapse of lung cancer ahead of standard-of-care radiologic imaging (11, 17, 39). Our reported data from the LUNGCA cohort focused on the perioperative period, which is critical for decision-making on adjuvant therapies and follow-up strategies. Our results showed that MRD positivity within 1 month after surgery was a strong predictor for relapse in NSCLC. Patients with positive MRD were 11 times more likely to experience relapse than those with negative MRD. Our multivariate Cox analysis has also indicated that the MRD status was an independent factor for RFS and had a higher relative contribution for RFS prediction than all clinical variables including the TNM stage. These results suggest that postoperative MRD can serve as a strong prognostic marker for patient stratification in terms of relapse risk.

Given the proven prognostic power of MRD, we further tested the hypothesis that detection of ctDNA-based MRD may better identify NSCLC patients to benefit from adjuvant therapies. In the nonrandomized LUNGCA cohort, adjuvant therapy was administered following standard clinical guidelines based on prognostic stratification by TNM stage (7), and hence with a majority of the patients at stages II and III receiving adjuvant therapy, whereas most stage I patients free of therapy. Therefore, it is not surprising that adjuvant therapy in the entire cohort is associated with shortened RFS. However, in subgroup analysis, adjuvant therapy completely reversed this tendency in the MRD-positive population: MRD-positive patients receiving adjuvant therapies had improved RFS over those not receiving adjuvant therapy, whereas MRD-negative patients receiving adjuvant therapies had worse RFS than their counterparts not receiving adjuvant therapy. After adjusting for clinicopathologic variables, adjuvant therapy was an independent favorable factor for RFS only in the MRD-positive population. The result held true when the analysis was further limited to stage II–III subpopulation. These results clearly suggest that MRD testing can be a powerful tool for selecting patients for adjuvant therapies.

Despite of the relatively large cohort size, the cohort composition as well as the noninterventional design of the current study have posed some limitations to our quest for findings of finer granularity. Particularly, the relatively smaller population sizes of stage II and III patients precluded thorough subgroup analysis. Our findings need to be further validated and refined in individual disease stages through randomized clinical trials in the future.

Taken together, in this large prospective, multicenter NSCLC cohort study, we demonstrated that perioperative ctDNA testing can effectively detect MRD and identify patients with a high risk of relapse. Our findings shed light on the potential clinical utility of perioperative ctDNA-based MRD detection in NSCLC patients and warrant further clinical trials to validate its utility in disease management.

W. Chen reports a patent for Genecast Biotechnology pending; and W. Chen is an employee at Genecast Biotechnology. No disclosures were reported by the other authors.

L. Xia: Data curation, formal analysis, investigation, visualization, writing–original draft, project administration, writing–review and editing. J. Mei: Resources, supervision, funding acquisition, investigation, writing–original draft, writing–review and editing. R. Kang: Formal analysis, investigation, methodology, writing–original draft, writing–review and editing. S. Deng: Resources, investigation, writing–original draft, project administration, writing–review and editing. Y. Chen: Supervision, investigation, writing–original draft, writing–review and editing. Y. Yang: Formal analysis, visualization, methodology, writing–original draft, writing–review and editing. G. Feng: Investigation, project administration, writing–review and editing. Y. Deng: Formal analysis, investigation, methodology, writing–review and editing. F. Gan: Data curation, investigation, writing–review and editing. Y. Lin: Investigation, writing–review and editing. Q. Pu: Investigation, writing–review and editing. L. Ma: Investigation, writing–review and editing. F. Lin: Investigation, writing–review and editing. Y. Yuan: Investigation, writing–review and editing. Y. Hu: Investigation, writing–review and editing. C. Guo: Investigation, writing–review and editing. H. Liao: Investigation, writing–review and editing. C. Liu: Investigation, writing–review and editing. Y. Zhu: Investigation, writing–review and editing. W. Wang: Investigation, writing–review and editing. Z. Liu: Investigation, writing–review and editing. Y. Xu: Investigation, writing–review and editing. K. Li: Investigation, writing–review and editing. C. Li: Investigation, writing–review and editing. Q. Li: Data curation, writing–review and editing. J. He: Methodology, writing–review and editing. W. Chen: Data curation, methodology, writing–review and editing. X. Zhang: Investigation, project administration, writing–review and editing. Y. Kou: Investigation, writing–review and editing. Y. Wang: Investigation, writing–review and editing. Z. Wu: Investigation, writing–review and editing. G. Che: Investigation, writing–review and editing. L. Chen: Investigation, writing–review and editing. L. Liu: Conceptualization, resources, supervision, funding acquisition, investigation, project administration, writing–review and editing.

This work was supported by 1.3.5 Project for Disciplines of Excellence, West China Hospital, Sichuan University (No. ZYGD18021 for L. Liu and No. ZYJC18009 for J. Mei), and the Major Research Project of Sichuan Province (No. 2021YFS0024 for L. Liu).

The costs of publication of this article were defrayed in part by the payment of page charges. This article must therefore be hereby marked advertisement in accordance with 18 U.S.C. Section 1734 solely to indicate this fact.