Enhanced Detection of Landmark Minimal Residual Disease in Lung Cancer Using Cell-free DNA Fragmentomics

Currently, approximately 30%–55% of the patients with non–small cell lung cancer (NSCLC) develop recurrence due to minimal residual disease (MRD) after receiving surgical resection of the tumor. This study aims to develop an ultrasensitive and affordable fragmentomic assay for MRD detection in patients with NSCLC. A total of 87 patients with NSCLC, who received curative surgical resections (23 patients relapsed during follow-up), enrolled in this study. A total of 163 plasma samples, collected at 7 days and 6 months postsurgical, were used for both whole-genome sequencing (WGS) and targeted sequencing. WGS-based cell-free DNA (cfDNA) fragment profile was used to fit regularized Cox regression models, and leave-one-out cross-validation was further used to evaluate models’ performance. The models showed excellent performances in detecting patients with a high risk of recurrence. At 7 days postsurgical, the high-risk patients detected by our model showed an increased risk of 4.6 times, while the risk increased to 8.3 times at 6 months postsurgical. These fragmentomics determined higher risk compared with the targeted sequencing–based circulating mutations both at 7 days and 6 months postsurgical. The overall sensitivity for detecting patients with recurrence reached 78.3% while using both fragmentomics and mutation results from 7 days and 6 months postsurgical, which increased from the 43.5% sensitivity by using only the circulating mutations. The fragmentomics showed great sensitivity in predicting patient recurrence compared with the traditional circulating mutation, especially after the surgery for early-stage NSCLC, therefore exhibiting great potential to guide adjuvant therapeutics. Significance: The circulating tumor DNA mutation-based approach shows limited performance in MRD detection, especially for landmark MRD detection at an early-stage cancer after surgery. Here, we describe a cfDNA fragmentomics–based method in MRD detection of resectable NSCLC using WGS, and the cfDNA fragmentomics showed a great sensitivity in predicting prognosis.


Introduction
Non-small cell lung cancer (NSCLC) is the leading cause of worldwide cancer-related deaths (1). After surgical resection of the tumor, approximately 30%-55% of the patients with NSCLC eventually develop recurrence due to minimal residual disease (MRD; refs. 1,2). Up-to-date, Guardant 360 and Foun-dationOne Liquid CDx, which focused on mutation detection in 73 and 311 genes, remain the only two liquid cell-free DNA (cfDNA) assays approved by the FDA for utilization in NSCLC (3,4). However, they both suffer from low sensitivity and are unsuitable for detecting MRD. Therefore, there is an and colleagues designed a patient-specific targeted sequencing panel for ctDNA mutation-based MRD detection (6). However, their panel suffered from low sensitivity, especially at landmark postsurgery, yielding only 36% sensitivity at 90% specificity (6). Chen and colleagues demonstrated a limited performance in detecting landmark MRD (44% sensitivity at 88% specificity) using their circulating single-molecule amplification and resequencing technology assay (9), which was more similar to the results by Abbosh and colleagues (6). Moreover, Zhang and colleagues investigated the utilization of MRD detection assay in a comprehensive cohort of 261 patients with NSCLC, yet only showing a 36.2% sensitivity at 98.0% specificity for MRD detection at landmark (7). Likewise, Waldeck and colleagues 2022 reported a 40% MRD detection rate using 1-2 weeks of postsurgical samples from a relatively small cohort of 20 patients with resectable NSCLC (12). Li and colleagues 2021 reported only a 23% positive MRD detection rate while using the somatic mutation approach on the landmark plasma samples (13). A similar sensitivity (21.4%) was observed by Xia and colleagues in the LUNGCA-1 study, which detected MRD in 15 of 70 patients with progression using 3 days of postsurgical samples (14). Gale and colleagues also showed a 45.4% sensitivity for landmark MRD detection in a cohort of 88 patients with NSCLC using a whole-exome sequencing approach (8).
Overall, the ctDNA mutation-based approach shows limited sensitivity in MRD detection, especially for landmark MRD detection at an early-stage cancer after surgery. Such limitation is potentially contributed by the combination of low ctDNA abundance and relatively low sensitivity of the target sequencing approach (15)(16)(17). The whole-genome sequencing (WGS) approach, however, has shown a higher sensitivity as it is not limited by the number of ctDNA molecules and their specific location (16). Mathios and colleagues showed that fragmentomics machine learning modeling using WGS data could detect patients with NSCLC with extremely high sensitivity (95%). Moreover, Bao and colleagues developed an ultrasensitive assay for multi-cancer early detection incorporating plasma cfDNA fragmentomic features derived from WGS data, showing a 90.5% sensitivity at 95.5% specificity for detecting very early-stage lung cancer (17). A recent article by Wang and colleagues reported an integrated model utilizing five machine learning algorithms on five different cfDNA fragmentomics features for detecting early-stage lung cancer. The integrated model yielded high AUCs in both validation cohorts (0.984 and 0.987; ref. 18). Furthermore, Zviran and colleagues showed an excellent 100% sensitivity while detecting MRD in a small cohort of 22 patients with WGS, albeit limited by a relatively low specificity of 71% (19). Therefore, fragmentomics profiling may have great potential in MRD detection.
In this study, we aim to explore the utilization of plasma cfDNA fragmentomic profiling for MRD detection, especially for landmark detection. We developed a cohort of 100 patients with NSCLC who received curative tumor resections as the standard of care. Fragmentomics feature profiling was retrieved on the basis of WGS data using plasma samples collected at both landmark and longitudinal timepoints, defined as around 7 days and 6 months postsurgery, respectively. Machine learning models were evaluated and then constructed using these fragmentomics profiles.

Patient Enrollment and Sample Collection
A total of 100 patients with NSCLC were initially enrolled in this study cohort between April 2017 and January 2019. All study protocols were approved by the ethics committee of Jiangsu Cancer Hospital and in accordance with the Declaration of Helsinki. Written informed consents were provided by all patients.
Among the 100 enrolled patients, 2 patients later withdrew consents and 10 patients were lost during the follow-up period, resulting in a cohort size of 88 patients (Fig. 1A). These patients were pathologically diagnosed with NSCLC (AJCC, 8th edition) and received curative tumor resections as the standard of care. The histologic subtype and tumor-node-metastasis staging were identified and reviewed by two pathologists. Postsurgical plasma samples, which were collected were scheduled at 7 days and 6 months after surgeries, were used for both target sequencing (425 cancer-associated gene panel, GeneseeqPrime) and WGS. Primary tumor tissue samples, as well as paired leukocyte samples, were also collected and sequenced by the 425-gene panel. One patient was removed from the cohort due to the failure of plasma samples during the quality control process.

Library Preparation and Sequencing
Genomic DNA from tissue samples was extracted with QIAamp DNA formalin-fixed paraffin-embedded (FFPE) Tissue Kit, while DNeasy Blood & Tissue Kit (Qiagen) was used for plasma samples. The Qubit 3.0 fluorometer and dsDNA HS Assay Kit (Thermo Fisher Scientific) were used to quantify the extracted DNA, following the manufacturer's instruction. A deparaffin procedure using xylene was performed on FFPE samples before the genomic DNA extraction with QIAamp DNA FFPE Tissue Kit (Qiagen) per manufacturer's instruction. To extract genomic DNA from the plasma samples, centrifugation was performed at high speed to remove any cell debris. The supernatant was then used for the cfDNA extraction using QIAamp Circulating Nucleic Acid Kit (Qiagen) per manufacturer's instruction. Library preparations were performed with KAPA Hyper Prep Kit (KAPA Biosystems) according to manufacturer's suggestions for different sample types.
For target sequencing, sequential operations of end-repairing, A-tailing, and indexed adapter ligation were performed to 6.08-200 ng (median: 70.5 ng) of cfDNA or 1 μg of fragmented genomic DNA, followed by size selection using Agencourt AMPure XP beads (Beckman Coulter). Hybridization-based target enrichment was carried out with GeneseeqPrime pan-cancer gene panel (425 cancer-relevant genes), which covers a total of approximately 1. according to the manufacturer's instructions. PE150 was used as it was shown by Cristiano and colleagues 2019 that the peak frequency of cfDNA fragment size was around 167 bp (20).

Fragmentomics Machine Learning Model
The raw reads for the WGS data were first trimmed (trailing reads only, quality <20) with Trimmomatic (21). The Picard toolkit (http://broadinstitute.github. io/picard/) was then used for PCR duplicates removal. The reads were then mapped onto the human reference genome (GRCh37/UCSC hg19) using the sequence aligner Burrows-Wheeler Aligner (BWA, v0.7.12; ref. 22). Fragment size was calculated using the mapped distances between 5 ends of the read pairs and should not be significantly impacted by the trimming of trailing reads. To eliminate the potential impact on the predictive power of the different coverages among the WGS data, we downsampled the coverages to a unified 5 ×.
The Fragment Size Ratio (FSR) profile, which was adapted from previous reports by Ma and colleagues and Zhang and colleagues (23,24), examined the ratios of different size fragments across the human genome, as the cfDNA fragments are reported to be aberrantly short and long in cancer patient's plasma samples (25). In-house scripts were used to construct these fragmentation profiles as reported previously (23,24), which focused on comparing the short/long reads around the first and second peaks, as shown in Fig. 1B. The ShortPeak1, LongPeak1, ShortPeak2, and LongPeak2 fragments were defined as 100-150 bp, 151-220 bp, 221-300 bp, and 311-400 bp, according to the overall fragment length profile in our cohorts (Fig. 1B). The ratios of the short/long fragments of both peaks for each sample were examined in 5 Mb bins, resulting in a total of 2,164 (541 bins × 4) FSR features from the 541 bins genome-wide, excluding the sex chromosomes (22 autosomes) and masking any repeat regions, which were then used by the machine learning algorithms for model construction. Following the previous report by Cristiano and colleagues, a local smoother was applied to remove any potential bias in the coverage due to the guanine-cytosine (GC) content (20).
A total of 87 patients with NSCLC, including 23 patients with recurrence during follow-up, were used to construct predictive models for MRD. Two machine learning models, which used either 7 days or 6 months postsurgical samples, selected for its advantages against other competing methods, including efficacy, stability, significantly shorter runtime and, most importantly, its ability to process high-dimensional input features (27). Both models used leave-oneout cross-validation to evaluate the predictive performance of MRD. As shown in Fig. 1A, each of the samples (87 and 76 for the 7 days and 6 months postsurgical samples, respectively) were used once as a validation set during the leave-one-out cross-validation. The remaining samples (86 or 75) were used as the training set to fit a Coxnet model, which were then used for predicting the MRD risk score for the validation set. This process was then repeated 87 (76) times till the risk score for every sample was generated. To determine the cutoffs for fragmentomics-predicted MRD-positive status, ROC curves were constructed by the pROC package (v. 1.17.0.1) using the sample risk scores. The cutoffs for both models were defined as the best sensitivity at a minimal of 90% specificity.

Identification of ctDNA Mutation Profile
The ctDNA mutation profile was constructed using a previously reported method (2, 10) by tracking somatic mutations identified from cancer tissue samples in patients' plasma samples. Variant calling of plasma samples was tumor-informed by comparing the identified mutations against mutations detected in the primary tumor tissue sample and the leukocyte samples. Tissue clonality was annotated for each variant detected in plasma. Somatic singlenucleotide variant (sSNV) and insertion/deletion (InDel) detected in plasma samples were filtered if (i) they were not in the paired tissue mutational profile, (ii) they fell in an in-house list of clonal hematopoiesis variants, or (iii) they were detected in paired leukocyte controls with at least one variant read. Furthermore, a normal pool was constructed using 100 healthy individuals to determine whether the mutation found in plasma samples was significantly higher than background noise.

Availability of Data and Materials
The sequencing data reported in the study have been deposited in the European Genome-phenome Archive (EGA) database as EGAD00001010300. The data are deposited under controlled access for access to the data contact Dr. Rong Yin, rong_yin@njmu.edu.cn. All the other data supporting the findings of this study are available within Supplementary Data and from the corresponding author upon reasonable request.

Participant Characteristics in the Cohort
A total of 100 patients with pathologically diagnosed NSCLC, all received curative tumor resections as the standard of care. Among these 100 patients, 10 patients were lost during the follow-up period, while 2 patients withdrew their consents and 1 patient's samples failed the quality-control process, resulting in a final cohort size of 87 (Fig. 1). All patients had not received neoadjuvant therapy.

Fragmentomics Coxnet Models for MRD Detection
We developed two Coxnet models by using the FSR feature profile of both 7 days and 6 months postsurgical to predict the risk status for recurrence. ROC curves were constructed with predicted risk scores from leave-one-out cross-validation results.
As shown in Fig. 2A, both Coxnet models showed excellent AUCs (7 days postsurgical: 0.817, 95% CI: 0.724-0.909; 6 months postsurgical: 0.837, 95% CI: 0.725-0.950) for distinguishing patients with recurrence from recurrence-free patients. The risk score cutoff for identifying patients with a high risk of recurrence was selected to ensure maximizing sensitivity while maintaining over 90% specificity. As shown in Fig. 2B and C, patients with recurrence showed higher risk scores than recurrence-free patients for both 7 days and 6 months postsurgical models. As shown in Supplementary     Furthermore, as shown in Fig. 4A and B, 10 patients with recurrence were successfully identified as high risk by the fragmentomics model at 7 days postsurgical. The median lead time was 293 days (145-514 days) for the modelpredicted recurrence than the radiographic recurrence. As shown in Fig. 5A and B, the risk scores predicted by the Coxnet models were positively correlated with the max variant allele frequency determined by the ctDNA mutationbased method. Moreover, it is shown that the profiles of short/long fragment ratio at Peak1 were distinguishable between patients who progressed during the follow-up and those who did not yet progress (Fig. 5C). The median short/long ratio was larger in the progressed patient group, indicating that they have shorter cfDNA fragments compared with the yet-to-progress patients.
These results were consistent with the previous finding that the plasma cfDNA fragments were aberrantly short in cancer patient samples (25).

Comparing Fragmentomics Models and ctDNA Mutation Profiling
We then compared our model-predicted results against the ctDNA mutation profiling results. As shown in Supplementary Coxnet models (Fig. 6). At 7 days postsurgical, the Coxnet model predicted 16 patients with a high risk of recurrence, compared with the only 10 ctDNApositive patients by mutation profiling, as shown in Fig. 6A. This resulted in a 1.6 times sensitivity for the fragmentomics model at 7 days postsurgical. The same trend was observed at 6 months postsurgical, with the fragmentomics model-predicted 16 patients with high risk of recurrence compared with the 9 ctDNA-positive patients by mutation profiling, yielding a 1.8 times higher sensitivity ( Fig. 6B and C).
The overall sensitivity was even higher (56.5%, 95% CI: 34.5-76.8) at 7 days postsurgical, when combining the fragmentomics model prediction with the ctDNA mutation profiling results ( Fig. 6A and C). A patient was labeled as high risk for progression while having either a high-risk status predicted by the fragmentomics model or a positive ctDNA status based on the target panel sequencing. A low progression risk status would require a patient to have a predicted low-risk status by fragmentomics as well as a negative ctDNA status by mutational profiling (Fig. 3E and F). However, there was no increase for the 6 months postsurgical timepoint, as the sensitivity of combined results was identical for using only fragmentomics model. Furthermore, after combining both 7 days and 6 months postsurgical results, the Coxnet model predicted 24 patients with high risk of recurrence, compared with the 16 ctDNA-positive patients, as shown in Supplementary Fig. S1A Fig. S1B).

Discussion
Despite receiving curative surgeries, a large number of patients with NSCLC can still develop recurrence. Therefore, there is an urgent clinical need for MRD detection in patients with postsurgical NSCLC, which can assist in treatment decision-making progress to both maximize the chance of cure and minimize the risk of overtreatment. However, the current MRD detection methods are limited by several significant barriers: low sensitivity and high turnaround time.
Herein, we reported an ultrasensitive assay for enhanced detection of landmark MRD in patients with postsurgical NSCLC. By using the more sensitive WGS Moreover, as the WGS approach requires no existing knowledge of mutation profile in the cancer tissue samples, the fragmentomics model is, in theory, of more clinical potential by having less turnaround time and limitation due to tumor tissue insufficiency. For example, Li and colleagues 2021 failed to detect any mutation in the baseline plasma sample for approximately 40% of their cohort, contributing to the limited sensitivity of MRD detection (13). One significant advantage of our fragmentomics model is its ability to predict landmark MRD status at a very early stage (7 days) postsurgery compared with other studies (2 weeks to 4 months; refs. 6,7,9,19), which could potentially assist timely decision-making for postsurgery treatment.
However, this study is still limited, especially by the small cohort size and the lack of an independent test cohort. The leave-one-out cross-validation approach, despite having less bias than a K-fold cross-validation approach, can still be subject to overfitting as it relies solely on the training cohort. A large multicenter study is needed to validate the predictive power of cfDNA fragmentomics in landmark MRD detection using an independent validation cohort. We have already started recruiting patients for the independent validation cohort. However, such a cohort is not available in time for this study due to the long follow-up period required to obtain the data. Furthermore, it will be interesting to investigate the effectiveness of postsurgery treatment decision-making assisted by model-predicted MRD status.