Circulating-tumor DNA (ctDNA) present in plasma cell–free DNA (cfDNA) is a valuable tool for the noninvasive characterization and monitoring of cancer dynamics. However, the sensitivity of detection of ctDNA is challenging in low-disease-burden cases such as early-stage disease or minimal residual disease (MRD). 1 Ultra-deep sequencing of a limited target set of common cancer drivers has improved the detection of ctDNA present at low concentrations. Yet, ultra-deep targeted sequencing still fails to discover a signal in cases where there is an absence of physical tumor-derived DNA fragments that support a targeted single nucleotide variant (SNV), and single mutated reads are rarely sufficient for confident variant detection.2 More recently, personalized patient-specific panels capable of tracking hundreds of mutations were shown to improve detection of MRD using cfDNA.3,4 Nevertheless, the sensitivity of this approach is still driven by the number of tumor mutations available to track. Innovative sequencing approaches and computational tools are necessary to maximize the information obtained from blood samples, especially from patients with low disease burden.
In a recent article, Dr. Asaf Zviran and colleagues developed an ultra-sensitive, machine learning–powered DNA-sequencing strategy named MRDetect. The algorithm searches across the genome for cumulative patterns of mutations and region-specific changes in the number of copies of ctDNA based on whole-genome sequencing (WGS) information from tumor samples. This study elegantly demonstrates how increased breadth or an expanded number of mutations sequenced a given number of times could effectively overcome ctDNA sampling limitations. This study not only tested the trade-off between breadth and depth, but also challenged commonly used locus-centric mutation callers. The researchers developed a read-centric classification framework that uses every individual read as input to discriminate between sequencing reads containing true variants and sequencing artifacts.5,6 The machine learning model predicts the tumor fraction (TF) present in cfDNA as a function of the number of detected sites, mutation load, and sequencing coverage depth.7,8 MRDetect also generates SNV and copy number alteration (CNA) classifiers as independent variables that can be integrated into a single statistical detection score to improve detection power. This model’s performance is independent of coverage and variant allele frequency (VAF), and thus is expected to perform well even in ultra-low VAF cases.7
To validate MRDetect, ctDNA samples were simulated using in silico admixtures of tumor and normal WGS data from various cancers. More than 700 admixtures were generated with TFs ranging from 10−5 to 0.2. MRDetect achieved accurate and sensitive ctDNA detection in TFs as low 10−5, with 35× sequencing depth and assay-level specificity of 95 percent. Subsequently, the TF’s lowest limit of detection (LLOD) was evaluated in relationship with mutational load and sequencing depth by using melanoma admixtures that were generated with varying depth of coverage (10-120×), mutation load (2,000-63,000), and TFs (10−6-10−3). At a high mutation load, TF detection was possible as low as 10−6 with 120× sequencing depth. Notably, MRDetect also extended the detection range of CNAs by two orders of magnitude compared to leading benchmarked CNA algorithms.9
MRDetect was evaluated in the clinical setting using plasma samples from patients with lung adenocarcinoma (LUAD), colorectal cancer (CRC), and metastatic melanoma. Patients with LUAD and CRC underwent surgery, while those with melanoma were treated with nivolumab. Plasma from healthy controls were used to characterize the noise background and estimate clinical specificity. As expected, MRDetect indicated a decrease in disease burden in patients who underwent both surgery and treatment, and showed a correlation with disease stage. In treated melanoma patients, MRDetect effectively tracked tumor responses and detected residual disease consistent with radiographic changes. In CRC patients, MRDetect-SNV analysis in the pre-operative plasma showed an area under the curve (AUC) of 0.97, sensitivity of 90 percent, and specificity of 98 percent. MRDetect-CNA analysis showed lower performance consistent with the lower CNA load, with an AUC of 0.73, sensitivity of 40 percent, and specificity of 92 percent. Twelve patients did not recur and ctDNA was not detected in their postoperative plasma. MRDetect detected ctDNA in four patient’s postoperative samples, but the patients did not show evidence of recurrence. However, three of these patients received adjuvant therapy that may have eliminated residual disease prior to detection. Lastly, MRDetect inferred ctDNA in stage I-IIa LUAD at 10–3, placing genome-wide integration for early disease monitoring at the same level or better than deep targeted approaches.
In Brief
As sequencing technologies and machine learning methods continue to mature, the analysis of ctDNA could be applied to multiple facets of patient care, such as cancer screening, risk stratification, and real-time monitoring of residual or progressive disease. Ultra-low tumor fractions such as those observed in early-stage disease and MRD continue to challenge the prevailing mutation-calling paradigm. Despite deep targeted sequencing having a recognized ceiling of depth that can ultimately be achieved, few groups have been dedicated to advancing alternative approaches that can be applied to cases with low tumor burden. The development of MRDetect by Dr. Landau’s research group transforms the detection landscape of ctDNA. By using a low-input genome-wide integration approach, with novel computational tools capable of reading and interpreting each fragment in the sample, a 100× higher sensitivity than benchmarked methods was achieved while still using broad sequencing efforts. This study emphasizes the need to challenge conventional ideas regarding sequencing techniques and classifiers. Notably, MRDetect is still a tumor-informed approach, which holds its own limitations. Although the data are limited, pre-operative plasma was used as a replacement for tissue biopsies for melanoma samples assessed in this study. Further exploration and confirmation of the usability of pretreatment or preoperative plasma samples to serve as baseline reference samples for MRDetect would be useful for cohorts where sufficient tumor biopsy specimens are difficult to obtain. Further validation in large cohorts of other cancer types with longer follow-up data is needed; thus, more cases of recurrence remain to be tested using MRDetect, especially for hematologic malignancies with variable mutation loads that were not investigated here. Moving forward, ultra-sensitive ctDNA tests such as MRDetect need to be implemented into the standard design of clinical trials. Once blood tests can provide integral biomarkers in interventional therapeutic trials, they will have the potential to advance into routine clinical care.
References
Competing Interests
Drs. Lightbody, Dutta, and Ghobrial indicated no relevant conflicts of interest.
