Clinical Validation of a Cell-Free DNA Fragmentome Assay for Augmentation of Lung Cancer Early Detection

Abstract

Lung cancer screening via annual low-dose computed tomography has poor adoption. We conducted a prospective case-control study among 958 individuals eligible for lung cancer screening to develop a blood-based lung cancer detection test that when positive is followed by a low-dose computed tomography. Changes in genome-wide cell-free DNA fragmentation profiles (fragmentomes) in peripheral blood reflected genomic and chromatin characteristics of lung cancer. We applied machine learning to fragmentome features to identify individuals who were more or less likely to have lung cancer. We trained the classifier using 576 cases and controls from study samples and validated it in a held-out group of 382 cases and controls. The validation demonstrated high sensitivity for lung cancer and consistency across demographic groups and comorbid conditions. Applying test performance to the screening eligible population in a 5-year model with modest utilization assumptions suggested the potential to prevent thousands of lung cancer deaths. Significance: Lung cancer screening has poor adoption. Our study describes the development and validation of a novel blood-based lung cancer screening test utilizing a highly affordable, low-coverage genome-wide sequencing platform to analyze cell-free DNA fragmentation patterns. The test could improve lung cancer screening rates leading to substantial public health benefits. See related commentary by Haber and Skates, p. 2025.

Figure 1.

Overall approach to clinical validation of a cell-free DNA fragmentome assay for augmentation of lung cancer early detection. A, Illustration representing the DELFI approach for lung cancer through noninvasive assessment of cell-free DNA fragmentation profiles (ratio of short to long cfDNA fragments). Nucleosomal DNA with variable length of linker DNA is released by dying lung cancer cells into the circulation. Genome-wide mapping of the cfDNA fragments demonstrates more aberrant profiles with cancer cell cfDNA fragments compared with the cfDNA in noncancer individuals. B, The DNA Evaluation of Fragments for Early Interception–Lung Cancer Training Study, DELFI-L101 study was a prospective case–control study (NCT04825834, including two institutional supplementary protocols NCT00301119 and NCT01775072). The flow diagram illustrates the inclusion and exclusion of L101 participants based on clinical, sample, and assay eligibility criteria and the assignment of evaluable participants to the classifier training (n = 576) and clinical validation (n = 382) sets. Machine learning of genome-wide cfDNA fragmentation profiles from the training set was used to develop a locked classifier that was evaluated in the clinical validation set.

Figure 2.

Genome-wide fragmentation profiles are altered in patients with cancer and reflect underlying chromatin structure. A, The fragmentation profile (ratio of short to long cfDNA fragments in 5 Mb bins) across the genome was evaluated in the classifier training plasma samples of lung cancer (n = 181) and noncancer individuals (n = 395). The noncancer individuals had similar fragmentation profiles, whereas patients with lung cancer exhibited significant variation. B, Comparison of cfDNA profiles with Hi-C A/B chromatin compartment reference data from lung cancer tissue or peripheral blood cells. Track 1 shows A/B compartments extracted from LUSC cancer tissue (48). Track 2 shows a median lung cancer component extracted from the LUSC plasma samples of 7 patients with lung cancer from the classifier training set with high tumor fraction by ichorCNA (49). The 7 LUSC cases with high ichorCNA have values of 0.051, 0.439, 0.230, 0.259 0.439, 0.167, and 0.057. Track 3 shows the median profile for 10 noncancer plasma samples from the training set. Track 4 shows A/B compartments for lymphoblast cells (48). These four tracks show chromosome 22 as an example, with darker shading indicating informative regions of the genome where the two reference tracks differ in domain (open/closed). C, 100-kb regions were selected using the reference LUSC and lymphoblast A/B tracks as having the same chromatin state or opposite chromatin state. Within these regions, the deviation of the fragmentation value from a noncancer cfDNA reference (n = 10) was plotted per region per individual (noncancer n = 20, LUSC n = 7). Values around 0 have little variation from the noncancer reference. Negative values indicate a region has a more open chromatin state than the reference and positive values indicate a region has a more closed chromatin state than the reference. These data suggest that although cfDNA profiles of healthy individuals reflect the chromatin structure of blood cells, those of patients with lung cancer represent a mixture of cfDNA patterns of chromatin compartments from lung cancer as well as blood cells.

Figure 3.

High-dimensional fragmentation features reflect lung cancer biology and are incorporated in the machine learning classifier. A, Heatmap representation of the deviation of cfDNA fragmentation features across the genome for the classifier training set with lung cancer or noncancer individuals compared with the mean of classifier training noncancer individuals. Each row represents a sample, whereas columns show individual genomic features. The cross-validated DELFI score and clinical characteristics are indicated to the left of the fragmentation deviation heatmap. B, Left, TCGA-derived observations of chromosomal arm gains (red) and losses (blue) in lung adenocarcinoma (LUAD; n = 518) and squamous cell cancer tissues (LUSC; n = 501). Right, the observed chromosome arm gains (red) and losses (blue) in the classifier training individuals separated by histology. C, A heatmap representation of the principal component eigenvectors of the fragmentation profile features. Regression coefficients from the final classifier indicating how the principal components of the fragmentation profiles and z-scores of the chromosomal arms were combined are provided in the top and right margins of the heatmap, respectively. Positive values for the coefficients are represented in red, whereas negative values are represented in blue. Agreement across copy number chromosomal gains and losses in TCGA lung cancers, observed z-scores in the cfDNA of patients with lung cancer, and chromosome arm model coefficients reflect biologic consistency between chromosomal changes in lung cancer, cfDNA fragmentation profiles, and classifier features.

Figure 4.

Performance of blood-based lung cancer screening test. A, Sensitivity and specificity of the test in the clinical validation set (N = 382) overall and by clinical subgroup. Point estimates are reported with 95% Wilson confidence intervals. Overall sensitivity and specificity denoted by solid vertical lines. B, Sensitivity of the test in the lung cancer cases in the clinical validation set (N = 248) evaluated across cancer histology, and T, N, and M categories. Point estimates are reported with 95% Wilson confidence intervals. Overall sensitivity of 84% denoted by the solid horizontal line. C, Left, sensitivity of the test in the lung cancer cases in the clinical validation set (N = 246) by cancer group stage. Middle, bar plot showing the stage distribution of lung cancer as observed in populations undergoing lung cancer screening with LDCT (based on NLST study) that are used to weigh observed stage-specific sensitivities. Right, lung cancer screening relevant stage-weighted sensitivity in clinical validation set. D, Comparison of the NNS with LDCT conditioned on test positive or negative result when applied in the lung cancer screening eligible population. Test performance showed consistency across clinical subgroups and expected increased performance with increasing burden of disease (tumor (T), node (N), metastasis (M) and group staging). After weighting, the stage distribution to reflect a screening population, test performance remained high and demonstrated the ability to reliably identify those individuals more likely to have lung cancer detected on LDCT.

Figure 5.

Population health benefits of a blood-based test in lung cancer screening. A, Care pathway reflecting the recommended standard of care for lung cancer screening with LDCT that is received by 6%–10% of eligible individuals annually, as well as potential pathway employing initial blood-based test and follow-on events. B, The predicted number of cancers detected by screening scenario: LDCT alone (“base case”); LDCT + low test uptake; LDCT + high test uptake. C, Predicted cancers diagnosed at stage I versus Stage IV by screening scenario: LDCT alone (“base case”); LDCT + low test uptake; LDCT + high test uptake. D, Predicted decrease in lung cancer deaths represented by screening scenario: LDCT alone (“base case”); LDCT + low test uptake; LDCT + high test uptake. E, Simulated comparison of the predicted number needed to scan with LDCT to detect one lung cancer: LDCT alone (“base case”); LDCT + low test uptake; LDCT + high test uptake. Population-level modeling demonstrates significant health benefits when a blood-based test is available as an alternative for lung cancer screening.