Multi-modal cell-free DNA genomic and fragmentomic patterns enhance cancer survival and recurrence analysis

Abstract

The structure of cell-free DNA (cfDNA) is altered in the blood of patients with cancer. From whole-genome sequencing, we retrieve the cfDNA fragment-end composition using a new software (FrEIA [fragment end integrated analysis]), as well as the cfDNA size and tumor fraction in three independent cohorts (n = 925 cancer from >10 types and 321 control samples). At 95% specificity, we detect 72% cancer samples using at least one cfDNA measure, including 64% early-stage cancer (n = 220). cfDNA detection correlates with a shorter overall (p = 0.0086) and recurrence-free (p = 0.017) survival in patients with resectable esophageal adenocarcinoma. Integrating cfDNA measures with machine learning in an independent test set (n = 396 cancer, 90 controls) achieve a detection accuracy of 82% and area under the receiver operating characteristic curve of 0.96. In conclusion, harnessing the biological features of cfDNA can improve, at no extra cost, the diagnostic performance of liquid biopsies.

Keywords: cancer; cell-free DNA; fragmentomics; liquid biopsy; multi-modal; sequencing.

Graphical abstract

Figure 1

Measures of cfDNA biological features are altered in cancer (A) The number of cancer, nodule, and control samples, the biological signatures of cfDNA, and the extracted measures used in this study. (B) Log₁₀ cancer/control fold changes (FCs) of the 5′ trinucleotide fragment end sequence proportions. Trinucleotides with a p < 0.01 and a log₁₀FC below the 25^th percentile (red) or above the 75^th percentile (blue) are shown. (C) The log₁₀FC of trinucleotides significantly altered (∗p < 0.01) in various cancer types pre-treatment. (D and E) The increase in (D) the FrEIA score and (E) the Gini diversity index by cancer type in pre-treatment samples. (F) Aberrant normalized size distribution of cfDNA fragments in pre-treatment cancer samples compared to control samples. The vertical dashed lines outline the size interval used to calculate the P20-150 measure. (G and H) The (G) P20-150 and (H) the ichorCNA TF increased by cancer type in pre-treatment samples. Bd, bile duct cancer; Br, breast cancer; Cr, colorectal cancer; Es, esophageal cancer; Ga, gastric cancer; Gl, glioblastoma; Lu, lung cancer; Ov, ovarian cancer; Pa, pancreatic cancer. Numbers below the cancer type abbreviation represent the sample count. Cancer types with less than 10 samples are in the “other” category. p values were calculated using two-sided Mann-Whitney U test: ns, not significant, ∗p < 0.05, ∗∗p < 0.01, ∗∗∗p < 0.005, ∗∗∗∗p < 0.001. When multiple hypotheses were tested, alpha values were adjusted using the Bonferroni method. #, mean passing the threshold of 3% tumor fraction. No biological or technical replicates were used.

Figure 2

cfDNA biological signatures from xenograft mouse models (n = 16) grafted with a human colorectal cancer cell line (A) The schematic of the workflow. Signatures were computed from the reads aligning to the human reference genome (GRCh38; graft). (B) Spearman correlation of trinucleotide fragment-end proportions from the patient-derived samples (cancer) and the xenograft-derived samples (graft). Trinucleotides with a p < 0.01 and a log₁₀FC below the 25^th percentile (red) or above the 75^th percentile (blue) are shown, as computed from the patient-derived samples with a tumor fraction >10% (see Figure 1B). (C–E) The increase in (C) the FrEIA score, (D) the Gini diversity index, and (E) the proportion of short fragments (P20-150) of the graft (in magenta) compared to patient-derived control and cancer samples. No biological or technical replicates were used.

Figure 3

Correlation between cfDNA biological features and with physiological variables (A) Spearman correlation between cfDNA measures in pre-treatment cancer samples. (B) Spearman correlation of cfDNA biological variables with the mutant allele fraction, where available (n = 196 samples). (C) Spearman correlation between cfDNA measures in control samples. ns, not significant, other values p < 0.01. (D–F) Spearman correlation of age and (D) the FrEIA score, (E) the Gini diversity index, and (F) the P20-150 of controls. (G–I) The FrEIA score (G), the Gini diversity index (H), and the P20-150 (I) by gender of control individuals. p values were calculated using two-sided Mann-Whitney U test: ns, not significant, ∗∗∗∗p < 0.001. No biological or technical replicates were used.

Figure 4

Cancer detection and classification using cfDNA biological features (A) Receiver operating characteristic (ROC) curve of the detection performance of pre-treatment samples using distinct cfDNA measures individually or in combination (all metrics). The vertical dashed line marks 95% specificity. (B) The proportion of detected pre-treatment lung and esophageal adenocarcinoma samples by stage. The numbers below the stages represent the detection rate. (C) Detection rates by at least one of the cfDNA measure or by the MAF, where available, of pre-treatment samples (n = 196 samples). (D) Schematic representation of the machine learning approach. (E) ROC curve from predictions on an independent dataset of a logistic regression classifier based on individual or the combination of cfDNA measures. The vertical dashed line marks 95% specificity. (F) Prediction probabilities of the logistic regression classifier of pre-treatment lung and esophageal adenocarcinoma samples by stage. Samples above the detection threshold (the horizontal dashed line) are considered detected. C, controls; NC, nodules; I, stage I; II, stage II; III, stage III; IV, stage IV. Numbers below the stages represent detection rates. No biological or technical replicates were used.

Figure 5

cfDNA biological patterns enable monitoring and prediction of recurrence for esophageal carcinoma (A) Schematic representation of the clinical timeline and sampling of patients with EAC. (B) The change in FrEIA score between pre-CRT and post-CRT samples based on the pathological complete response (pCR). p values were calculated using two-sided Mann-Whitney U test: ns, not significant, ∗∗p < 0.01, ∗∗∗p < 0.005. (C) Clinical timeline of patients with EAC undergoing resection from the PERFECT subcohort (n = 33) centered around the time point of resection. EOT, end of treatment; CSD, cancer-specific death. (D) Kaplan-Meier curves of the recurrence-free survival probabilities for patients with EAC from the PERFECT subcohort from the postresection time point. Samples with one of the measures higher than the threshold were considered “detected” (FrEIA score: 2.54, Gini diversity index: 0.98, P20-150: 0.26, ichorCNA: 3% sensitivity threshold). p values were calculated using log-rank test statistics. Dashed lines show the median survival time. No biological or technical replicates were used.

Figure 6

Integrating cfDNA biological patterns improve survival prognostication (A) Kaplan-Meier curves of the survival probability for patients with EAC from the postsurgery time point. Dashed lines represent the median survival. (B) Kaplan-Meier curves of the survival probability for patients with lung cancer from the time of initial sampling. Samples with one of the measures higher than the threshold were considered “detected” (FrEIA score: 2.54, Gini diversity index: 0.98, P20-150: 0.26, ichorCNA: 3% sensitivity threshold). p values were calculated using log-rank test statistics. Dashed lines show the median survival time. No biological or technical replicates were used.