Addressing the high false-positive rate of conventional low-dose computed tomography (LDCT) for lung cancer diagnosis, the efficacy of incorporating blood-based noninvasive testing for assisting practicing clinician's decision making in diagnosis of pulmonary nodules (PNs) is investigated. In this prospective observative study, next generation sequencing-(NGS-) based cell-free DNA (cfDNA) mutation profiling, NGS-based cfDNA methylation profiling, and blood-based protein cancer biomarker testing are performed for patients with PNs, who are diagnosed as high-risk patients through LDCT and subsequently undergo surgical resections, with tissue sections pathologically examined and classified. Using pathological classification as the gold standard, statistical and machine learning methods are used to select molecular markers associated with tissue's malignant classification based on a 98-patient discovery cohort (28 benign and 70 malignant), and to construct an integrative multianalytical model for tissue malignancy prediction. Predictive models based on individual testing platforms have shown varying levels of performance, while their final integrative model produces an area under the receiver operating characteristic curve (AUC) of 0.85. The model's performance is further confirmed on a 29-patient independent validation cohort (14 benign and 15 malignant, with power > 0.90), reproducing AUC of 0.86, which translates to an overall sensitivity of 80% and specificity of 85.7%.

This study aims to assess the potential clinical application of targeted next generation sequencing (NGS)-based deep sequencing for the detection of clinically relevant mutations in circulating tumor DNA (ctDNA) obtained from non-small cell lung cancer (NSCLC) patients. Targeted deep sequencing was performed to identify High Confidence Somatic Variants (HCSVs) in matched tumor tissue DNA (tDNA) and ctDNA in 50 NSCLC patients. Our results demonstrated that NSCLC patients with Stage IV (61.5%) exhibited a higher concordance rate at the mutation level between plasma ctDNA and tDNA samples than patients with Stage I-III (14.5%). Moreover, it is noteworthy that the allele frequency of these detected HCSVs in ctDNA increased with the advance in tumor stage. Besides, using tDNA as a reference, the sensitivity of plasma ctDNA analyzed by deep NGS for actionable EGFR was much higher in patients with Stage IV (66.6%) than in patients with Stage I-III (7.7%). In conclusion, it appears that ctDNA NGS-based deep sequencing is a feasible approach to identify mutations in patients with Stage IV NSCLC. However, additional methods with higher sensitivity and specificity are needed to improve the successful application of this platform in the earlier stages of NSCLC.

Plasma cell-free DNA (cfDNA) methylation and fragmentation signatures have been shown to be valid biomarkers for blood-based cancer detection. However, conventional methylation sequencing assays are inapplicable for fragmentomic profiling due to bisulfite-induced DNA damage. Here using enzymatic conversion-based low-pass whole-methylome sequencing (WMS), we developed a novel approach to comprehensively interrogate the genome-wide plasma methylation, fragmentation, and copy number profiles for sensitive and noninvasive multi-cancer detection. With plasma WMS data from a clinical cohort comprising 497 healthy controls and 780 patients with both early- and advanced-stage cancers of the breast, colorectum, esophagus, stomach, liver, lung, or pancreas, genomic features including methylation, fragmentation size, copy number alteration, and fragment end motif were extracted individually and subsequently integrated to develop an ensemble cancer classifier, called THEMIS, using machine learning algorithms. THEMIS outperformed individual biomarkers for differentiating cancer patients of all seven types from healthy individuals and achieved a combined area under the curve value of 0.971 in the independent test cohort, translating to a sensitivity of 86% and early-stage (I and II) sensitivity of 77% at 99% specificity. In addition, we built a cancer signal origin classifier with true-positive cancer samples at 100% specificity based on methylation and fragmentation profiling of tissue-specific accessible regulatory elements, which localized cancer-like signal to a limited number of clinically informative sites with 66% accuracy. Overall, this proof-of-concept work demonstrates the feasibility of extracting and integrating multi-modal biomarkers from a single WMS run for noninvasive detection and localization of common cancers across stages.

e13554 Background: Approximately 3-5% of human cancers are cancer of unknown primary (CUP). Treatment of a cancer patient is largely dependent on the tumor origin. Therefore, identification of the tumor origin can improve the survival of patients with CUP. We developed a multi-class classification model using DNA methylation profile as biomarker to determine the primary site of CUP. Methods: We split 7,082 primary tumor samples of 19 cancers and 679 normal samples of 15 tissues from TCGA into a 75% training set and a 25% testing set to develop the classification model. We started with multiple support vector machine (SVM) models, and then combined them into an optimal multi-class ensemble model. Predictors included tumor-specific markers and tissue-specific markers, which were filtered by comparing between groups. Only the training samples were used for feature selection and model development. A validation dataset consisting of 150 primary tissues, 54 metastasis tissues, 105 plasma samples with known cancer site origins from 12 classes was generated in house by a self-designed panel. Performance was measured by area under the curve (AUC) using the one-vs-all approach. Results: 7,453 tumor-specific and 1,533 tissue-specific markers were selected for model construction. AUCs of all cancer types were high in TCGA training and testing set (AUC≥0.96 for all classes). In our validation tissues, esophageal cancer, pancreatic cancer, colorectal cancer, lung adenocarcinoma, breast cancer and liver cancer achieved high AUC in both primary (0.83, 0.83, 0.82, 0.82, 0.80 and 0.79 respectively) and metastasis (0.74, 0.92, 0.86, 0.61, 0.92 and 0.65 respectively). Lung adenocarcinoma, colorectal cancer, liver cancer, breast cancer and esophageal cancer even achieved high AUC in the plasmas. Conclusions: Performance of our model in tissue and plasma samples indicated the potential clinical application of DNA methylation profile in unknown cancer origin identification.

e13051 Background: Targeted sequencing of circulating tumor DNA (ctDNA) has been used in early tumor detection and guidance of clinical cancer treatment. Accurate somatic copy number variation (SCNV) estimation is useful for better decisions in clinics, but remain challenging due to low percentage of tumor-cell-released DNA in circulating blood, in addition to the low signal-noise ratio and usual lack of normal control. Methods: To overcome these challenges and call SCNV at gene level, we develop a novel bioinformatics tool ctCNV. After correction for GC content, target region length and read counts, genome local scores at the gene level (GCS) were calculated, using 30 normal blood samples as control. Statistically significant cutoffs were determined using control. Results: The new method has been benchmarked with other SCNV calling tools such as CNVkit, which have been developed mainly for whole exome sequencing instead of targeted DNAseq. The validation experiment involves 6 mixed samples consisting of different proportions of known cell lines and real blood samples. ROC curve obtained through comparing the SCNV calling results with Droplet Digital PCR results show improvement of our method comparing to existing methods. Conclusions: As a summary, our study has three main results/contributions. First, we developed a novel computational tool for SCNV calling in targeted sequencing of ctDNA, with good sensitivity and specificity. Second, we thoroughly evaluated the performance of current available CNV calling tools in targeted sequenced ctDNA. Third, the raw data generated from our six real samples of gradient mixture of different cell lines and human’s samples can serve as an evaluation standard to other further computational tools for CNV calling in targeted sequenced ctDNA.