Carcinoma of unknown primary (CUP) is a type of metastatic cancer, the primary tumor site of which cannot be identified. CUP occupies approximately 5% of cancer incidences in the United States with usually unfavorable prognosis, making it a big threat to public health. Traditional methods to identify the tissue-of-origin (TOO) of CUP like immunohistochemistry can only deal with around 20% CUP patients. In recent years, more and more studies suggest that it is promising to solve the problem by integrating machine learning techniques with big biomedical data involving multiple types of biomarkers including epigenetic, genetic, and gene expression profiles, such as DNA methylation. Different biomarkers play different roles in cancer research; for example, genomic mutations in a patient’s tumor could lead to specific anticancer drugs for treatment; DNA methylation and copy number variation could reveal tumor tissue of origin and molecular classification. However, there is no systematic comparison on which biomarker is better at identifying the cancer type and site of origin. In addition, it might also be possible to further improve the inference accuracy by integrating multiple types of biomarkers. In this study, we used primary tumor data rather than metastatic tumor data. Although the use of primary tumors may lead to some biases in our classification model, their tumor-of-origins are known. In addition, previous studies have suggested that the CUP prediction model built from primary tumors could efficiently predict TOO of metastatic cancers (Lal et al., 2013; Brachtel et al., 2016). We systematically compared the performances of three types of biomarkers including DNA methylation, gene expression profile, and somatic mutation as well as their combinations in inferring the TOO of CUP patients. First, we downloaded the gene expression profile, somatic mutation and DNA methylation data of 7,224 tumor samples across 21 common cancer types from the cancer genome atlas (TCGA) and generated seven different feature matrices through various combinations. Second, we performed feature selection by the Pearson correlation method. The selected features for each matrix were used to build up an XGBoost multi-label classification model to infer cancer TOO, an algorithm proven to be effective in a few previous studies. The performance of each biomarker and combination was compared by the 10-fold cross-validation process. Our results showed that the TOO tracing accuracy using gene expression profile was the highest, followed by DNA methylation, while somatic mutation performed the worst. Meanwhile, we found that simply combining multiple biomarkers does not have much effect in improving prediction accuracy.

Metastatic cancers require further diagnosis to determine their primary tumor sites. However, the tissue-of-origin for around 5% tumors could not be identified by routine medical diagnosis according to a statistics in the United States. With the development of machine learning techniques and the accumulation of big cancer data from The Cancer Genome Atlas (TCGA) and Gene Expression Omnibus (GEO), it is now feasible to predict cancer tissue-of-origin by computational tools. Metastatic tumor inherits characteristics from its tissue-of-origin, and both gene expression profile and somatic mutation have tissue specificity. Thus, we developed a computational framework to infer tumor tissue-of-origin by integrating both gene mutation and expression (TOOme). Specifically, we first perform feature selection on both gene expressions and mutations by a random forest method. The selected features are then used to build up a multi-label classification model to infer cancer tissue-of-origin. We adopt a few popular multiplelabel classification methods, which are compared by the 10-fold cross validation process. We applied TOOme to the TCGA data containing 7,008 non-metastatic samples across 20 solid tumors. Seventy four genes by gene expression profile and six genes by gene mutation are selected by the random forest process, which can be divided into two categories: (1) cancer type specific genes and (2) those expressed or mutated in several cancers with different levels of expression or mutation rates. Function analysis indicates that the selected genes are significantly enriched in gland development, urogenital system development, hormone metabolic process, thyroid hormone generation prostate hormone generation and so on. According to the multiple-label classification method, random forest performs the best with a 10-fold cross-validation prediction accuracy of 96%. We also use the 19 metastatic samples from TCGA and 256 cancer samples downloaded from GEO as independent testing data, for which TOOme achieves a prediction accuracy of 89%. The cross-validation validation accuracy is better than those using gene expression (i.e., 95%) and gene

Sequencing-based identification of tumor tissue-of-origin (TOO) is critical for patients with cancer of unknown primary lesions. Even if the TOO of a tumor can be diagnosed by clinicopathological observation, reevaluations by computational methods can help avoid misdiagnosis. In this study, we developed a neural network (NN) framework using the expression of a 150-gene panel to infer the tumor TOO for 15 common solid tumor cancer types, including lung, breast, liver, colorectal, gastroesophageal, ovarian, cervical, endometrial, pancreatic, bladder, head and neck, thyroid, prostate, kidney, and brain cancers. To begin with, we downloaded the RNA-Seq data of 7,460 primary tumor samples across the above mentioned 15 cancer types, with each type of cancer having between 142 and 1,052 samples, from the cancer genome atlas. Then, we performed feature selection by the Pearson correlation method and performed a 150-gene panel analysis; the genes were significantly enriched in the GO:2001242 Regulation of intrinsic apoptotic signaling pathway and the GO:0009755 Hormonemediated signaling pathway and other similar functions. Next, we developed a novel NN model using the 150 genes to predict tumor TOO for the 15 cancer types. The average prediction sensitivity and precision of the framework are 93.36 and 94.07%, respectively, for the 7,460 tumor samples based on the 10-fold cross-validation; however, the prediction sensitivity and precision for a few specific cancers, like prostate cancer, reached 100%. We also tested the trained model on a 20-sample independent dataset with metastatic tumor, and achieved an 80% accuracy. In summary, we present here a highly accurate method to infer tumor TOO, which has potential clinical implementation.

No abstract

Circulating tumor cells (CTCs) derived from primary tumors and/or metastatic tumors are markers for tumor prognosis, and can also be used to monitor therapeutic efficacy and tumor recurrence. Circulating tumor cells enrichment and screening can be automated, but the final counting of CTCs currently requires manual intervention. This not only requires the participation of experienced pathologists, but also easily causes artificial misjudgment. Medical image recognition based on machine learning can effectively reduce the workload and improve the level of automation. So, we use machine learning to identify CTCs. First, we collected the CTC test results of 600 patients. After immunofluorescence staining, each picture presented a positive CTC cell nucleus and several negative controls. The images of CTCs were then segmented by image denoising, image filtering, edge detection, image expansion and contraction techniques using python's openCV scheme. Subsequently, traditional image recognition methods and machine learning were used to identify CTCs. Machine learning algorithms are implemented using convolutional neural network deep learning networks for training. We took 2300 cells from 600 patients for training and testing. About 1300 cells were used for training and the others were used for testing. The sensitivity and specificity of recognition reached 90.3 and 91.3%, respectively. We will further revise our models, hoping to achieve a higher sensitivity and specificity.