BackgroundWith improvements in next‐generation DNA sequencing technology, lower cost is needed to collect genetic data. More machine learning techniques can be used to help with cancer analysis and diagnosis.MethodsWe developed an ensemble machine learning system named performance‐weighted‐voting model for cancer type classification in 6,249 samples across 14 cancer types. Our ensemble system consists of five weak classifiers (logistic regression, SVM, random forest, XGBoost and neural networks). We first used cross‐validation to get the predicted results for the five classifiers. The weights of the five weak classifiers can be obtained based on their predictive performance by solving linear regression functions. The final predicted probability of the performance‐weighted‐voting model for a cancer type can be determined by the summation of each classifier’s weight multiplied by its predicted probability.ResultsUsing the somatic mutation count of each gene as the input feature, the overall accuracy of the performance‐weighted‐voting model reached 71.46%, which was significantly higher than the five weak classifiers and two other ensemble models: the hard‐voting model and the soft‐voting model. In addition, by analyzing the predictive pattern of the performance‐weighted‐voting model, we found that in most cancer types, higher tumor mutational burden can improve overall accuracy.ConclusionThis study has important clinical significance for identifying the origin of cancer, especially for those where the primary cannot be determined. In addition, our model presents a good strategy for using ensemble systems for cancer type classification.

Though whole exome sequencing (WES) is the gold-standard for measuring tumor mutational burden (TMB), the development of gene-targeted panels enables cost-effective TMB estimation. With the growing number of panels in clinical trials, developing a statistical method to effectively evaluate and compare the performance of different panels is necessary. The mainstream method uses R-squared value to measure the correlation between the panel-based TMB and WES-based TMB. However, the performance of a panel is usually overestimated via R-squared value based on the long-tailed TMB distribution of the dataset. Herein, we propose angular distance, a measurement used to compute the extent of the estimated bias. Our extensive in silico analysis indicates that the R-squared value reaches a plateau after the panel size reaches 0.5 Mb, which does not adequately characterize the performance of the panels. In contrast, the angular distance is still sensitive to the changes in panel sizes when the panel size reaches 6 Mb. In particular, R-squared values between the hypermutation-included dataset and the non-hypermutation dataset differ widely across many cancer types, whereas the angular distances are highly consistent. Therefore, the angular distance is more objective and logical than R-squared value for evaluating the accuracy of TMB estimation for gene-targeted panels.

Deciphering the population structure of SARS-CoV-2 is critical to inform public health management and reduce the risk of future dissemination. With the continuous accruing of SARS-CoV-2 genomes worldwide, discovering an effective way to group these genomes is critical for organizing the landscape of the population structure of the virus. Taking advantage of recently published state-of-the-art machine learning algorithms, we used an unsupervised deep learning clustering algorithm to group a total of 16,873 SARS-CoV-2 genomes. Using single nucleotide polymorphisms as input features, we identified six major subtypes of SARS-CoV-2. The proportions of the clusters across the continents revealed distinct geographical distributions. Comprehensive analysis indicated that both genetic factors and human migration factors shaped the specific geographical distribution of the population structure. This study provides a different approach using clustering methods to study the population structure of a never-seen-before and fast-growing species such as SARS-CoV-2. Moreover, clustering techniques can be used for further studies of local population structures of the proliferating virus.

Spatially resolved transcriptomics performs high-throughput measurement of transcriptomes while preserving spatial information about the cellular organizations. However, many spatially resolved transcriptomic technologies can only distinguish spots consisting of a mixture of cells instead of working at single-cell resolution. Here, we present STdGCN, a graph neural network model designed for cell type deconvolution of spatial transcriptomic (ST) data that can leverage abundant single-cell RNA sequencing (scRNA-seq) data as reference. STdGCN is the first model incorporating the expression profiles from single cell data as well as the spatial localization information from the ST data for cell type deconvolution. Extensive benchmarking experiments on multiple ST datasets showed that STdGCN outperformed 13 published state-of-the-art models. Applied to a human breast cancer Visium dataset, STdGCN discerned spatial distributions between stroma, lymphocytes and cancer cells for tumor microenvironment dissection. In a human heart ST dataset, STdGCN detected the changes of potential endothelial-cardiomyocyte communications during tissue development. Our results demonstrate that STdGCN can serve as a robust and versatile tool for cell type deconvolution across multiple ST platforms and tissues. STdGCN is available as open source Python software at https://github.com/luoyuanlab/stdgcn.