Identifying tumor cells at the single-cell level using machine learning.

Tumors are complex tissues of cancerous cells surrounded by a heterogeneous cellular microenvironment with which they interact. Single-cell sequencing enables molecular characterization of single cells within the tumor. However, cell annotation-the assignment of cell type or cell state to each sequenced cell-is a challenge, especially identifying tumor cells within single-cell or spatial sequencing experiments. Here, we propose ikarus, a machine learning pipeline aimed at distinguishing tumor cells from normal cells at the single-cell level. We test ikarus on multiple single-cell datasets, showing that it achieves high sensitivity and specificity in multiple experimental contexts.

Functional interplay of Epstein-Barr virus oncoproteins in a mouse model of B cell lymphomagenesis.

Epstein-Barr virus (EBV) is a B cell transforming virus that causes B cell malignancies under conditions of immune suppression. EBV orchestrates B cell transformation through its latent membrane proteins (LMPs) and Epstein-Barr nuclear antigens (EBNAs). We here identify secondary mutations in mouse B cell lymphomas induced by LMP1, to predict and identify key functions of other EBV genes during transformation. We find aberrant activation of early B cell factor 1 (EBF1) to promote transformation of LMP1-expressing B cells by inhibiting their differentiation to plasma cells. EBV EBNA3A phenocopies EBF1 activities in LMP1-expressing B cells, promoting transformation while inhibiting differentiation. In cells expressing LMP1 together with LMP2A, EBNA3A only promotes lymphomagenesis when the EBNA2 target Myc is also overexpressed. Collectively, our data support a model where proproliferative activities of LMP1, LMP2A, and EBNA2 in combination with EBNA3A-mediated inhibition of terminal plasma cell differentiation critically control EBV-mediated B cell lymphomagenesis.

Evaluation of colorectal cancer subtypes and cell lines using deep learning.

Colorectal cancer (CRC) is a common cancer with a high mortality rate and a rising incidence rate in the developed world. Molecular profiling techniques have been used to better understand the variability between tumors and disease models such as cell lines. To maximize the translatability and clinical relevance of in vitro studies, the selection of optimal cancer models is imperative. We have developed a deep learning-based method to measure the similarity between CRC tumors and disease models such as cancer cell lines. Our method efficiently leverages multiomics data sets containing copy number alterations, gene expression, and point mutations and learns latent factors that describe data in lower dimensions. These latent factors represent the patterns that are clinically relevant and explain the variability of molecular profiles across tumors and cell lines. Using these, we propose refined CRC subtypes and provide best-matching cell lines to different subtypes. These findings are relevant to patient stratification and selection of cell lines for early-stage drug discovery pipelines, biomarker discovery, and target identification.

PiGx: reproducible genomics analysis pipelines with GNU Guix.

In bioinformatics, as well as other computationally intensive research fields, there is a need for workflows that can reliably produce consistent output, from known sources, independent of the software environment or configuration settings of the machine on which they are executed. Indeed, this is essential for controlled comparison between different observations and for the wider dissemination of workflows. However, providing this type of reproducibility and traceability is often complicated by the need to accommodate the myriad dependencies included in a larger body of software, each of which generally comes in various versions. Moreover, in many fields (bioinformatics being a prime example), these versions are subject to continual change due to rapidly evolving technologies, further complicating problems related to reproducibility. Here, we propose a principled approach for building analysis pipelines and managing their dependencies with GNU Guix. As a case study to demonstrate the utility of our approach, we present a set of highly reproducible pipelines called PiGx for the analysis of RNA sequencing, chromatin immunoprecipitation sequencing, bisulfite-treated DNA sequencing, and single-cell resolution RNA sequencing. All pipelines process raw experimental data and generate reports containing publication-ready plots and figures, with interactive report elements and standard observables. Users may install these highly reproducible packages and apply them to their own datasets without any special computational expertise beyond the use of the command line. We hope such a toolkit will provide immediate benefit to laboratory workers wishing to process their own datasets or bioinformaticians seeking to automate all, or parts of, their analyses. In the long term, we hope our approach to reproducibility will serve as a blueprint for reproducible workflows in other areas. Our pipelines, along with their corresponding documentation and sample reports, are available at http://bioinformatics.mdc-berlin.de/pigx.

netSmooth: Network-smoothing based imputation for single cell RNA-seq.

Single cell RNA-seq (scRNA-seq) experiments suffer from a range of characteristic technical biases, such as dropouts (zero or near zero counts) and high variance. Current analysis methods rely on imputing missing values by various means of local averaging or regression, often amplifying biases inherent in the data. We present netSmooth, a network-diffusion based method that uses priors for the covariance structure of gene expression profiles on scRNA-seq experiments in order to smooth expression values. We demonstrate that netSmooth improves clustering results of scRNA-seq experiments from distinct cell populations, time-course experiments, and cancer genomics. We provide an R package for our method, available at: https://github.com/BIMSBbioinfo/netSmooth.