An integrated single-cell RNA-seq map of human neuroblastoma tumors and preclinical models uncovers divergent mesenchymal-like gene expression programs.

Neuroblastoma is a common pediatric cancer, where preclinical studies suggest that a mesenchymal-like gene expression program contributes to chemotherapy resistance. However, clinical outcomes remain poor, implying we need a better understanding of the relationship between patient tumor heterogeneity and preclinical models. Here, we generated single-cell RNA-seq maps of neuroblastoma cell lines, patient-derived xenograft models (PDX), and a genetically engineered mouse model (GEMM). We developed an unsupervised machine learning approach ('automatic consensus nonnegative matrix factorization' (acNMF)) to compare the gene expression programs found in preclinical models to a large cohort of patient tumors. We confirmed a weakly expressed, mesenchymal-like program in otherwise adrenergic cancer cells in some pre-treated high-risk patient tumors, but this appears distinct from the presumptive drug-resistance mesenchymal programs evident in cell lines. Surprisingly however, this weak-mesenchymal-like program was maintained in PDX and could be chemotherapy-induced in our GEMM after only 24 hours, suggesting an uncharacterized therapy-escape mechanism. Collectively, our findings improve the understanding of how neuroblastoma patient tumor heterogeneity is reflected in preclinical models, provides a comprehensive integrated resource, and a generalizable set of computational methodologies for the joint analysis of clinical and pre-clinical single-cell RNA-seq datasets.

A spatial cell atlas of neuroblastoma reveals developmental, epigenetic and spatial axis of tumor heterogeneity.

Neuroblastoma is a pediatric cancer arising from the developing sympathoadrenal lineage with complex inter- and intra-tumoral heterogeneity. To chart this complexity, we generated a comprehensive cell atlas of 55 neuroblastoma patient tumors, collected from two pediatric cancer institutions, spanning a range of clinical, genetic, and histologic features. Our atlas combines single-cell/nucleus RNA-seq (sc/scRNA-seq), bulk RNA-seq, whole exome sequencing, DNA methylation profiling, spatial transcriptomics, and two spatial proteomic methods. Sc/snRNA-seq revealed three malignant cell states with features of sympathoadrenal lineage development. All of the neuroblastomas had malignant cells that resembled sympathoblasts and the more differentiated adrenergic cells. A subset of tumors had malignant cells in a mesenchymal cell state with molecular features of Schwann cell precursors. DNA methylation profiles defined four groupings of patients, which differ in the degree of malignant cell heterogeneity and clinical outcomes. Using spatial proteomics, we found that neuroblastomas are spatially compartmentalized, with malignant tumor cells sequestered away from immune cells. Finally, we identify spatially restricted signaling patterns in immune cells from spatial transcriptomics. To facilitate the visualization and analysis of our atlas as a resource for further research in neuroblastoma, single cell, and spatial-omics, all data are shared through the Human Tumor Atlas Network Data Commons at www.humantumoratlas.org.

A CRISPR-drug perturbational map for identifying compounds to combine with commonly used chemotherapeutics.

Combination chemotherapy is crucial for successfully treating cancer. However, the enormous number of possible drug combinations means discovering safe and effective combinations remains a significant challenge. To improve this process, we conduct large-scale targeted CRISPR knockout screens in drug-treated cells, creating a genetic map of druggable genes that sensitize cells to commonly used chemotherapeutics. We prioritize neuroblastoma, the most common extracranial pediatric solid tumor, where ~50% of high-risk patients do not survive. Our screen examines all druggable gene knockouts in 18 cell lines (10 neuroblastoma, 8 others) treated with 8 widely used drugs, resulting in 94,320 unique combination-cell line perturbations, which is comparable to the largest existing drug combination screens. Using dense drug-drug rescreening, we find that the top CRISPR-nominated drug combinations are more synergistic than standard-of-care combinations, suggesting existing combinations could be improved. As proof of principle, we discover that inhibition of PRKDC, a component of the non-homologous end-joining pathway, sensitizes high-risk neuroblastoma cells to the standard-of-care drug doxorubicin in vitro and in vivo using patient-derived xenograft (PDX) models. Our findings provide a valuable resource and demonstrate the feasibility of using targeted CRISPR knockout to discover combinations with common chemotherapeutics, a methodology with application across all cancers.

Random-effects meta-analysis of effect sizes as a unified framework for gene set analysis.

Gene set analysis (GSA) remains a common step in genome-scale studies because it can reveal insights that are not apparent from results obtained for individual genes. Many different computational tools are applied for GSA, which may be sensitive to different types of signals; however, most methods implicitly test whether there are differences in the distribution of the effect of some experimental condition between genes in gene sets of interest. We have developed a unifying framework for GSA that first fits effect size distributions, and then tests for differences in these distributions between gene sets. These differences can be in the proportions of genes that are perturbed or in the sign or size of the effects. Inspired by statistical meta-analysis, we take into account the uncertainty in effect size estimates by reducing the influence of genes with greater uncertainty on the estimation of distribution parameters. We demonstrate, using simulation and by application to real data, that this approach provides significant gains in performance over existing methods. Furthermore, the statistical tests carried out are defined in terms of effect sizes, rather than the results of prior statistical tests measuring these changes, which leads to improved interpretability and greater robustness to variation in sample sizes.

Cell type identification in spatial transcriptomics data can be improved by leveraging cell-type-informative paired tissue images using a Bayesian probabilistic model.

Spatial transcriptomics technologies have recently emerged as a powerful tool for measuring spatially resolved gene expression directly in tissues sections, revealing cell types and their dysfunction in unprecedented detail. However, spatial transcriptomics technologies are limited in their ability to separate transcriptionally similar cell types and can suffer further difficulties identifying cell types in slide regions where transcript capture is low. Here, we describe a conceptually novel methodology that can computationally integrate spatial transcriptomics data with cell-type-informative paired tissue images, obtained from, for example, the reverse side of the same tissue section, to improve inferences of tissue cell type composition in spatial transcriptomics data. The underlying statistical approach is generalizable to any spatial transcriptomics protocol where informative paired tissue images can be obtained. We demonstrate a use case leveraging cell-type-specific immunofluorescence markers obtained on mouse brain tissue sections and a use case for leveraging the output of AI annotated H&E tissue images, which we used to markedly improve the identification of clinically relevant immune cell infiltration in breast cancer tissue. Thus, combining spatial transcriptomics data with paired tissue images has the potential to improve the identification of cell types and hence to improve the applications of spatial transcriptomics that rely on accurate cell type identification.

The chemotherapeutic CX-5461 primarily targets TOP2B and exhibits selective activity in high-risk neuroblastoma.

Survival in high-risk pediatric neuroblastoma has remained around 50% for the last 20 years, with immunotherapies and targeted therapies having had minimal impact. Here, we identify the small molecule CX-5461 as selectively cytotoxic to high-risk neuroblastoma and synergistic with low picomolar concentrations of topoisomerase I inhibitors in improving survival in vivo in orthotopic patient-derived xenograft neuroblastoma mouse models. CX-5461 recently progressed through phase I clinical trial as a first-in-human inhibitor of RNA-POL I. However, we also use a comprehensive panel of in vitro and in vivo assays to demonstrate that CX-5461 has been mischaracterized and that its primary target at pharmacologically relevant concentrations, is in fact topoisomerase II beta (TOP2B), not RNA-POL I. This is important because existing clinically approved chemotherapeutics have well-documented off-target interactions with TOP2B, which have previously been shown to cause both therapy-induced leukemia and cardiotoxicity-often-fatal adverse events, which can emerge several years after treatment. Thus, while we show that combination therapies involving CX-5461 have promising anti-tumor activity in vivo in neuroblastoma, our identification of TOP2B as the primary target of CX-5461 indicates unexpected safety concerns that should be examined in ongoing phase II clinical trials in adult patients before pursuing clinical studies in children.

Facilitating Drug Discovery in Breast Cancer by Virtually Screening Patients Using In Vitro Drug Response Modeling.

(1) Background: Drug imputation methods often aim to translate in vitro drug response to in vivo drug efficacy predictions. While commonly used in retrospective analyses, our aim is to investigate the use of drug prediction methods for the generation of novel drug discovery hypotheses. Triple-negative breast cancer (TNBC) is a severe clinical challenge in need of new therapies. (2) Methods: We used an established machine learning approach to build models of drug response based on cell line transcriptome data, which we then applied to patient tumor data to obtain predicted sensitivity scores for hundreds of drugs in over 1000 breast cancer patients. We then examined the relationships between predicted drug response and patient clinical features. (3) Results: Our analysis recapitulated several suspected vulnerabilities in TNBC and identified a number of compounds-of-interest. AZD-1775, a Wee1 inhibitor, was predicted to have preferential activity in TNBC (p < 2.2 × 10-16) and its efficacy was highly associated with TP53 mutations (p = 1.2 × 10-46). We validated these findings using independent cell line screening data and pathway analysis. Additionally, co-administration of AZD-1775 with standard-of-care paclitaxel was able to inhibit tumor growth (p < 0.05) and increase survival (p < 0.01) in a xenograft mouse model of TNBC. (4) Conclusions: Overall, this study provides a framework to turn any cancer transcriptomic dataset into a dataset for drug discovery. Using this framework, one can quickly generate meaningful drug discovery hypotheses for a cancer population of interest.

Long non-coding RNA transcriptome of uncharacterized samples can be accurately imputed using protein-coding genes.

Long non-coding RNAs (lncRNAs) play an important role in gene regulation and are increasingly being recognized as crucial mediators of disease pathogenesis. However, the vast majority of published transcriptome datasets lack high-quality lncRNA profiles compared to protein-coding genes (PCGs). Here we propose a framework to harnesses the correlative expression patterns between lncRNA and PCGs to impute unknown lncRNA profiles. The lncRNA expression imputation (LEXI) framework enables characterization of lncRNA transcriptome of samples lacking any lncRNA data using only their PCG profiles. We compare various machine learning and missing value imputation algorithms to implement LEXI and demonstrate the feasibility of this approach to impute lncRNA transcriptome of normal and cancer tissues. Additionally, we determine the factors that influence imputation accuracy and provide guidelines for implementing this approach.

Discovering long noncoding RNA predictors of anticancer drug sensitivity beyond protein-coding genes.

Large-scale cancer cell line screens have identified thousands of protein-coding genes (PCGs) as biomarkers of anticancer drug response. However, systematic evaluation of long noncoding RNAs (lncRNAs) as pharmacogenomic biomarkers has so far proven challenging. Here, we study the contribution of lncRNAs as drug response predictors beyond spurious associations driven by correlations with proximal PCGs, tissue lineage, or established biomarkers. We show that, as a whole, the lncRNA transcriptome is equally potent as the PCG transcriptome at predicting response to hundreds of anticancer drugs. Analysis of individual lncRNAs transcripts associated with drug response reveals nearly half of the significant associations are in fact attributable to proximal cis-PCGs. However, adjusting for effects of cis-PCGs revealed significant lncRNAs that augment drug response predictions for most drugs, including those with well-established clinical biomarkers. In addition, we identify lncRNA-specific somatic alterations associated with drug response by adopting a statistical approach to determine lncRNAs carrying somatic mutations that undergo positive selection in cancer cells. Lastly, we experimentally demonstrate that 2 lncRNAs, EGFR-AS1 and MIR205HG, are functionally relevant predictors of anti-epidermal growth factor receptor (EGFR) drug response.

Cancer expression quantitative trait loci (eQTLs) can be determined from heterogeneous tumor gene expression data by modeling variation in tumor purity.

Expression quantitative trait loci (eQTLs) identified using tumor gene expression data could affect gene expression in cancer cells, tumor-associated normal cells, or both. Here, we have demonstrated a method to identify eQTLs affecting expression in cancer cells by modeling the statistical interaction between genotype and tumor purity. Only one third of breast cancer risk variants, identified as eQTLs from a conventional analysis, could be confidently attributed to cancer cells. The remaining variants could affect cells of the tumor microenvironment, such as immune cells and fibroblasts. Deconvolution of tumor eQTLs will help determine how inherited polymorphisms influence cancer risk, development, and treatment response.

Discovering novel pharmacogenomic biomarkers by imputing drug response in cancer patients from large genomics studies.

Obtaining accurate drug response data in large cohorts of cancer patients is very challenging; thus, most cancer pharmacogenomics discovery is conducted in preclinical studies, typically using cell lines and mouse models. However, these platforms suffer from serious limitations, including small sample sizes. Here, we have developed a novel computational method that allows us to impute drug response in very large clinical cancer genomics data sets, such as The Cancer Genome Atlas (TCGA). The approach works by creating statistical models relating gene expression to drug response in large panels of cancer cell lines and applying these models to tumor gene expression data in the clinical data sets (e.g., TCGA). This yields an imputed drug response for every drug in each patient. These imputed drug response data are then associated with somatic genetic variants measured in the clinical cohort, such as copy number changes or mutations in protein coding genes. These analyses recapitulated drug associations for known clinically actionable somatic genetic alterations and identified new predictive biomarkers for existing drugs.

Exploring the Link between the Germline and Somatic Genome in Cancer.

Carter and colleagues propose a systematic analysis of the germline and somatic genome in cancer. They identify interactions that occur between germline and somatic variants. This elucidates the function of the germline genome in the context of cancer risk and development. Cancer Discov; 7(4); 354-5. ©2017 AACRSee related article by Carter et al., p. 410.

Institutional Profile: Pharmacogenomic research in R Stephanie Huang Laboratory.

The Huang Lab was established in 2009 at the University of Chicago and has since been active in conducting pharmacogenomic research. Our laboratory's main research focus is translational pharmacogenomics with a particular interest in the pharmacogenomics of anticancer agents. By systematically evaluating the human genome and its relationships to drug response and toxicity, our goal is to develop clinically useful models that predict risk for adverse drug reactions and nonresponse prior to administration of chemotherapy. Specifically, the theme of our research evolved around the idea of cell-based pharmacogenomics, which utilizes in vitro models for biomarker discovery and prediction-model construction, followed by in vivo validation. We routinely use cell lines (derived from healthy and diseased individuals as well as commercially available cancer cell lines) and clinical samples to discover and functionally characterize genetic variation and gene, miRNA, and long noncoding RNA expression for their roles in drug sensitivity.

Cancer biomarker discovery is improved by accounting for variability in general levels of drug sensitivity in pre-clinical models.

We show that variability in general levels of drug sensitivity in pre-clinical cancer models confounds biomarker discovery. However, using a very large panel of cell lines, each treated with many drugs, we could estimate a general level of sensitivity to all drugs in each cell line. By conditioning on this variable, biomarkers were identified that were more likely to be effective in clinical trials than those identified using a conventional uncorrected approach. We find that differences in general levels of drug sensitivity are driven by biologically relevant processes. We developed a gene expression based method that can be used to correct for this confounder in future studies.

Shadow Enhancers Are Pervasive Features of Developmental Regulatory Networks.

Embryogenesis is remarkably robust to segregating mutations and environmental variation; under a range of conditions, embryos of a given species develop into stereotypically patterned organisms. Such robustness is thought to be conferred, in part, through elements within regulatory networks that perform similar, redundant tasks. Redundant enhancers (or "shadow" enhancers), for example, can confer precision and robustness to gene expression, at least at individual, well-studied loci. However, the extent to which enhancer redundancy exists and can thereby have a major impact on developmental robustness remains unknown. Here, we systematically assessed this, identifying over 1,000 predicted shadow enhancers during Drosophila mesoderm development. The activity of 23 elements, associated with five genes, was examined in transgenic embryos, while natural structural variation among individuals was used to assess their ability to buffer against genetic variation. Our results reveal three clear properties of enhancer redundancy within developmental systems. First, it is much more pervasive than previously anticipated, with 64% of loci examined having shadow enhancers. Their spatial redundancy is often partial in nature, while the non-overlapping function may explain why these enhancers are maintained within a population. Second, over 70% of loci do not follow the simple situation of having only two shadow enhancers-often there are three (rols), four (CadN and ade5), or five (Traf1), at least one of which can be deleted with no obvious phenotypic effects. Third, although shadow enhancers can buffer variation, patterns of segregating variation suggest that they play a more complex role in development than generally considered.

Seq-ing improved gene expression estimates from microarrays using machine learning.

BACKGROUND: Quantifying gene expression by RNA-Seq has several advantages over microarrays, including greater dynamic range and gene expression estimates on an absolute, rather than a relative scale. Nevertheless, microarrays remain in widespread use, demonstrated by the ever-growing numbers of samples deposited in public repositories. RESULTS: We propose a novel approach to microarray analysis that attains many of the advantages of RNA-Seq. This method, called Machine Learning of Transcript Expression (MaLTE), leverages samples for which both microarray and RNA-Seq data are available, using a Random Forest to learn the relationship between the fluorescence intensity of sets of microarray probes and RNA-Seq transcript expression estimates. We trained MaLTE on data from the Genotype-Tissue Expression (GTEx) project, consisting of Affymetrix gene arrays and RNA-Seq from over 700 samples across a broad range of human tissues. CONCLUSION: This approach can be used to accurately estimate absolute expression levels from microarray data, at both gene and transcript level, which has not previously been possible. This methodology will facilitate re-analysis of archived microarray data and broaden the utility of the vast quantities of data still being generated.

Predicting Response to Histone Deacetylase Inhibitors Using High-Throughput Genomics.

BACKGROUND: Many disparate biomarkers have been proposed as predictors of response to histone deacetylase inhibitors (HDI); however, all have failed when applied clinically. Rather than this being entirely an issue of reproducibility, response to the HDI vorinostat may be determined by the additive effect of multiple molecular factors, many of which have previously been demonstrated. METHODS: We conducted a large-scale gene expression analysis using the Cancer Genome Project for discovery and generated another large independent cancer cell line dataset across different cancers for validation. We compared different approaches in terms of how accurately vorinostat response can be predicted on an independent out-of-batch set of samples and applied the polygenic marker prediction principles in a clinical trial. RESULTS: Using machine learning, the small effects that aggregate, resulting in sensitivity or resistance, can be recovered from gene expression data in a large panel of cancer cell lines.This approach can predict vorinostat response accurately, whereas single gene or pathway markers cannot. Our analyses recapitulated and contextualized many previous findings and suggest an important role for processes such as chromatin remodeling, autophagy, and apoptosis. As a proof of concept, we also discovered a novel causative role for CHD4, a helicase involved in the histone deacetylase complex that is associated with poor clinical outcome. As a clinical validation, we demonstrated that a common dose-limiting toxicity of vorinostat, thrombocytopenia, can be predicted (r = 0.55, P = .004) several days before it is detected clinically. CONCLUSION: Our work suggests a paradigm shift from single-gene/pathway evaluation to simultaneously evaluating multiple independent high-throughput gene expression datasets, which can be easily extended to other investigational compounds where similar issues are hampering clinical adoption.

pRRophetic: an R package for prediction of clinical chemotherapeutic response from tumor gene expression levels.

We recently described a methodology that reliably predicted chemotherapeutic response in multiple independent clinical trials. The method worked by building statistical models from gene expression and drug sensitivity data in a very large panel of cancer cell lines, then applying these models to gene expression data from primary tumor biopsies. Here, to facilitate the development and adoption of this methodology we have created an R package called pRRophetic. This also extends the previously described pipeline, allowing prediction of clinical drug response for many cancer drugs in a user-friendly R environment. We have developed several other important use cases; as an example, we have shown that prediction of bortezomib sensitivity in multiple myeloma may be improved by training models on a large set of neoplastic hematological cell lines. We have also shown that the package facilitates model development and prediction using several different classes of data.

Genome-wide interrogation of longitudinal FEV1 in children with asthma.

RATIONALE: Most genomic studies of lung function have used phenotypic data derived from a single time-point (e.g., presence/absence of disease) without considering the dynamic progression of a chronic disease. OBJECTIVES: To characterize lung function change over time in subjects with asthma and identify genetic contributors to a longitudinal phenotype. METHODS: We present a method that models longitudinal FEV1 data, collected from 1,041 children with asthma who participated in the Childhood Asthma Management Program. This longitudinal progression model was built using population-based nonlinear mixed-effects modeling with an exponential structure and the determinants of age and height. MEASUREMENTS AND MAIN RESULTS: We found ethnicity was a key covariate for FEV1 level. Budesonide-treated children with asthma had a slight but significant effect on FEV1 when compared with those treated with placebo or nedocromil (P < 0.001). A genome-wide association study identified seven single-nucleotide polymorphisms nominally associated with longitudinal lung function phenotypes in 581 white Childhood Asthma Management Program subjects (P < 10(-4) in the placebo ["discovery"] and P < 0.05 in the nedocromil treatment ["replication"] group). Using ChIP-seq and RNA-seq data, we found that some of the associated variants were in strong enhancer regions in human lung fibroblasts and may affect gene expression in human lung tissue. Genetic mapping restricted to genome-wide enhancer single-nucleotide polymorphisms in lung fibroblasts revealed a highly significant variant (rs6763931; P = 4 × 10(-6); false discovery rate < 0.05). CONCLUSIONS: This study offers a strategy to explore the genetic determinants of longitudinal phenotypes, provide a comprehensive picture of disease pathophysiology, and suggest potential treatment targets.