Assessment of modelling strategies for drug response prediction in cell lines and xenografts.

Data from several large high-throughput drug response screens have become available to the scientific community recently. Although many efforts have been made to use this information to predict drug sensitivity, our ability to accurately predict drug response based on genetic data remains limited. In order to systematically examine how different aspects of modelling affect the resulting prediction accuracy, we built a range of models for seven drugs (erlotinib, pacliatxel, lapatinib, PLX4720, sorafenib, nutlin-3 and nilotinib) using data from the largest available cell line and xenograft drug sensitivity screens. We found that the drug response metric, the choice of the molecular data type and the number of training samples have a substantial impact on prediction accuracy. We also compared the tasks of drug response prediction with tissue type prediction and found that, unlike for drug response, tissue type can be predicted with high accuracy. Furthermore, we assessed our ability to predict drug response in four xenograft cohorts (treated either with erlotinib, gemcitabine or paclitaxel) using models trained on cell line data. We could predict response in an erlotinib-treated cohort with a moderate accuracy (correlation ≈ 0.5), but were unable to correctly predict responses in cohorts treated with gemcitabine or paclitaxel.

Molecular Classification Substitutes for the Prognostic Variables Stage, Age, and MYCN Status in Neuroblastoma Risk Assessment.

BACKGROUND: Current risk stratification systems for neuroblastoma patients consider clinical, histopathological, and genetic variables, and additional prognostic markers have been proposed in recent years. We here sought to select highly informative covariates in a multistep strategy based on consecutive Cox regression models, resulting in a risk score that integrates hazard ratios of prognostic variables. METHODS: A cohort of 695 neuroblastoma patients was divided into a discovery set (n=75) for multigene predictor generation, a training set (n=411) for risk score development, and a validation set (n=209). Relevant prognostic variables were identified by stepwise multivariable L1-penalized least absolute shrinkage and selection operator (LASSO) Cox regression, followed by backward selection in multivariable Cox regression, and then integrated into a novel risk score. RESULTS: The variables stage, age, MYCN status, and two multigene predictors, NB-th24 and NB-th44, were selected as independent prognostic markers by LASSO Cox regression analysis. Following backward selection, only the multigene predictors were retained in the final model. Integration of these classifiers in a risk scoring system distinguished three patient subgroups that differed substantially in their outcome. The scoring system discriminated patients with diverging outcome in the validation cohort (5-year event-free survival, 84.9±3.4 vs 63.6±14.5 vs 31.0±5.4; P<.001), and its prognostic value was validated by multivariable analysis. CONCLUSION: We here propose a translational strategy for developing risk assessment systems based on hazard ratios of relevant prognostic variables. Our final neuroblastoma risk score comprised two multigene predictors only, supporting the notion that molecular properties of the tumor cells strongly impact clinical courses of neuroblastoma patients.

A comprehensive assessment of somatic mutation detection in cancer using whole-genome sequencing.

As whole-genome sequencing for cancer genome analysis becomes a clinical tool, a full understanding of the variables affecting sequencing analysis output is required. Here using tumour-normal sample pairs from two different types of cancer, chronic lymphocytic leukaemia and medulloblastoma, we conduct a benchmarking exercise within the context of the International Cancer Genome Consortium. We compare sequencing methods, analysis pipelines and validation methods. We show that using PCR-free methods and increasing sequencing depth to ∼ 100 × shows benefits, as long as the tumour:control coverage ratio remains balanced. We observe widely varying mutation call rates and low concordance among analysis pipelines, reflecting the artefact-prone nature of the raw data and lack of standards for dealing with the artefacts. However, we show that, using the benchmark mutation set we have created, many issues are in fact easy to remedy and have an immediate positive impact on mutation detection accuracy.

Decision-tree based model analysis for efficient identification of parameter relations leading to different signaling states.

In systems biology, a mathematical description of signal transduction processes is used to gain a more detailed mechanistic understanding of cellular signaling networks. Such models typically depend on a number of parameters that have different influence on the model behavior. Local sensitivity analysis is able to identify parameters that have the largest effect on signaling strength. Bifurcation analysis shows on which parameters a qualitative model response depends. Most methods for model analysis are intrinsically univariate. They typically cannot consider combinations of parameters since the search space for such analysis would be too large. This limitation is important since activation of a signaling pathway often relies on multiple rather than on single factors. Here, we present a novel method for model analysis that overcomes this limitation. As input to a model defined by a system of ordinary differential equations, we consider parameters for initial chemical species concentrations. The model is used to simulate the system response, which is then classified into pre-defined classes (e.g., active or not active). This is combined with a scan of the parameter space. Parameter sets leading to a certain system response are subjected to a decision tree algorithm, which learns conditions that lead to this response. We compare our method to two alternative multivariate approaches to model analysis: analytical solution for steady states combined with a parameter scan, and direct Lyapunov exponent (DLE) analysis. We use three previously published models including a model for EGF receptor internalization and two apoptosis models to demonstrate the power of our approach. Our method reproduces critical parameter relations previously obtained by both steady-state and DLE analysis while being more generally applicable and substantially less computationally expensive. The method can be used as a general tool to predict multivariate control strategies for pathway activation and to suggest strategies for drug intervention.

A highly standardized, robust, and cost-effective method for genome-wide transcriptome analysis of peripheral blood applicable to large-scale clinical trials.

The use of peripheral blood mononuclear cells (PBMC) for transcriptome analysis has already been proven valuable for assessing disease-associated and drug-response-related gene signatures. While these proof-of-principle studies have been critically important, the instability of RNA within PBMC prohibits their use in large-scale multicenter trials for which samples have to be transported for a prolonged time prior to RNA isolation. Therefore, a prerequisite for transcriptome analysis of peripheral blood in clinical trials will be a standardized and valid method to stabilize the RNA profile immediately after blood withdrawal. Moreover, to be able to perform such large-scale clinical studies routinely in several hundred patients more cost-effective array technologies are required. To address these critical issues, we have combined a whole-blood RNA stabilization technology with a method to reduce globin mRNA, followed by genome-wide transcriptome analysis using a newly introduced BeadChip oligonucleotide technology. We demonstrate that the globin mRNA reduction method results in significantly improved data quality of stabilized RNA samples with low intragroup variance and a detection rate of expressed genes similar to that in PBMC. More important, even small differences in gene expression such as are observed between females and males were detected and sufficient to predict gender in whole-blood samples. We therefore propose the combination of globin mRNA reduction after whole-blood RNA stabilization with a newly introduced cost-effective BeadChip array as the preferred approach for large-scale multicenter trials, especially when establishing predictive markers for disease and treatment outcome in peripheral blood.