Increasing clinical usefulness of gene signature prediction rules through simplification and validation

生物医学文献中已经提出了数百个基于高维数据（HDD）的基因签名，并发表了数百篇方法论文，讨论这些数据集的推导、验证和应用预测方法的问题。然而，人们意识到HDD信息的临床价值是有限的，生物计量学家在低维环境中广泛认可的重要原则往往在HDD环境中被忽视。然而，即使在低维数据（LDD）中，也有几个问题需要额外的研究，一些调查将集中在低维数据领域，并考虑采用HDD。该项目的主要重点是与多变量回归模型的开发和验证相关的关键统计问题。方法学部分的发现将用于考虑临床和遗传数据的综合价值可能带来的改进。对于一些被认为是乳腺癌患者有用的预测规则的基因特征，我们的目标是对其临床应用的证据进行批判性评估。后者需要系统的审查，包括评估报告的质量和用于荟萃分析的方法的适用性。在工作包1 （WP1）中，我们研究了识别有影响的观测值及其对选择多变量模型的影响的问题。首先，将在几个例子中比较在低维上下文中提出的方法。在第二步中，我们将考虑它们对HDD的用处。需要调整关键参数。WP2侧重于与多元回归中模型选择的重采样技术相关的方法学问题，这已被证明对推导可解释的预测模型很有用。特别是，开发了将这些技术扩展到HD环境的方法，特别强调模型平均程序，并在显然同样重要（可能相关）的预测因素的情况下支持决策。WP3是第一个资助期工作包的直接延伸，致力于构建和评估利用临床和组学信息的预测模型。WP4旨在通过批判性地考虑文献中发表的现有证据来评估一些基因特征的临床应用。相关的方法学问题是报告的质量，以及用于验证和荟萃分析的方法的适用性。

Hundreds of gene signatures based on high-dimensional data (HDD) have been proposed in the biomedical literature and hundreds of methodological papers are published on issues around deriving, validating and applying prediction methods for such data sets. Nevertheless it is realized that the clinical value of information from HDD is limited and that important principles widely recognized by biometricians in low-dimensional settings tend to be often ignored in the HDD context. However, even in low-dimensional data (LDD) several issues require additional research and some investigations will concentrate in the LD world and consider adoption to HDD. The main emphasis of this project is on key statistical issues related to the development and validation of multivariable regression models. Findings from the methodological part will be used to consider possible improvements attributed to the combined value of clinical and genetic data. For some gene signatures proposed as useful prediction rules in patients with breast cancer we aim to evaluate critically the evidence for their clinical utility. The latter requires systematic reviews including assessment of quality of reporting and suitability of the methods used for meta-analysis. In work package 1 (WP1) we study the problem of identifying influential observations and their impact on selecting a multivariable model. First, approaches proposed in the low-dimensional context will be compared in several examples. In a second step we will consider their usefulness for HDD. Adaption of key parameters will be required. WP2 focuses on methodological issues related to resampling techniques for model selection in multivariate regression, which have been demonstrated to be useful for deriving interpretable prediction models. In particular, methods are developed to extend these techniques to the HD context with a particular emphasis on model averaging procedures and to support decisions in the case of apparently equally important (possibly correlated) predictors. WP3, a direct extension of a work package from the first funding period, is devoted to the construction and evaluation of prediction models making use of both clinical and omics information. WP4 aims to assess the clinical utility of some gene signatures by critically considering available evidence published in the literature. Relevant methodological issues are the quality of reporting, and the suitability of methods used for validation and meta-analysis.