Im Yunju, Li Rong, Ma Shuangge
Department of Biostatistics, University of Nebraska Medical Center, Omaha, Nebraska, USA.
Department of Biostatistics, Yale School of Public Health, New Haven, Connecticut, USA.
Stat Med. 2025 Feb 10;44(3-4):e10350. doi: 10.1002/sim.10350.
With the increasing maturity of genetic profiling, an essential and routine task in cancer research is to model disease outcomes/phenotypes using genetic variables. Many methods have been successfully developed. However, oftentimes, empirical performance is unsatisfactory because of a "lack of information." In cancer research and clinical practice, a source of information that is broadly available and highly cost-effective comes from pathological images, which are routinely collected for definitive diagnosis and staging. In this article, we consider a Bayesian approach for selecting relevant genetic variables and modeling their relationships with a cancer outcome/phenotype. We propose borrowing information from (manually curated, low-dimensional) pathological imaging features via reinforcing the same selection results for the cancer outcome and imaging features. We further develop a weighting strategy to accommodate the scenario where information borrowing may not be equally effective for all subjects. Computation is carefully examined. Simulations demonstrate competitive performance of the proposed approach. We analyze TCGA (The Cancer Genome Atlas) LUAD (lung adenocarcinoma) data, with overall survival and gene expressions being the outcome and genetic variables, respectively. Findings different from the alternatives and with sound properties are made.
随着基因谱分析日益成熟，癌症研究中的一项重要常规任务是利用基因变量对疾病结局/表型进行建模。许多方法已成功开发出来。然而，由于“信息不足”，实证表现往往不尽人意。在癌症研究和临床实践中，广泛可得且性价比高的一种信息来源是病理图像，这些图像是为明确诊断和分期而常规收集的。在本文中，我们考虑一种贝叶斯方法来选择相关基因变量并对它们与癌症结局/表型的关系进行建模。我们建议通过强化癌症结局和影像特征的相同选择结果，从（人工整理的、低维的）病理影像特征中借用信息。我们进一步开发了一种加权策略，以适应信息借用对所有受试者可能并非同样有效的情况。对计算进行了仔细研究。模拟结果表明所提方法具有竞争力。我们分析了TCGA（癌症基因组图谱）的LUAD（肺腺癌）数据，分别将总生存期和基因表达作为结局和基因变量。得出了与其他方法不同且具有良好性质的结果。
