Functional regression framework with applications to drug response prediction

个体患者的药物敏感性预测是精准医学面临的重大挑战。目前的建模方法考虑预测药物反应曲线的单个特征，如曲线下面积或IC50。然而，剂量反应曲线的单一特征总结并不能提供整个药物敏感性概况，因为有些特征随细胞系而系统性地变化，而另一些特征则随药物而变化。本提案的总体目标是设计一个基于靶反应曲线和遗传特征的剂量反应曲线预测的数学和计算框架。对于个别患者，问题是制定为功能预测器的功能预测。功能预测因子是指药物抑制的特定靶点的剂量反应，可以从化学数据库和药物激酶活性研究中获得。为了实现我们的目标，我们提出了三个具体目标：在目标1中，我们将设计一个贝叶斯框架，用于估计疾病的重要驱动因素，同时设计一个单一和联合药物反应预测模型。在目标2中，我们将通过纳入遗传特征形式的非功能预测因子来增强模型，并开发一个联合模型，以遗传特征和靶反应曲线作为输入来预测剂量反应曲线。我们提出了一种高效的搜索这种极高维预测空间的方法。在目标3中，我们将开发一种混合预测机制，该机制结合了基于推理动机的模型技术和计算效率高的机器学习技术，以改进预测并同时获得重要的预测因子。所提出的框架适用于建模和分析细胞对单一或组合扰动剂的反应，成功完成目标将使我们能够设计个性化治疗，同时增加对细胞对外部扰动的功能反应的理解。

Drug sensitivity prediction for individual patients is a significant challenge for precision medicine. Current modeling approaches consider prediction of a single feature of the drug response curve such as Area Under the Curve or IC50. However, the single feature summary of the dose response curve does not provide the entire drug sensitivity profile as some features vary systematically with cell lines while others with drugs. The overall goal of this proposal is to design a mathematical and computational framework for dose response curve prediction based on target response curves and genetic characterizations. For individual patients, the problem is formulated as functional prediction from functional predictors. The functional predictors refer to the dose response of specific targets inhibited by the drug that can be obtained from chemical databases and drug kinase activity studies. To achieve our goal, we propose three specific aims: In Aim 1, we will design a Bayesian framework for estimating the significant drivers of a disease along with the design of a model for single and combination drug response prediction. In Aim 2, we will enhance the model by incorporating non-functional predictors in the form of genetic characterizations and develop a joint model to predict dose response curves with both genetic characteristics as well as target response curves as inputs. We propose to develop an efficient way of searching such extremely high dimensional predictor space. In Aim 3, we will develop a hybrid prediction mechanism that combines inferentially motivated model-based techniques and computationally efficient machine learning techniques to improve predictions as well as obtain significant predictors simultaneously. The proposed framework is appropriate for modeling and analyzing cellular response to single or combination of perturbation agents and the successful completion of the aims will allow us to design personalized therapy along with increased understanding of functional response of cells to external perturbations.