Generating realistic Multiomic data

从高通量表达数据中逆向工程基因调控网络的问题是生物信息学中最大的挑战之一。为了对网络推理算法进行基准测试，通常需要具有良好特征的表达数据集的模拟器。然而，现有的模拟器受到批评，因为它们不能模拟基因表达数据的关键特性（Maier等人，2013年）。在我的研究项目中，我的目标是解决两个问题。首先，我希望研究和提出机制，忠实地评估合成表达数据集的真实性。其次，我希望设计一个基于生成对抗网络的表达数据的对抗模拟器（GAN; Goodfellow埃塔尔.，2014年）。该框架描述了一种通过玩两个玩家的游戏来估计生成模型的方法，其中第一个玩家学习从特定分布中生成样本，第二个玩家试图将它们与来自真实数据分布的样本区分开来。这种新颖的深度学习框架已经在图像或音频生成等任务中显示出了有希望的结果，据我所知，GAN还没有被应用于构建基因表达数据的模拟器。

The problem of reverse engineering gene regulatory networks from high-throughput expression data is one of the biggestchallenges in bioinformatics. In order to benchmark network inference algorithms, simulators of well-characterizedexpression datasets are often required. However, existing simulators have been criticized because they fail to emulatekey properties of gene expression data (Maier et al., 2013). In my research project I aim to address two problems. First, Iwish to study and propose mechanisms to faithfully assess the realism of a synthetic expression dataset. Second, I wishto design an adversarial simulator of expression data based on a generative adversarial network (GAN; Goodfellow etal., 2014). This framework describes a method for estimating a generative model by playing a two-player game, in whichthe first player learns to generate samples from a particular distribution, and the second tries to discriminate them fromthe samples coming from the true data distribution. This novel deep learning framework has shown promising results fortasks such as image or audio generation, and to the best of my knowledge GANs have not yet been applied to build asimulator of gene expression data.