Deep learning for understanding gene regulation in diseases via 'omics' integration

项目摘要 我们建议开发和完善神经网络基因表达，以了解基因 疾病的调节。我们将设计深度学习框架来整合各种数据集 （组蛋白修饰，3D构象，序列和SNP），并建立它们之间的关系模型 与基因表达有关。我们提出的模型将明确地捕捉底层结构 和生物数据的复杂性来学习有意义的联系。例如，我们将使用 基于图形的神经网络将DNA的3D构象建模为图形并学习 从不同基因组区域之间的连接来预测基因表达。我们的一 使用这些方法的关键目标是提取可能有助于 基因的上调和下调。我们将通过应用解释来实现这一目标 神经网络的方法这些方法将使我们能够将重要性分数分配给 对感兴趣的特定预测贡献最大的输入特征。比较 这些健康和疾病细胞系基因的得分将提供对基因的深入了解。 错误调节，并作为生物实验的假设驱动工具。我们也 提出了一种新的贝叶斯推理为基础的解释方法，以改善解释 基于图形的神经网络，可以应用于各种任务。最后，鉴于 单细胞技术和估算方法的改进，我们将扩大我们的深度 学习框架来模拟信号之间的关系，如染色质可及性， DNA甲基化与基因表达这个方向将使我们能够探索 在消除噪音和产生高质量的单细胞样本的插补方法， 用于基因调控的深度学习建模。看看模型化的关系 在细胞的发育阶段，可以确定潜在的失调的时间点， 疾病因此，本提案旨在制定利用数据集的统一方法 跨越多个存储库，利用他们的集体知识， 以数据驱动的方式了解疾病。

PROJECT SUMMARY We propose to develop and refine neural networks gene expression to understand gene regulation in diseases. We will design deep learning frameworks to integrate various datasets (histone modifications, 3D conformation, sequences, and SNPs) and model their relationship with the gene expression. Our proposed models will explicitly capture the underlying structure and complexity of the biological data to learn meaningful connections. For example, we will use a graph-based neural network to model the 3D conformation of the DNA as a graph and learn from the connections between different genomic regions to predict gene expression. One of our critical goals for using these methods is to extract relevant signals that could be contributing to the up- and down-regulation of genes. We will accomplish this goal by applying interpretation methods for neural networks. These methods will allow us to assign importance scores to the input features that contribute the most towards a particular prediction of interest. Comparing these scores for genes across healthy and disease cell lines will provide insights into gene misregulation and serve as a hypothesis driving tool for biological experiments. We also propose a novel Bayesian inference-based interpretation method to improve explanations of graph-based neural networks that could be applied to various tasks. Finally, given the improvement of single-cell technologies and imputation methods, we will extend our deep learning frameworks to model relationships between signals like chromatin accessibility and DNA methylation with gene expression. This direction will allow us to explore the effectiveness of the imputation methods in removing noise and generating high-quality single-cell samples for usage in deep learning modeling of gene regulation. Looking at the modeled relationships across the cell's developmental stages could pinpoint timepoints for potential misregulation in diseases. Therefore, this proposal aims to develop unified approaches that utilize datasets spanning multiple repositories to leverage their collective knowledge and improve our understanding of diseases in a data-driven manner.