Reliable post hoc interpretations of deep learning in genomics

项目总结 了解转录因子的协调如何与非编码DNA结合提供了机制 对转录调控的洞察。深度神经网络(DNN)的最新发展已经 彻底改变了我们研究调控基因组学的能力。虽然他们已经证明了更好的预测 与以前基于传统计算基因组学的方法相比，它们的低可解释性为 他们有黑匣子的名声.为了解决这一差距，出现了后自组织模型可解释性方法 询问该网络已经了解到的重要特征。其中，归因地图已经证明了 承诺，为给定序列中的每个核苷酸提供重要性分数；这些分数具有自然的 解释为单核苷酸变异效应。原则上，归属图应该包含以下信息 确定对细胞类型特定调控功能重要的基序，并注释它们在碱基上的位置- 决议。然而，属性映射在实践中往往是噪声的；除了主题之外，它们还包含虚假的 任意核苷酸的重要性分数，其原因尚未完全确定。尽管他们做出了承诺， 通过归因映射来解释DNN仍然是具有挑战性的。在这里，我们提出了三个互补的目标 这有助于我们从基因组DNN的属性图中获得最大限度的生物学洞察力。在……里面 目标1，我们将开发一个模型选择框架来从一组候选DNN中识别出最优的DNN 这产生了高泛化性能和可解释的属性图。在目标2中，我们将开发健壮的 为基因组学量身定做的基于正规化和数据扩充的培训战略，目标更广泛 确保DNN产生高质量的属性图和高度概括性。在目标3中，我们将发展和 使用可解释的计算方法直接分析属性图，以便于发现 功能主题并注解它们的位置。每个目标都将作为开源软件在 TensorFlow和PyTorch。随着基因组学中深度学习应用程序的数量迅速增加， 生物医学社区将极大地受益于这些用户友好的计算工具，因为它使 为任何接受过功能基因组分析培训的DNN部署可靠的培训和可解释性分析。 反过来，这将推动顺式调控生物学的新发现，跨越许多深入的生物系统 学习已经被应用于，并且在未来将继续出现新的应用。

PROJECT SUMMARY Understanding how the coordination of transcription factors bind to non-coding DNA provides mechanistic insights into transcriptional regulation. Recent developments in deep neural networks (DNNs) have revolutionized our ability to study regulatory genomics. While they have demonstrated improved predictions compared to previous methods based on traditional computational genomics, their low interpretability has earned them a reputation as a black box. To address this gap, post hoc model interpretability methods have emerged to interrogate important features that the network has learned. Of these, attribution maps have demonstrated promise, providing importance scores for each nucleotide in a given sequence; these have a natural interpretation as single-nucleotide variant effects. In principle, attribution maps should contain information to identify motifs that are important for cell-type specific regulatory functions and annotate their positions at base- resolution. However, attribution maps are often noisy in practice; in addition to motifs, they contain spurious importance scores for arbitrary nucleotides for reasons that are not well established. Despite their promise, interpreting a DNN through attribution maps remains challenging. Here we propose three complementary aims that serve to maximize the biological insights that we can achieve from attribution maps for genomic DNNs. In Aim 1, we will develop a model selection framework to identify the optimal DNN from a set of candidate DNNs that yields high generalization performance and interpretable attribution maps. In Aim 2, we will develop robust training strategies based on regularization and data augmentations tailored for genomics, with the broader aim of ensuring that DNNs yield high-quality attribution maps and high generalization. In Aim 3, we will develop and employ interpretable computational methods to directly analyze attribution maps to facilitate discovery of functional motifs and annotate their positions. Each aim will be implemented as open-source software in TensorFlow and PyTorch. As the number of deep learning applications in genomics is rising quickly, the biomedical community will greatly benefit from these user-friendly computational tools by enabling the deployment of robust training and interpretability analysis for any DNN trained on functional genomics assays. This, in turn, will drive new discoveries in cis-regulatory biology across the many biological systems that deep learning has already been applied to and the new applications that will continue to emerge in the future.