CRII:III:Development of deep learning methods for high-resolution 3D genome structure spatial reconstruction

来自不同个体和细胞类型、多细胞、单细胞或亚细胞定位的数据量和分辨率的增加，因此增加了这些数据集的复杂性。这种增长意味着我们需要更复杂和自动化的方法来从这些数据集中推断模式和结构。细胞的三维（3D）染色体和基因组组织结构是我们这个数据丰富的社会中可演绎的重要结构之一。重建基因组的三维结构是一项复杂而具有挑战性的任务。然而，有必要了解许多细胞活动，如基因表达，基因稳定性和调控。为了提高对细胞内染色体组织的理解，基于染色体构象捕获技术的基因组技术，特别是Hi-C技术得到了发展。这一发展极大地改善了细胞研究，并导致多年来多种3D染色体结构重建方法的发展。虽然已经提出了许多3D染色体结构重建方法，但缺乏高分辨率（=5kb）的染色体和基因组结构结构的详细见解。该项目旨在开发一种先进的可扩展算法，用于从Hi-C数据中进行高分辨率3D基因组重建，以提供足够的细节来解释精细尺度上的生物活动，如基因-基因相互作用。该项目将作为开源工具和软件开发。此外，该项目的多学科性质提供了一种培训机制，为本科生和研究生的教学和研究提供了实践学习和指导机会。具体来说，该项目的最终目标是开发基于计算和机器学习的框架，通过高分辨率（=5kb） 3D结构重建来阐明基因组内分层组织与其功能之间的相互作用。该项目预计将开发算法和计算工具，以推进以下方面的三维基因组组织研究。首先，开发一种鲁棒、可靠、灵活的高分辨率三维染色体和基因组结构重建算法。该算法将图形卷积神经网络（GCNNs）作为空间结构重构的核心方法。其次，它将开发一种新的非基于实例的3D结构重建方法，能够使用泛化方法进行高分辨率3D基因组结构模型预测，其中训练好的模型可用于重建多个染色体或用于跨分辨率的预测。提出的基于非实例的三维结构方法利用节点嵌入算法对每个染色体位点对应的图节点特征表示。使用GCNN训练这些特征，以生成对应于每个染色体位点的染色体和基因组三维结构坐标的预测。总的来说，本项目将研究GCNN的复杂性和模型深度问题，为染色体和基因组的高分辨率三维结构重建提供新的途径。该项目的成功实施将为染色体和基因组的三维组织提供独特的见解，并激励染色体三维结构预测的非实例方法的未来发展。该奖项反映了美国国家科学基金会的法定使命，并通过使用基金会的知识价值和更广泛的影响审查标准进行评估，被认为值得支持。

An increase in the amount and resolution of data from different individuals and cell types, from multiple cells, single cells, or subcellular localizations, has consequently increased the complexity of these datasets. This increase means we need more sophisticated and automated approaches to infer patterns and structures from these datasets. The cell's three-dimensional (3D) chromosome and genome organization structure is one of the vital structures deducible in our data-rich society. Reconstructing the 3D organization of the genome is a complicated and challenging task. Nevertheless, it is necessary to understand many cellular activities, such as gene expression, gene stability, and regulation. To improve the understanding of chromosome organization within a cell, genomic technologies based on chromosome conformation capture techniques, particularly Hi-C, were developed. This development greatly improved cellular study and led to the development of multiple 3D chromosome structure reconstruction methods over the years. Although many 3D chromosome structure reconstruction methods have been proposed, detailed insight into the structural architecture of the chromosome and genome at a high resolution (=5kb) is lacking. This project aims to develop an advanced and scalable algorithm for high-resolution 3D genome reconstruction from Hi-C data that provides enough detail to explain biological activities such as gene-gene interaction at a refined scale. This project will be developed as open-source tools and software. Also, the multidisciplinary nature of this project provides a mechanism for training and providing hands-on learning and mentoring opportunities in teaching and research to students at both undergraduate and graduate levels.Specifically, the ultimate objective of this project is to develop computational and machine learning-based frameworks to elucidate the interplay between the hierarchical organization within the genome and its functions through high-resolution(=5kb) 3D structure reconstruction. This project is expected to develop algorithms and computational tools to advance 3D genome organization research in the following aspects. First, a robust, reliable, and flexible high-resolution 3D chromosome and genome structure reconstruction algorithm will be developed. This algorithm will use graph convolutional neural networks (GCNNs) as the core method for spatial structure reconstruction. Second, it will develop a novel noninstance-based approach for 3D structure reconstruction capable of high-resolution 3D genome structure model prediction using a generalization approach, where a trained model can be used to reconstruct multiple chromosomes or used for prediction across resolutions. The proposed noninstance-based approach for 3D structure utilizes a node embedding algorithm for the graph node feature representation corresponding to each chromosomal locus. These features are trained with a GCNN to generate predictions for chromosome and genome 3D structure coordinates corresponding to each chromosomal locus. Overall, this project will investigate the GCNN complexity and model depth questions to provide a new approach for chromosome and genome 3D structure reconstruction at a high resolution. The successful implementation of this project will give a unique insight into the 3D organization of chromosomes and genomes and motivate the future development of noninstance-based methods for chromosome 3D structure prediction.This award reflects NSF's statutory mission and has been deemed worthy of support through evaluation using the Foundation's intellectual merit and broader impacts review criteria.