近年来，快速发展的单细胞空间转录组分析技术为理解各类生理或病理组织的形成与功能、异质性细胞的组装与互作提供了丰富信息。空间数据的特殊复杂性与高度的非理想性为针对性的数据深度解析方法学研究提出了更高的要求和挑战。具体来说，学术界长期关注的关键科学问题之一是生理组织中异质性细胞之间的复杂组装与互作关系。如何在不依赖于配体-受体关系等先验知识的前提下，通过对单细胞空间转录组数据的深度分析，系统、全面地重构和预测细胞互作；并且基于空间转录组数据的学习和迁移，扩展到缺失了空间定位信息的常规单细胞转录组数据中直接预测细胞互作关系？这成为领域内亟待解决的前沿问题。 本论文针对以上不同层次的生物学信息挖掘需求，以新型、高效的深度学习模型为核心，开发了两个深入递进的人工智能分析方法。首先，我们基于变分图自编码器结合对抗策略，开发了一种新方法DeepLinc（deep-learning framework for landscapes of interacting cells），用于从单细胞空间转录组数据中从头重建细胞互作网络 (Genome Biology, 2022) (Li and Yang, 2022)。DeepLinc能够在高度非理想、不完整的空间转录组数据中，高效地学习、过滤错误互作以及填补缺失的近远端互作关系。这一成果扩展了利用空间组学数据研究细胞互作的思路和策略，验证了深度学习模型用于挖掘细胞内分子特征与细胞间互作的隐藏关联关系的可行性。 进一步，通过对DeepLinc的敏感性测试，我们推断出可能参与塑造细胞互作网络的特征基因。在对细胞互作关系的测试中，这些新的特征基因表现出比已知配-受体基因更好的区分能力。这也支持了方法的设计基础，即细胞内的分子特征谱（不仅限于已知的配-受体）与细胞空间互作关系之间存在复杂、紧密的关联。 更进一步，在对细胞互作关系进行学习、重构的基础上，我们受增长图建模的启发，开发了仅使用细胞转录组数据直接预测互作关系的深度学习方法DeepTALK。在模拟数据和真实数据集上的大量测试显示，DeepTALK具有良好稳健的表现，在常规单细胞转录组数据中的应用结果进一步验证了模型的性能。 总而言之，本文基于新的人工智能方法理论，对细胞互作分析方法进行了层层深入、递进的研究工作，扩展了单细胞空间转录组数据解析的思路与策略，也为进一步系统解决细胞互作的重构和预测问题明确了可行的方向。

In recent years, the rapid development of single-cell spatial transcriptome technology has provided rich information for understanding the formation and function of various physiological or pathological tissues, as well as the assembly and interaction of heterogeneous cells. The complexity and high degree of non-ideality for spatial data put forward higher requirements and challenges for data analysis methodology. Specifically, one of the key scientific issues that the academic community has been paying attention to for a long time is complex assembly and interaction relationships among heterogeneous cells in physiological tissues. One question is how to systematically reconstruct and predict cell interactions through in-depth analysis of single-cell spatial transcriptome data without relying on prior knowledge such as ligand-receptor relationships. Another question is how to directly predict cell interaction relationships in conventional single-cell transcriptome data without spatial information? This has become an urgent frontier problem in this field.Focusing on the above different levels of biological information mining needs, this thesis develops in-depth and progressive artificial intelligence analysis methods with some efficient deep-learning models as core. First, we develop a new method named DeepLinc （deep-learning framework for landscapes of interacting cells）, based on a variational graph autoencoder combined with an adversarial strategy, for de novo reconstruction of cell interaction networks from single-cell spatial transcriptome data （Li and Yang, 2022）. DeepLinc can efficiently filter false interactions and impute missing local and distal interactions from non-ideal and incomplete spatial transcriptome data. This achievement expands the ideas and strategies of studying cell interactions with spatial omics data, and verifies the feasibility of utilizing deep learning models to mine hidden correlations between intracellular molecular features and cell interactions.Further, through the sensitivity test of DeepLinc, we deduce signature genes that may be involved in shaping cell interaction network. Besides, these new signature genes show better discrimination effect than known ligand-receptor genes. It also supports the design basis of our methods that there are complex and tight relations between intracellular molecular profiles （not limited to known ligand-receptors） and cell spatial interaction relationships.Furthermore, on the basis of learning and reconstructing cell interaction landscapes, inspired by the growing graph modeling, we develop a deep learning method named DeepTALK. It can directly infer cell interactions using only single-cell transcriptome information. A large number of tests on simulated data and real datasets show that DeepTALK has a good and robust performance, and the application results in conventional single-cell transcriptome data further verify its effectiveness.In summary, based on the artificial intelligence methods, this thesis conducts in-depth and progressive research on cell interaction analysis methods, expands the ideas and strategies for analyzing single-cell spatial transcriptome data, and clarifies the feasible direction for systematically solving the reconstruction and prediction problem of cell interactions.