The article introduces CytoCommunity, a novel algorithm for identifying tissue cellular neighborhoods (TCNs) based on cell phenotypes and spatial distributions. Unlike existing methods that rely on clustering or intermediate steps, CytoCommunity uses a graph neural network (GNN) model to directly map cell phenotypes to TCN space without clustering. It enables both unsupervised and supervised identification of condition-specific TCNs, leveraging graph pooling and sample labels for improved accuracy. The algorithm was tested on various spatial omics datasets, including proteomics and transcriptomics, and demonstrated superior performance in identifying TCNs of varying sizes and in detecting condition-specific patterns, such as granulocyte-enriched and cancer-associated fibroblast-enriched TCNs in high-risk tumors. CytoCommunity also revealed altered interactions between neoplastic and immune/stromal cells within and between TCNs. The algorithm's ability to integrate cell-type information and use differentiable graph pooling allows it to address challenges in graph alignment and improve TCN detection. CytoCommunity outperformed existing methods in identifying TCNs in both spatial proteomics and transcriptomics data, particularly in distinguishing condition-specific TCNs in cancer tissues. The study highlights the importance of TCNs in understanding tissue function and disease progression, and CytoCommunity provides a scalable and effective approach for de novo identification of TCNs in various tissues.The article introduces CytoCommunity, a novel algorithm for identifying tissue cellular neighborhoods (TCNs) based on cell phenotypes and spatial distributions. Unlike existing methods that rely on clustering or intermediate steps, CytoCommunity uses a graph neural network (GNN) model to directly map cell phenotypes to TCN space without clustering. It enables both unsupervised and supervised identification of condition-specific TCNs, leveraging graph pooling and sample labels for improved accuracy. The algorithm was tested on various spatial omics datasets, including proteomics and transcriptomics, and demonstrated superior performance in identifying TCNs of varying sizes and in detecting condition-specific patterns, such as granulocyte-enriched and cancer-associated fibroblast-enriched TCNs in high-risk tumors. CytoCommunity also revealed altered interactions between neoplastic and immune/stromal cells within and between TCNs. The algorithm's ability to integrate cell-type information and use differentiable graph pooling allows it to address challenges in graph alignment and improve TCN detection. CytoCommunity outperformed existing methods in identifying TCNs in both spatial proteomics and transcriptomics data, particularly in distinguishing condition-specific TCNs in cancer tissues. The study highlights the importance of TCNs in understanding tissue function and disease progression, and CytoCommunity provides a scalable and effective approach for de novo identification of TCNs in various tissues.