AI Foundation Model Maps Cellular Interactions for Spatial Genomics

A new AI foundation model, the Cellular Interaction Foundation Model (CI-FM), has been developed to analyze and simulate complex cellular interactions within tissues. Leveraging geometric graph neural networks and self-supervised learning, CI-FM predicts gene expressions based on a cell’s microenvironment. The model was trained on 23 million cells, achieving high correlation with experimental data and a low error rate (MSE of 1.1%). It successfully identified tumor microenvironment signatures and simulated immune cell responses, such as T-cell infiltration, a key factor in disease progression. This scalable framework advances AI-driven spatial genomics, paving the way for modeling tissue-level cellular responses and enhancing our understanding of disease mechanisms.