An artificial intelligence system can identify patterns in the arrangement of cells and proteins that are associated with cancer outcomes, while showing which features influence its predictions.
The method, called SpaCEy, was published in Nature Communications on September 26. It tackles a challenge in spatial biology: connecting molecular maps of tissue with what happens to patients.
SpaCEy represents tissue as a graph, allowing a neural network to learn relationships between neighboring cells and their protein markers. It does not require predefined cell categories or anatomical regions as inputs.
An explanation component identifies spatial patterns and combinations of markers relevant to the model’s predictions. This helps researchers examine the tissue features behind a result.
In a lung cancer cohort, the system identified spatial and protein expression patterns associated with disease progression. Across breast cancer datasets, it separated patients into groups with different overall survival, including within clinical subtypes.
These are research findings about associations and prediction, rather than evidence that the system improves treatment decisions.
The approach offers researchers a way to investigate how cellular neighborhoods contribute information beyond molecular measurements.