FlowSP: AI-Powered Prediction of 3D Protein Distributions in Tissue

Stanford researchers have developed FlowSP, a generative AI model that predicts high-resolution 3D spatial maps of protein markers in tissue from limited imaging data. This technology enables researchers and clinicians to generate comprehensive protein information, including predictions of how tissues respond to genetic or therapeutic changes, without the time and cost of extensive experimental imaging.

Current prediction methods only work for 2D images and typically at low spatial resolution and limited number of protein markers. Researchers currently lack the tools to efficiently explore how proteins would change under different genetic or therapeutic conditions, forcing them to conduct costly trial-and-error experiments. Existing AI approaches using older generative models such as GANs cannot scale to capture the complexity of 3D protein distributions or predict how individual proteins respond to specific perturbations.

FlowSP uses a novel flow-based generative model, a type of AI architecture more powerful than previous approaches, trained on massive amounts of tissue imaging data to learn biologically meaningful patterns of how proteins co-express and distribute in 3D space. The model can accept multiple types of input conditions (tissue type, species, genetic knockouts, or therapeutic treatments as text descriptions, as well as partial imaging data), and then generates complete, high-resolution 3D protein maps at subcellular resolution. In a single model, it can predict over 100 different protein markers across multiple tissue types and species, and uniquely enables researchers to simulate "in silico perturbations" and forecast how protein distributions would change under genetic or drug interventions to prioritize targets for wet lab testing. The approach extends to other spatial omics modalities, including spatial transcriptomics.

Stage of Development: Prototype