Census Models

PINNACLE is a flexible geometric deep learning model that is trained on protein interaction networks contextualized by 156 cell types from a human single-cell transcriptomic atlas in order to generate context-aware protein representations. It generates 394,760 protein representations (~18X more than existing methods) that reflect the organization of 156 cell types spanning 62 tissues. We demonstrate that PINNACLE’s contextualized protein representations 1) enable zero-shot retrieval of the tissue hierarchy, 2) improve 3D structure predictions of PD-1/PD-L1 and B7-1/CTLA-4 protein interactions, two critical immune checkpoint interactors targeted by cancer immunotherapies, and 3) outperform state-of-the-art, yet context-free, models in nominating therapeutic targets of RA and IBD while pinpointing their most predictive cell type contexts.

Neurological disorders are the leading driver of global disability and cause 16.8% of global mortality. Unfortunately, most lack disease-modifying treatments or cures. To address disease complexity and heterogeneity in neurological disease, we developed scCIPHER, an AI approach for Contextually Informed Precision HEalthcaRe using deep learning on single-cell knowledge graphs. We created the Neurological Disease Knowledge Graph (NeuroKG), a neurobiological knowledge graph with 132 thousand nodes and 3.98 million edges, by integrating 20 high-quality primary data sources with single-cell RNA-sequencing data from 3.37 million cells across 106 regions of the adult human brain. Next, we pre-trained a heterogeneous graph transformer on NeuroKG to create scCIPHER. We leverage scCIPHER to make precision medicine-based predictions in neurological disorders across patient phenotyping, therapeutic response prediction, and causal gene discovery tasks, with validation in large-scale patient cohorts.

A major question in cancer biology is how cancer cells rewire the regulatory network to support uncontrollable cell growth. We developed a computational framework, D-SPIN, that decodes regulatory network models of single-cell profiling from multiple samples and various conditions. With a collection of healthy and cancerous tissue profiling from Census data, we identified a hundred gene programs shared between healthy and cancer tissues, constructed regulatory network models of the cell population, and observed a series of regulatory shifts in key functions such as energy metabolism and cell adhesion in cancer tissues. Our study can help identify network-level signatures of cancer progression, and propose critical regulatory shifts as targets of therapeutics.