PQG Seminar

Bo Xia
Assistant Professor, Harvard Medical School
Gene Regulation Observatory Fellow, Broad Institute

Predictive Genomics for Gene Regulation and Cell Fate Determination

How does the human genome encode gene activities to determine the thousands of cell types and functions? Genome regulation and cell fate determination are intrinsically multimodal, integrating DNA sequence, protein complexes, and their intricate interactions. Traditional experimental approaches for studying gene regulation, particularly in in vivo contexts, are frequently limited by sample availability, assay feasibility, scalability, and efficiencies. To accelerate the investigation of cell fate determination, we have built foundational multimodal genomics AI models—including C.Origami and Chromnitron—that understand the key principles of genome regulation and enable high-throughput in silico screens to accelerate discoveries. First, I will talk about C.Origami, a multimodal deep neural network that learned the rules of genome organization and thus enabled accurate prediction of chromatin interaction maps in unseen cell types. Applying an in silico screening strategy, we discovered two uncharacterized proteins that contribute to the core mechanism of chromatin domain formation. Second, I will talk about our recent development of a foundation model for studying global chromatin-associated proteins (CAPs). This multimodal deep neural network, Chromnitron, learned the key rules of how a protein binds to chromatin, such as base-resolution DNA sequence features, protein-DNA interaction features, and the chromatin’s biophysical background. Chromnitron model enables many new types of in silico investigation of CAPs in various conditions, such as characterizing unstudied CAPs, predicting the impact of non-coding variants, and discovering putative regulators of cell fate transition in unseen cell types.

Justin Tubbs
Instructor in Psychology in the Department of Psychiatry
Massachusetts General Hospital

Real-time Dynamic Updating of Polygenic Scores Improves Clinical Prediction and Utility

Polygenic risk scores (PRSs) are promising tools for advancing precision medicine. However, existing PRS construction methods rely on static summary statistics derived from genome-wide association studies (GWASs), which are often updated at lengthy intervals. With genetic data and health outcomes continuously being generated, the current PRS training and deployment paradigm is suboptimal in maximizing prediction accuracy for incoming patients in healthcare settings. We introduce real-time PRS-CS (rtPRS-CS), which enables online, dynamic refinement and standardization of PRS as each new sample is collected. Extensive simulation studies evaluate the performance of rtPRS-CS across various genetic architectures and training sample sizes. Leveraging quantitative traits from two large biobanks, we show that rtPRS-CS can integrate massive streaming data to enhance PRS prediction over time. We apply rtPRS-CS to schizophrenia cohorts across 7 Asian regions, demonstrating the clinical utility of rtPRS-CS in dynamically capturing health status changes and predicting disease risk across diverse genetic ancestries.