Lujia Chen

Precision Medicine Research Lab

University of Pittsburgh

Department of Biomedical Informatics


Integrate deep hierarchical and causal discovery models for discovering single-cell transcriptomic programs and molecular mechanisms of intercellular communication.


The transcriptome of individual cells is a mixture of genes regulated by active cellular signaling pathways, and gene expression models serve as the readout for downstream transcriptomic changes. A signaling pathway could be activated through the interaction of ligand and receptor (LR). Our lab aims to discover such signals and communications by
1) advancing the nHDP model tailored for inferring transcriptomic programs of individual cells and
2) developing discovery algorithms for searching population-wide cell-cell interactions mediated by LR interaction.


Develop computational tools for discovering individualized cell-cell communication networks (CCCNs)


Each instance (a tissue sample) has its unique cell composition and an instance-specific CCCN. Our lab aims to deconvolute the signals in the transcriptomes using high-resolution signals learned from single-cell transcriptomics. We build individualized causal model to detect the CCCNs in each instance and make the model more transparent and biologically explainable by incorporating the knowledge of biological components into the model. The identification of patterns in CCCNs could be used as informative features to explain tissue environment heterogeneity and evaluate their utility in precision medicine.


Build multimodal AI model to predict the drug response/clinical outcomes based on informative signatures learned from different data sources, such as tumor biopsy, liquid biopsy and whole image slides (WSI)


The integration of multiple data modalities in cancer research presents a promising avenue for improving patient outcomes through complementary features. By developing sophisticated deep learning architectures that can effectively combine and analyze these heterogeneous data sources, we aim to identify robust predictive signatures that could not be detected through single-modality analysis alone. The model learns to recognize complex patterns and relationships across these different data types, potentially uncovering novel biomarkers and improving our ability to predict treatment responses and patient outcomes. This integrated approach could significantly enhance clinical decision-making by providing more comprehensive and accurate predictions, ultimately leading to more personalized treatment strategies in cancer care.