Current Research Directions

Multi-omic deconvolution to study DNA–RNA dynamics in cancer

Cancer is driven by genetic mutations, including single nucleotide variations (SNV), copy number alterations (CNA) and structural variations (SV), which influence tumor behavior, such as growth rate, treatment resistance and metastasis. Identifying these mutations is critical for cancer research. While whole-genome sequencing (WGS) and whole-exome sequencing (WES) are key tools, basic steps like somatic mutation calling can be slow, limiting large-scale analysis. My lab is addressing this with [MuSE2], a fast and efficient mutation calling method that facilitates large dataset analysis and advances precision medicine.

We are also interested in improving methods for reconstructing subclonal structures, which are critical for understanding cancer evolution and treatment resistance. Our effort on developing software tools like [CliPP] helps overcome limitations in previous methods by significantly reducing computational resources and time, through penalized likelihoods [Characterizing ITH]. These advancements are critical to understanding intratumor heterogeneity and cancer evolution, providing important evidence for translational research to improve patient outcomes.

Tissues, including tumors, contain diverse cell types, each with unique transcriptional patterns that can be studied through RNA expression data. While single-cell RNA sequencing (scRNA-seq) provides detailed insights, it is often costly and challenging for large-scale use. Bulk RNA-seq is more affordable but mixes signals from different cell types. To address this, deconvolution methods like [DeMixSC] help separate these signals, improving analysis of cell proportions and disease mechanisms. In cancer research, deconvolution differentiates tumor from non-tumor cells, offering insights into pathways, prognosis, and heterogeneity [DeMixT].

We further developed an integrative transcriptomic/genomic deconvolution method to calculate [TmS] (tumor-specific total mRNA expression), a feature of cancer cell plasticity, with a striking ability to predict prognosis across cancers. Spatial transcriptomics data builds on this by adding another dimension, preserving the spatial arrangement of cells to help map tumor microenvironments (TME). This spatial context provides crucial insights into how cells interact within their environments, which is essential for understanding tumor progression. We recently developed DeMixNB to characterize spatial distributions of tumor-specific gene expression. By integrating bulk, single-cell and spatial data, we can achieve deeper insights, advancing more effective and personalized cancer treatment strategies.

Cancer risk modeling (TP53) using machine learning and Bayesian models

Cancer survivors represent a fast-growing yet under-studied population with respect to cancer risk, particularly for second primary cancers, which frequently occur in survivors of breast and bladder cancer. Current risk assessments often overlook prior cancers due to limitations in large databases like SEER, which mainly account for age and sex. To address this, my lab studies patients with Li-Fraumeni syndrome (LFS), a hereditary condition linked to higher cancer risk. LFS patients often develop multiple primary cancers, offering a unique opportunity to study cancer risk while accounting for additional factors like mutation status. Using LFS data, we developed [LFSPRO] to predict both first and second primary tumors in LFS families. These insights can help physicians and genetic counselors provide personalized treatment and screening plans, aiming for early detection of cancers in survivors and LFS patients' personalized risk prediction.

We are also particularly interested in the biological annotation of TP53 mutations, as the germline mutations of TP53 are the main cause of LFS. Known as the “guardian of the genome,” the TP53 gene plays a critical role in cell signaling, apoptosis, metabolism, DNA repair and transcription, and in the meantime it is the most frequently mutated gene in human cancer. We developed survival-based clustering of predictors [SCP] using penalized likelihoods for survival outcomes to cluster hundreds of TP53 missense mutations in terms of their associated early, medium and late onset of cancer in LFS. This research aims to uncover new patterns in cancer susceptibility and improve predictive models, offering deeper insights into the genetic underpinnings of cancer risk in LFS patients.