Modern machine learning (ML) models can accurately predict patient progress, an individual's phenotype, or molecular events such as transcription factor binding. However, they do not explain why selected features make sense or why a particular prediction was made. For example, a model may predict that a patient will get chronic kidney disease, which can lead to kidney failure. The lack of explanations about which features drove the prediction – e.g., high systolic blood pressure, high BMI, or others – hinders medical professionals in making diagnoses and decisions on appropriate clinical actions. I will briefly describe my group’s efforts to develop interpretable ML techniques for varied biological and medical applications, including treating cancer based on a patient’s own molecular profile, identifying therapeutic targets for Alzheimer’s, predicting kidney diseases, preventing complications during surgery, enabling pre-hospital diagnoses for trauma patients, and improving our understanding of pan-cancer biology and genome biology. My talk will focus in greater detail on interpretable ML techniques to identify molecular markers for anti-cancer drugs for acute myeloid leukemia, published in Nature Communications (Jan 2018) and our more recent works in collaboration with Harvard Medical School; our explainable artificial intelligence system, Prescience, for preventing hypoxemia in patients under anesthesia, featured on the cover of Nature Biomedical Engineering (Oct 2018); SHAP, our general ML framework on model interpretability, published as a full oral presentation at NeurIPS (2017; cited >500 over 2 years after publication); Tree explainer, our polynomial time algorithm for computing SHAP values for tree models, published in Nature Machine Intelligence (Jan 2020) and our ML approach to identifying Alzheimer's disease therapeutic targets.

Although the immunogenicity of tubo-ovarian high-grade serous cancer (HGSC) is well documented, responses to immunotherapy for this malignancy have been disappointing. To select patients likely to gain the most benefit from these therapies a deeper understanding of the HGSOC tumor immune microenvironment is required, the ultimate aim being to identify predictive mechanistic markers.

We recently published the first CyTOF study of newly diagnosed HGSC tumors that focused only on the tumor cells. We have now analyzed the CyTOF data of the tumor immune cell infiltrate within with the same set of tumors.

To examine relationships between tumor and immune cell types we performed extensive pairwise correlation analyses. In addition to observing positive correlations between tumor cell types and exhausted T cell sub-populations, we observed strong positive correlations between NK cell subpopulations with overall tumor cell abundance and with tumor cells co-expressing E-cadherin and vimentin (EV cells). The latter are cells are likely undergoing epithelial to mesenchymal transition and so contributing to tumor progression.

Intriguingly, the phenotype of the NK cell sub-populations resembled decidual NK cells. Decidual NK cells play a critical role in early pregnancy by conferring immune tolerance toward the hemi-allogeneic fetus. Our analysis suggests that the same features of immune tolerance could be subverted for HGSC tumor maintenance and/or progression from the epithelial to metastatic state.

I will discuss some of our mechanistic insight into how this could be happening and critically, the implications of for NK cell immunotherapy.

Tumor-infiltrating lymphocytes (TIL) are associated with survival in virtually every human cancer, but the mechanisms by which they confer protective immunity remain incompletely understood. Focusing on ovarian cancer, our group applies genomic and molecular pathology approaches to define the mechanisms by which the human immune system responds to the evolving tumor genome over space and time. We have evidence that T cell clones track tumor clones over space and time and apply selective pressure that leads to reduced tumor clone diversity and progressive loss of immune recognition through several mechanisms. We find that optimal anti-tumor immunity involves interactions between T cells, antibody-producing B-lineage cells and macrophages in the tumor microenvironment. Our findings suggest new approaches to immunotherapy that overcome the inherent limitations of the T cell response.

Natural killer (NK) cells comprise one subset of the innate lymphoid cell (ILC) family. Despite reported anti-tumor functions of NK cells, their tangible contribution to tumor control in humans remains controversial. This is due to incomplete understanding of the NK cell states within the tumor microenvironment. Using single-cell RNA-sequencing and in vitro and in vivo modeling, we found that peripheral circulating NK cells differentiate down two divergent pathways within the tumor microenvironment, resulting in different end-states. These data and the phenotype of these end states will be discussed.

Systems biology approaches, including computational models, provide a framework to test biological hypotheses and optimize effective therapeutic strategies to treat human diseases. My research group applies systems biology methods to quantitatively understand the dynamics of key signaling and metabolic networks in cancer. In particular, we study: 1) Cancer immunotherapy: the signaling pathways that regulate chimeric antigen receptor (CAR) mediated T cell activation required for therapeutic strategies; 2) Tumor angiogenesis: the effects of crosstalk amongst key proteins involved in regulating angiogenesis, the formation of new blood vessels; and 3) Cancer metabolism: central carbon metabolic pathways in cancer cells. Our modeling provides a quantitative framework to improve our understanding of cancer.

In this talk, I will present our latest work in modeling signaling in CAR T cells. Chimeric antigen receptors (CARs) are engineered receptors that mediate T cell activation. To do so, CARs are comprised of a variety of different activating domains (such as CD3ζ) and co-stimulatory domains (such as CD28) derived from endogenous T cells. These intracellular domains initiate signaling required for T cell activation, including ERK activation through the MAPK pathway. Although in vitro and preclinical studies show that the addition of co-stimulatory domains on CARs enhances T cell activity, the mechanisms by which this happens are not clear. Therefore, we have constructed a computational mechanistic model of CAR-mediated activation of ERK in T cells. We apply mathematical modeling to improve our understanding of how CAR structure influences downstream T cell signaling and to develop new hypotheses for the optimal design of CAR-engineered T cell systems.

Tumors are complex ecosystems of malignant, immune, and stromal elements whose dynamic interactions underlie patient survival and response to therapy. We applied CIBERSORTx, a recently described “digital cytometry” framework, to impute cell type-specific gene expression profiles of cancer cells and 11 major immune and stromal cell types from thousands of solid tumor transcriptomes profiled by The Cancer Genome Atlas (TCGA). We then devised and applied a novel machine learning framework to identify distinct transcriptional states for each cell type, validate them in scRNA-seq data, and uncover co-occurrence patterns between them in order to define tumor cellular ecosystems. We defined robust transcriptional states of 12 cell types, including epithelial, fibroblast, endothelial, and 9 immune subsets. These states included both known and novel cellular phenotypes, nearly all of which could be validated in published scRNA-seq tumor atlases. Further analysis of these communities unveiled unique cellular ecosystems that span most solid tumor types. Large-scale deconvolution of cell type-specific transcriptomes across thousands of solid tumors revealed a comprehensive atlas of cellular states in the TME. The majority were present across multiple tumor types. Specific cell states tend to co-occur in a pan-cancer manner, defining robust tumor cellular ecosystems with specific biological features and clinical outcomes. Overall, our results provide a high resolution portrait of cellular heterogeneity in the TME of multiple solid tumor types, with implications for novel diagnostics and immunotherapeutic targets

Exemplified by CRISPR gene-editing and single-cell sequencing, novel genomics technologies are a powerful collection of tools for probing the hidden mechanisms of cancer. My group is developing gene-editing and single-cell tools for studying dynamics of cancer cell evolution. Our recent work has been focusing on using novel Cas12a tools with single-cell genomics in cancer models, with a focus on studying the cellular regulation that lead to genetic and epigenetic drug resistance (BRAFi/MEKi). We aim to make them broadly useful for identifying mechanisms and new drug targets in a range of cancer types for future applications.

Combining cell cycle inhibitors with endocrine therapy improves outcomes of metastatic estrogen receptor positive (ER+) breast cancer patients. However, potentially curable earlier stage patients show mixed results to this drug combination. To evaluate this combination therapy regimen in early stage ER+ tumors, we built mathematical models to analyze evolutionary trajectories of patients’ tumors from a clinical trial using DNA and single cell transcriptomic profiling performed before, during and after treatment with endocrine therapy (letrozole) alone or in combination with the cell cycle inhibitor ribociclib. We established cancer phenotypes related to resistance with Gaussian process and hierarchical regression models and reconstructed tumor phenotypic evolution and cell cycle rewiring through integrative Gaussian mixture and hierarchical generalized additive modeling.

Independent of patient response, cancer cells undergo subclonal evolution during therapy. However, while endocrine therapy alone is associated with reduced tumor diversity over time combination therapy shows an increase. Despite subclonal heterogeneity, single nuclei RNA sequencing uncovered common phenotypic changes during treatment, including accelerated loss of estrogen signaling with convergent upregulation of the JNK pathway in tumors with progressive growth on combination but not endocrine-only therapy. In contrast, patient tumor cells that maintain estrogen signaling during therapy show potentiation of CDK4/6 activation consistent with ERBB4 and ERK upregulation. Lastly, cell cycle reconstruction identified that non-responder tumor cells can rebound during treatment, indicating stronger selection and promotion of a proliferative state. Taken together our results indicate that in early stage patients, combination therapy drives evolution of resistance via a shift from estrogen signaling to alternative growth factor receptor and JNK signaling concordant with a bypass in the G1 checkpoint.

The introduction of single-cell high-throughput measurement technologies (such as single-cell RNA-seq) have recently transformed the breadth and depth at which the immune component within tissues can be characterized and studied. Specifically, the ability to characterize heterogeneity of cellular states within tissues and the changes that occur following disease or therapeutic intervention are enabling the formation of novel prognostic tools as well as hypotheses regarding cellular function. Excitingly, the recent ability to generate coupled RNA-seq and TCR / BCR data at single-cell resolution and at high-throughput (using 10X technology) has enabled exploring relationships between cellular states and clonal distributions within and across tissues. 

In this talk we will discuss systemic aspects of T cell immunology during cancer and autoimmune disease, as they are revealed with coupled single-cell RNA-seq and TCR annotation. We will present ongoing works in which we assess the extent to which peripheral blood can be used for tracking a host response to tumor by characterization of the CD8+ T cell tumor-directed component in blood, and touch on a project characterizing Th17 T cell heterogeneity and migration during homeostasis and EAE disease in mouse models. We will explain how COMET, a computational tool developed by the Singer group to identify marker panels for cell populations of interest from single-cell RNA-seq data, was useful in both of the presented works. 

Tumors employ complex, multi-scale cellular and molecular interactions that evolve over the course of therapeutic response. The changes in these pathways enables tumors to overcome therapeutic regimens, and ultimately acquire resistance. New molecular profiling technologies, including notably single cell technologies, provide an unprecedented opportunity to characterize these molecular relationships. However, interpreting the specific cellular and molecular pathways in therapeutic response requires complementary computational analysis methods. We developed an unsupervised learning method, CoGAPS, that employs Bayesian non-negative matrix factorization to disentangle distinct biological processes from high-throughput molecular data. Notably, this algorithm discovers dynamic compensatory signaling in acquired therapeutic resistance from time course bulk RNA-seq data and novel NK cell activation in anti-CTLA4 response from post-treatment scRNA-seq data. To further demonstrate that the inferred pathways are biological rather than computational artifacts, we developed a complementary transfer learning method to relate learned patterns between datasets. We demonstrate that this approach identifies robust molecular processes between model systems and human tumors and enables multi-platform data integration to delineate the drivers of therapeutic response and resistance.