Abstract
Pancreatic cancer (PDAC) is a lethal disease in part because tumor cells exist in distinct transcriptional states (e.g. basal/mesenchymal v.s. classical/epithelial) with unique phenotypic properties that contribute to tumor growth and treatment resistance. Two major mechanisms have been suggested for treatment evasion: (1) the intrinsic resistance of an existing state to a therapy regimen and (2) plasticity of therapy-sensitive states to adopt more resistant states. The relative contribution of these mechanisms to treatment resistance is still poorly understood. Historically, measurement of plasticity in both human patients and mouse models has involved one of three principles: (1) observing a redistribution of cell states in tissue across timepoints or conditions; (2) identifying cells that have genomic, epigenetic or proteomic features of more than one state (mixed states); and (3) performing single-cell cloning of cells and observing the cell states adopted by clonal progeny. While these approaches are observationally consistent with the notion of plasticity, they either fail to definitively prove the existence of plasticity, are restricted in measurements of plasticity outside of native tissues or are unable to quantify the role of plasticity in treatment resistance.
Amongst the most well described forms of plasticity in human development and cancer is epithelial-mesenchymal plasticity (EMP), which includes epithelial to mesenchymal transition (EMT) and mesenchymal to epithelial transition (MET). To better understand and quantify the role of EMP in driving treatment resistance of human PDAC, we have developed single-cell multiomic, functional genomic and computational methods applied to patient-derived models and clinical biopsies. We first profiled twelve patient-derived PDAC cell lines by single-cell RNA-seq (scSeq) and learned convergent epithelial and mesenchymal gene programs that were consistent with programs observed in patient samples. We next performed lineage tracing experiments in three PDAC cell lines using an expressed lentiviral barcoding system (ClonMapper). By performing scSeq on these barcoded lines at weekly timepoints over four weeks, we proved the presence of EMP by showing a single cell can produce progeny in both epithelial and mesenchymal states.
We next developed a generative probabilistic model of our lineage tracing data. This demonstrated that clones (cells sharing a barcode) had different transition matrices (different EMT and MET rates), thus suggesting each clone has a distinct level of plasticity. Having established this, we focused on identifying genes that might explain the differing plasticity properties of clones. Using elastic net regression we identified 50 transcription factors (TFs) whose expression significantly explained the propensity for EMP over time across clonal populations. Among these were were several known EMP TFs (Zeb1 and Gata6), understudied TFs (Elf3, Sox2, Sox4, Klf3, Klf5 and Atf4) and novel TFs (Meis2, Meis3, FoxA1 and the interferon regulatory factors Irf6, Irf7 and Irf9). Using single-cell multiomics (paired scSeq and single-cell ATAC-seq) on our barcoded population, we found that 9 of the 50 predicted TFs, including Elf3, had differential accessibility between clones with different plasticity properties, suggesting a role for epigenetic regulation of these TFs in facilitating EMP. Importantly, we leveraged our multiomic data to infer gene regulatory networks influenced by these TFs and found an enrichment of binding motifs for these TFs in enhancer regions of genes in epithelial and mesenchymal programs.
To study the role of these predicted TFs in modulating EMP, we developed a CRISPRi system that enabled gene perturbations alongside lineage tracing. We performed a CRISPRi perturb-seq experiment (CRISPR perturbation with scSeq readouts), perturbing the 50 predicted TFs above and 10 control genes, and collected 1.5 million single-cell transcriptomic profiles, in addition to two other CRISPR KO perturb-seq experiments. We performed a negative binomial regression to estimate effect sizes of guide RNAs on all genes. 60% of our guides had significant perturbation effects on their target gene. We subsequently found Klf3 as an important regulator of PDAC proliferation independent of cell state. Importantly, we found that knockdown (KD) of several factors, such has Grhl2 and FoxA1, bias towards mesenchymal cell states, whereas KD of others such as Batf2, Snai1, Rel, Zeb1, Nr2f1 and Sox4 led to a bias towards epithelial cell states. Interestingly, KD of several TFs influenced transition properties of cells by decreasing rates of EMT and MET across barcodes without biasing the overall clonal distribution towards a single cell state. This suggests a role for these TFs in enabling plasticity and facilitating state transitions.
To study the effect of plasticity in treatment resistance, we treated four barcoded cell lines with the first-line chemotherapy combination FOLFIRINOX (5-fluorouracil, oxaliplatin and SN-38, the active metabolite of irinotecan) or targeted therapy and performed scSeq yielding over 600,00 single-cell transcriptomic profiles. We found an enrichment after treatment with FOLFIRINOX of barcodes that were biased for cells in mesenchymal states, consistent with selection against epithelial cells, but also of those barcodes with the highest inferred state transition rates. With targeted therapies, we found selective depletion of mesenchymal states. We next treated our CRISPR perturbed cell lines and found overall a significantly more restricted barcode diversity in cells containing guide RNAs targeting plasticity factors compared to non-targeting controls, suggestive of the role of plasticity in facilitation resistance.
To validate the role of these proposed plasticity factors in human patients, we collected paired biopsy samples from 23 patients in a phase 2 clinical trial of metastatic PDAC patients being treated with radiation therapy and dual checkpoint blockade (NCT03104439). We performed scSeq on these samples, and used a supervised Bayesian matrix factorization approach (Spectra) to learn epithelial and mesenchymal gene programs within tumor cells. We subsequently classified cells as epithelial, mesenchymal and intermediate cell types using a gaussian mixture model on gene expression features. We found the intermediate states were enriched in expression of our proposed plasticity factors, and importantly high expression of these factors in baseline samples correlated with a redistribution of states in follow-up biopsies.
Our efforts define a robust experimental and quantitative framework for studying tumor cell plasticity in patient-derived model systems with validation in human patient samples using single-cell and spatial transcriptomics. Collectively, we nominate several regulators that alter the propensity of EMP in PDAC, thus posing a paradigm whereby perturbations may be used to homogenize tumor populations towards treatment-sensitive phenotypes for combination therapy.
Citation Format: Arnav Mehta, Lynn Bi, Deepika Yeramosu, Michael Bogaev, Martin Jankowiak, Abigail Collins, Aziz Al'Khafaji, Milan Parikh, Mehrtash Babadi, Kyle Evans, Alex Bloemendal, Russell Kunnes, Marc Schwartz, Glen Munson, Elisa Donnard, Thouis R. Jones, Ben Z. Stanger, Jay Shendure, Jonathan Weissman, David T. Ting, Andrew Aguirre, Nir Hacohen, Dana Pe'er, Eric S. Lander. Dissecting and quantifying pancreatic cancer plasticity using single-cell multiomics, lineage tracing and functional genomics reveals novel mediators of therapy resistance [abstract]. In: Proceedings of the American Association for Cancer Research Annual Meeting 2024; Part 2 (Late-Breaking, Clinical Trial, and Invited Abstracts); 2024 Apr 5-10; San Diego, CA. Philadelphia (PA): AACR; Cancer Res 2024;84(7_Suppl):Abstract nr NG08.
Search from 250 M+ research papersSearch from 250 M+ research papers
Open Access
Just Published
Apr 13, 2024
Yoshiki Shinomiya + 7