Artificial Intelligence-Driven Morphology-Based Enrichment of Malignant Cells from Body Fluid

Cell morphology is a fundamental feature used to evaluate patient specimens in pathologic analysis. However, traditional cytopathology analysis of patient effusion samples is limited by low tumor cell abundance coupled with the high background of nonmalignant cells, restricting the ability of downstream molecular and functional analyses to identify actionable therapeutic targets. We applied the Deepcell platform that combines microfluidic sorting, brightfield imaging, and real-time deep learning interpretations based on multidimensional morphology to enrich carcinoma cells from malignant effusions without cell staining or labels. Carcinoma cell enrichment was validated with whole genome sequencing and targeted mutation analysis, which showed a higher sensitivity for detection of tumor fractions and critical somatic variant mutations that were initially at low levels or undetectable in presort patient samples. Our study demonstrates the feasibility and added value of supplementing traditional morphology-based cytology with deep learning, multidimensional morphology analysis, and microfluidic sorting.

Abstract 5926: High dimensional morphology analysis reveals new insights in melanoma cell heterogeneity

Melanomas are the deadliest skin cancers, in part due to cellular plasticity and heterogeneity within the tumors. These characteristics have made a deeper understanding of melanomas challenging. Classically, melanoma cells are characterized with a limited set of protein biomarkers. Gene expression signatures and mutational analysis (e.g., BRAF and NRAS genotyping) can provide a more detailed view of heterogeneity but may not translate to readily available biomarkers for functional studies. The Deepcell platform enables multi-dimensional morphology analysis and enrichment of unlabeled single cells using artificial intelligence (AI), advanced imaging, and microfluidics, enabling high resolution profiling of population heterogeneity. Label-free multi-dimensional morphology analysis may have higher resolution than a limited set of protein biomarkers, minimizes perturbation to the transcriptome, and reduces cell handling steps. We used patient-derived cell lines and dissociated biopsy samples to train a Deepcell AI model to identify and sort for melanoma cells based on morphology alone. The model was tested on metastatic melanoma biopsies, with identification and enrichment of melanoma cells verified by various downstream assays, including scRNASeq. In addition to melanoma populations, the AI model classified various cells of the microenvironment, such as stromal cells and immune subtypes, based on morphology only. To further characterize tumor heterogeneity, we imaged >25 patient-derived melanoma cell lines representing melanocytic, mesenchymal, and intermediate phenotypic states on the Deepcell platform. Morphology analysis of these images revealed distinct clusters of cells for each phenotype, indicating that there are morphological differences associated with each state. We developed a random forest (RF) classifier to identify the top differential morphological features between the different cell lines, thereby providing a label-free means of phenotyping melanoma samples. The morphology analysis of the cell lines uncovered significant variability in pigmentation; an RF classifier distinguished between pigmented vs non-pigmented cells with >90% accuracy. Pigmentation is a hallmark of melanoma cells, and it has been associated with the melanocytic phenotype and differential drug response in vitro. However, there is not currently a robust method to profile and study pigmentation in live cells. We further investigated this observation by correlating morphological profiles, molecular, and functional information with the level of cell pigmentation. The Deepcell platform presents a new method for sorting and characterizing cellular heterogeneity using morphology, including pigmentation status. As such, multi-dimensional morphology analysis will bolster the understanding of complex melanoma states and tumor microenvironment, particularly in patient derived biopsies.

Citation Format: Evelyn Lattmann, Aizhan Tastanova, Andreja Jovic, Kiran Saini, Tiffine Pham, Christian Corona, Jeanette Mei, Michael Phelan, Stephane C. Boutet, Ryan Carelli, Kevin B. Jacobs, Julie Kim, Manisha Ray, Chassidy Johnson, Nianzhen Li, Mahyar Salek, Maddison Masaeli, Mitchell P. Levesque. High dimensional morphology analysis reveals new insights in melanoma cell heterogeneity [abstract]. In: Proceedings of the American Association for Cancer Research Annual Meeting 2023; Part 1 (Regular and Invited Abstracts); 2023 Apr 14-19; Orlando, FL. Philadelphia (PA): AACR; Cancer Res 2023;83(7_Suppl):Abstract nr 5926.

Abstract 2392: Multi-dimensional morphology analysis enables identification and label-free enrichment of heterogeneous tumor cell populations

Studying tumor heterogeneity at single cell resolution is crucial for elucidating tumor progression and treatment response. Classically, tumor cells are characterized with a limited set of biomarkers that cannot cover the full extent of the tumor cell heterogeneity. The advent of single cell genomics has enabled scientists to study heterogeneity at high resolution, but gene expression signatures may not translate to readily available biomarkers for functional studies. Multi-dimensional morphology analysis has the potential for higher resolution than cell surface biomarkers while also reducing cell manipulations. The Deepcell platform enables multi-dimensional morphology analysis of unlabeled single cells using artificial intelligence (AI), imaging, and microfluidics, allowing for higher resolution of population heterogeneity beyond protein expression markers. Using the Deepcell platform, we trained an AI model to identify and sort for malignant cells from non-small cell lung cancer cells (NSCLC) tumor biopsies based on morphology. scRNA-seq, CNV, and targeted mutation analysis verified enrichment of carcinoma cells. Sorted cells had gene expression signatures indicative of tumor cells (e.g. EpCAM expression), increased genomic alterations by CNV analysis, and increased allele frequency of P53 and KRAS mutations relative to pre-sorted samples. Overall, the data indicate brightfield images of cells can be used to detect macro-level changes in cell morphology resulting from the molecular events involved in tumorigenesis. In addition, we induced chemotherapeutic drug resistance of lung cancer cell lines in vitro, imaged the resulting cultures on the Deepcell platform, and used the bright-field cell images to generate multi-dimensional morphological embeddings that can be visualized by UMAP. The resulting UMAPs showed distinct clusters of cells for both the parent and resistant cell lines, indicating that there are morphological differences associated with drug resistance. We developed a random forest classifier to identify the top differential morphological features between parental and resistant cell lines. These top features can distinguish between the populations with high accuracy (87%). Finally, we performed multi-dimensional morphology analysis to compare lung adenocarcinoma (LA) and squamous cell carcinoma (SCC) cell lines. The classifier predicted the correct cell type with 75% accuracy, suggesting that differences between these two carcinomas are reflected in their morphology. Together, our data suggest that AI-detected morphological differences between cell populations may show a biologically significant link between morphology and phenotype. As such, multi-dimensional morphology analysis will bolster the understanding of complex disease states and tumor microenvironment, particularly in patient derived biopsies.

Citation Format: Andreja Jovic, Kiran Saini, Ryan Carelli, Tiffine Pham, Christian Corona, Jeanette Mei, Michael Phelan, Stephane C. Boutet, Kevn Jacobs, Julie Kim, Manisha Ray, Chassidy Johnson, Nianzhen Li, Mahyar Salek, Maddison Masaeli, Matt Barnes, Cyril Ramathal. Multi-dimensional morphology analysis enables identification and label-free enrichment of heterogeneous tumor cell populations [abstract]. In: Proceedings of the American Association for Cancer Research Annual Meeting 2023; Part 1 (Regular and Invited Abstracts); 2023 Apr 14-19; Orlando, FL. Philadelphia (PA): AACR; Cancer Res 2023;83(7_Suppl):Abstract nr 2392.

Scalable and comprehensive characterization of antigen-specific CD8 T cells using multi-omics single cell analysis

Understanding the antigen binding specificities of lymphocytes is key to the development of effective therapeutics for cancers and infectious diseases. Recent technological advancements have enabled the integration of simultaneous cell-surface protein, transcriptome, immune repertoire and antigen specificity measurements at single cell resolution, providing comprehensive, scalable, high-throughput characterization of immune cells.

Using the 10x Genomics Single Cell Immune Profiling Solution with Feature Barcoding technology with 14 oligo-conjugated antibodies and 50 Immudex peptide-MHC I Dextramer reagents (pMHC) panels spanning different CMV, EBV, Influenza, HIV and Cancer antigens, we performed multi-omic characterization of ~100,000 CD8+ T cells from four MHC-matched donors. The multi-omic combination of gene expression, paired alpha/beta T cell receptor (TCR) repertoire, cell surface proteins and pMHC binding specificity allowed the identification of CD8+ T cell subpopulations with specificity for pMHCs within our panel. We observed multiple TCRs that bound the same pMHC and identified enriched amino acid motifs within TCR sequences that shared specificities. We compared the CDR3 amino acid sequences of the pMHC-specific TCR clonotypes with previously reported sequences with the same binding specificities to show that we could identify new and known CDR3 sequences. This analytical framework provides a systematic and scalable method for deciphering TCR–pMHC specificity combined with cellular phenotype identity which is critical for developing a better understanding of the adaptive immune response to cancer and infectious diseases and will be key in the development of successful immunotherapies.

Deep characterization of tumor microenvironments using single cell multi-omics analysis

Understanding the complexity of cell interactions in the tumor microenvironment (TME) requires the ability to distinguish each cell type, and is a prerequisite of personalized cancer treatments. Here, we use a fully integrated system for single cell RNA sequencing, to simultaneously profile the transcriptome, cell surface proteins and immune repertoire of cells from primary colorectal cancer (CRC), non-small cell lung cancer (NSCLC), and mucosa-associated lymphoid tissue (MALT) lymphoma. Each tumor varied in type and proportion of its cellular components, and in particular in the proportion of immune cells. The CRC tumor consisted of T (3% CD4+, 3% CD8+), B lymphocytes (5% CD79A+) and plasma B cells (11% IGH high). Repertoire sequencing identified a clonal expansion (>4% of B cell clonotypes) suggesting a strong B cell response in this tumor. The NSCLC tumor displayed a marked immune cell infiltration containing predominantly B cells (30% CD79A+ and 8% IGH high plasma B) with a very limited clonal expansion. The MALT lymphoma consisted of only T and B lymphocytes (31% CD4+, 8% CD8+, 57% CD79A+). In addition to the activated CD4+ and CD8+ T cells, Tfh and Treg cells were clearly identified as well as two distinct large B cell populations showing plasmacytic differentiation. Analysis of the B cell repertoire revealed a large expanded clone bearing an IGHV segment associated with parotid MALT lymphoma. These findings emphasize the importance of combining repertoire and gene expression sequencing data to determine the nature and clonality of an immune response. This method enables full characterization of tumor heterogeneity and the adaptive immune response to the TME and will be key in the development of successful immunotherapies.

An artificial niche preserves the quiescence of muscle stem cells and enhances their therapeutic efficacy

A promising therapeutic strategy for diverse genetic disorders involves transplantation of autologous stem cells that have been genetically corrected ex vivo. A major challenge in such approaches is a loss of stem cell potency during stem cell culture. Here we describe a system for maintaining muscle stem cells (MuSCs) in vitro in a potent, quiescent state. Using a machine learning method, we identified a molecular signature of quiescence and used it to screen for factors that could maintain mouse MuSC quiescence, thus defining a quiescence medium (QM). We also designed artificial muscle fibers (AMFs) that mimic the native myofiber of the MuSC niche. Mouse MuSCs maintained in QM on AMFs showed enhanced potential for engraftment, tissue regeneration and self-renewal after transplantation in mice. An artificial niche adapted to human MuSCs showed similarly prolonged quiescence in vitro and enhanced potency in vivo. Our approach for maintaining quiescence may be applicable to stem cells from a range of other tissues.

The eSNV-detect: a computational system to identify expressed single nucleotide variants from transcriptome sequencing data

Rapid development of next generation sequencing technology has enabled the identification of genomic alterations from short sequencing reads. There are a number of software pipelines available for calling single nucleotide variants from genomic DNA but, no comprehensive pipelines to identify, annotate and prioritize expressed SNVs (eSNVs) from non-directional paired-end RNA-Seq data. We have developed the eSNV-Detect, a novel computational system, which utilizes data from multiple aligners to call, even at low read depths, and rank variants from RNA-Seq. Multi-platform comparisons with the eSNV-Detect variant candidates were performed. The method was first applied to RNA-Seq from a lymphoblastoid cell-line, achieving 99.7% precision and 91.0% sensitivity in the expressed SNPs for the matching HumanOmni2.5 BeadChip data. Comparison of RNA-Seq eSNV candidates from 25 ER+ breast tumors from The Cancer Genome Atlas (TCGA) project with whole exome coding data showed 90.6-96.8% precision and 91.6-95.7% sensitivity. Contrasting single-cell mRNA-Seq variants with matching traditional multicellular RNA-Seq data for the MD-MB231 breast cancer cell-line delineated variant heterogeneity among the single-cells. Further, Sanger sequencing validation was performed for an ER+ breast tumor with paired normal adjacent tissue validating 29 out of 31 candidate eSNVs. The source code and user manuals of the eSNV-Detect pipeline for Sun Grid Engine and virtual machine are available at http://bioinformaticstools.mayo.edu/research/esnv-detect/.

Identification of an octamer element required for in vivo expression of the TIE1 gene in endothelial cells

TIE1, an endothelial-cell-specific tyrosine kinase receptor, is required for the survival and growth of microvascular endothelial cells during the capillary sprouting phase of vascular development. To investigate the molecular mechanisms that regulate the expression of TIE1 in the endothelium, we analysed transgenic mouse embryos carrying wild-type or mutant TIE1 promoter/LacZ constructs. Our data indicate that an upstream DNA octamer element (5'-ATGCAAAT-3') is required for the in vivo expression of TIE1 in embryonic endothelial cells. Transgenic embryos carrying the wild-type TIE1 promoter (-466 to +78 bp) fused to LacZ and spanning the octamer element demonstrate endothelial-cell-specific expression of the reporter transgene. Point mutations introduced within the octamer element result in a significant decrease of endothelial LacZ expression, suggesting that the octamer site functions as a positive regulator for TIE1 gene expression in endothelial cells. DNA-protein binding studies show that the octamer element exhibits an endothelial-cell-specific pattern of binding via interaction with endothelial-cell-restricted factor(s). Our findings suggest an important role for the octamer element in regulating the expression of the TIE1 receptor in the embryonic endothelium and suggest a common mechanism for the regulation of the angiogenic and cell-specific TIE1 and TIE2 genes during vascular development.

Endothelial cell-specific regulation of the murine endothelin-1 gene

Endothelin-1 (ET-1) plays an important role in the development, physiology and pathophysiology of the cardiovascular system in mammals. ET-1 is mainly expressed in endothelial cells thus making it an attractive model for the study of transcriptional regulation in this cell type. We have previously reported that expression of the human ET-1 gene is positively regulated by a cooperative interaction between GATA-2 and AP-1 transcription factors in cultured endothelial cells, however these factors are not sufficient to mediate cell type-specific expression. In vivo transcription studies of the murine ET-1 gene have demonstrated the presence of important cell-specific DNA elements in the 5.9 kb region upstream of the transcription initiation site. Using reporter gene transfection, site-directed mutagenesis and DNA-protein binding studies of the 5.9 kb region, we have identified a tripartite DNA element that positively regulates the expression of ET-1 specifically in cultured endothelial cells. This complex enhancer element demonstrates an endothelial cell-specific pattern of binding, suggesting that it interacts with cell-restricted regulatory factors. These findings provide important insights into the mechanisms that mediate the expression of ET-1 in the endothelium and a basis for future transgenic and cloning studies aimed at identifying the endothelial cell-specific binding site and transcription factor(s).

Octamer-dependent in vivo expression of the endothelial cell-specific TIE2 gene

The TIE2 gene, also known as TEK, encodes a tyrosine kinase receptor that is required for the normal development of the vascular system during embryogenesis. TIE2 is specifically expressed in endothelial cells; however, the transcriptional mechanisms that regulate this highly restricted pattern of expression remain unknown. Here we demonstrate that a consensus octamer element located in the 5'-flanking region of TIE2 is required for normal expression in embryonic endothelial cells. Transgenic embryos carrying a TIE2/LacZ construct spanning 2.1 kilobases of upstream regulatory sequences exhibit expression of the reporter transgene specifically in endothelial cells. Site-directed mutagenesis of a consensus octamer element located in this region results in the loss of enhancer activity and significantly impairs the endothelial expression of the reporter transgene. Consistent with the in vivo data, in vitro DNA-protein binding studies show that the consensus octamer element displays an endothelial cell-specific pattern of binding, suggesting an interaction with a protein complex consisting of Oct1 and an endothelial cell-restricted cofactor. These data identify a novel role for the octamer element as an essential regulator of TIE2 expression, define the first known transcriptional pathway that mediates the expression of a developmental endothelial cell gene, and provide insights into the transcriptional mechanisms that regulate development of the vasculature during embryogenesis.

Functional analysis of the endothelial cell-specific Tie2/Tek promoter identifies unique protein-binding elements

To investigate the molecular basis of endothelial cell-specific gene expression, we have examined the DNA sequences and the cognate DNA-binding proteins that mediate transcription of the murine tie2/tek gene. Reporter transfection experiments conformed with earlier findings in transgenic mice, indicating that the upstream promoter of Tie2/Tek is capable of activating transcription in an endothelial cell-specific fashion. These experiments have also allowed the identification of a single upstream inhibitory region (region I) and two positive regulatory regions (regions U and A) in the proximal promoter. Electrophoretic mobility-shift assays have allowed further characterization of three novel DNA-binding sequences associated with these regions and have provided preliminary characterization of the protein factors binding to these elements. Two of the elements (U and A) confer increased transcription on a heterologous promoter, with element U functioning in an endothelial-cell-selective manner. By employing embryonic endothelial-like yolk sac cells in parallel with adult-derived endothelial cells, we have identified differences in functional activity and protein binding that may reflect mechanisms for specifying developmental regulation of tie2/tek expression. Further study of the DNA and protein elements characterized in these experiments is likely to provide new insight into the molecular basis of developmental- and cell-specific gene expression in the endothelium.