Deciphering the cancer microenvironment from bulk data with EcoTyper

Abstract

In this issue of Cell, Luca, Steen et al. develop the EcoTyper software to deconvolve tumor-microenvironment interactions from high volume bulk transcriptomics data. They demonstrate its effectiveness in improving predictions for tumor progression and patient prognosis for a variety of tumor types from multiple data sources.

Poet John Donne said “no man is an island;” the same can be said of cancer cells. The dynamic interactions between cancer cells and their microenvironment determine trajectories of tumor growth, immune evasion, invasion, metastasis, and therapeutic resistance. The tumor microenvironment (TME) has multiple cellular components, including immune, vascular, neuronal, adipose, and stromal cells that can influence tumor initiation and progression (Wang et al., 2017). Recognizing the importance of these components, one of the current aims of the Human Tumor Atlas Network is to utilize single-cell and spatial technologies to build a catalog of interactions between tumor cells and the TME during critical transitional stages (Rozenblatt-Rosen et al., 2020). Association of these interactions with clinical outcomes would help identify predictive biomarkers and molecular features that are therapeutically relevant, resulting in knowledge that can improve patient-specific treatments, stratification, and prevention.

While single-cell RNA-sequencing (scRNA-seq) technologies can indeed decode tumor ecosystems (Pelka et al., 2021), we have yet to see sufficiently powered human cohort studies (~hundreds of specimens) at this resolution, which limits our ability to establish statistically robust associations between microenvironmental parameters and clinical parameters. There are important reasons why clinical studies are not yet conducted using scRNA-seq, mainly the practicality of analyzing large volumes of fresh patient specimens. Frozen banks of specimens can be processed by single-nuclei sequencing, yet the quality of the data is compromised compared to those generated from live cells isolated from fresh tissue. Establishing a logistical infrastructure to access and process fresh clinical specimens at scale remains a major challenge. Even if large volumes of fresh specimens are accessible, conducting scRNA-seq studies at such scale is cost prohibitive. Furthermore, the impact of disaggregation on cell type representation in single-cell datasets is highly variable and method dependent (Slyper et al., 2020), distorting the interpretation of the data compared to those generated without dissociation. The majority of clinical specimens are formalin fixed, paraffin embedded (FFPE) and cannot be successfully dissociated; this eliminates the usability of scRNA-seq on this routinely collected set of specimens. In contrast, bulk transcriptomics data—either via RNA-seq or microarray—can be generated at scale on frozen and FFPE specimens without dissociation artifacts. However, resolution at the cell level, and thus information about the microenvironment, is lost in bulk data.

In this issue of Cell, Luca, Steen et al. present the EcoTyper software, which solves the small sample size problem by converting bulk transcriptomics data into cell populational data complement to those generated by single cell studies (Figure 1; Luca et al., 2021). There exists a substantial quantity of accessible bulk transcriptomics data curated for prognosis and, in some cases, clinical trial responses amenable to data mining and reanalysis. Using scRNA-seq data as important landmarks, EcoTyper deconvolves bulk gene expression profiles to generate cell types and states through CIBERSORTx (Newman et al., 2019) and nonnegative matrix factorization. From large numbers of samples, groups of statistically co-occurring states are identified as “ecotypes,” defined as co-associated transcriptional programs across multicellular communities (Figure 1). Ecotypes act as strong classifiers of tumor subtypes and patient groups, and thus they can also be used for predicting tumor behavior from new data. Large-scale, high-volume bulk studies can be made amenable through EcoTyper for joint analysis with small-scale, focused single-cell studies by virtue of their shared data structures, representing a powerful tool to dissect tumors and their microenvironments.

Figure 1. The EcoTyper framework.

EcoTyper provides a means of deconvolving multi-sample and multi-platform cohorts of bulk data with insights gained from single-cell data. Cell types are further relegated as cell states and proportional representations. In this form, translationally meaningful ecotypes are composed.

Previous in silico approaches have been developed and applied to dissect cellular content and cell type-specific gene expression from bulk transcriptomic data, but the insights generated on microenvironmental interactions were limited (Newman et al., 2019; Li et al., 2020). Ojlert et al. previously used the xCell software to deconvolve immune cell populations from bulk expression data generated on 399 human non-small cell lung cancer samples, where adenocarcinoma and squamous cell carcinoma histologies with immune cell infiltration exhibited different compositions of immune cells (Öjlert et al., 2019). However, these types of analyses were limited to cell type enumeration with scant mechanistic insights coming from other data sources. Further, while many of such algorithms draw reference profiles from largescale studies, and even single-cell studies, they still largely rely on expert knowledge of marker genes to differentiate between well-defined (e.g., immune) cell types. EcoTyper takes deconvolution to the next level by further elucidating biological programs and states within cell types, which then facilitates the modeling of intercellular communication networks within the microenvironment. EcoTyper’s unique insights into deeper mechanistic details of the microenvironment beyond cell type enumeration permits a better prediction of cancer progression, outcome, and response over large numbers of patients.

In an accompanying Cancer Cell paper, Steen, Luca et al. provide an excellent case study of using EcoTyper for systematic identification of multicellular communities to predict clinical outcomes. By anchoring the analysis with a limited set of scRNA-seq data, the authors used EcoTyper to faithfully deconvolve cell types and states from bulk gene expression data generated from >1,500 diffuse large B cell lymphoma (DLBCL) specimens (Steen et al., 2021). Deconvolved data enabled the application of analyses designed for single-cell resolution data to generate tumor subtype-specific trajectories and prognosisclassifying multicellular ecosystems. Impressively, by revisiting bulk expression data from a study that previously found no clinical utility of the proteasome inhibitor bortezomib, the authors identified a specific DLBCL patient subgroup that appears to benefit from this treatment based on a unique ecotype. The focus of EcoTyper on the tumor microenvironment lends itself to emerging therapies for B cell lymphoma leveraging diverse immune cell types (Advani et al., 2018; Chao et al., 2010). More generally, given the wealth of bulk sequencing data available from previous clinical trials (Rodon et al., 2019), the same approach can likely be used to mine existing data at a finer resolution to identify responsive patient subclasses previously missed.

The strength of EcoTyper has been demonstrated by the high concordance between ecotypes and prognosis for different tumor types, including colon, ovarian, and gastric cancers, regardless of initial data sources from the Cancer Genome Atlas or PREdiction of Clinical Outcomes from Genomic Profiles. Given the increasingly widespread availability and cost-efficiency of bulk sequencing, EcoTyper has potential for broad application in clinical diagnostics by reconstructing cellular community structure at high definition and massive scale. The widereaching benefits of this framework are two-fold—in its application on data generated de novo and in the reanalysis of currently existing data. The ideal of precision medicine to provide treatments uniquely tailored to each patient’s tumors, which include diverse microenvironments, seems to be much closer to becoming a reality with the advent of cell atlasing efforts. The analogy being that instead of islands, cancers are more akin to interconnected biomes of a single continent, and EcoTyper accelerates continued cancer atlasing efforts by bridging the multitude of existing maps to new ones.