(P 397) A MESENCHYMAL TISSUE LANDSCAPE TO STUDY OSTEOSARCOMA TUMOR HETEROGENEITY

Objective: Osteosarcomas (OS) are a heterogenous class of bone tumors that can occur both in children and adults. OS tumors exhibit complex karyotypes and diverse genetic alterations, although few common pathognomonic drivers have been identified. The diversity of genetic drivers in OS poses a major challenge for treatment strategy because each tumor may possess unique therapeutic vulnerabilities.
Clinically, OS tumors are typically classified based on histological assessment including apparent predominant cell types (osteoblastic, chondroblastic) and relative differentiation status (low-grade, high-grade). However, a common feature of OS is that tumor cells can exhibit a range of differentiation states that may include multiple lineages (osteogenic, chondrogenic, adipogenic, etc.) derived from mesenchymal stem cells (MSCs). Instead of mutation-based drivers, differentiation states may depend on epigenetic and transcriptional changes that relate to normal cell differentiation resembling osteogenesis or other mesenchymal differentiation. Importantly, OS might arise from disrupted normal differentiation of MSCs or, alternatively, from a dedifferentiation of mature cells.
In order to sufficiently investigate how OS relate to normal mesenchymal differentiation, we first characterized a Mesenchymal Tissue Landscape (MTL) based on single-cell gene expression dynamics across time courses within multiple differentiation lineages (osteogenic, adipogenic, chondrogenic). Then, we utilized gene expression signatures learned from MTL to quantitatively assess differentiation states and lineage heterogeneity with a cohort of OS patient-derived xenografts and OS tumors. Herein, we describe the results of these analyses and highlight certain novel findings relating to tumor heterogeneity in OS.

Methods: First, we collected single-cell RNA-sequencing (scRNA-seq) data from three patient-derived xenograft (PDX) models of OS. We additionally incorporated a publicly available scRNA-seq dataset of 11 human OS tumors. Each PDX and tumor was classified based on a pathologist assessment as osteoblastic or chondroblastic in origin. We applied gene expression analysis to determine cell clusters and marker genes of each PDX, including known marker genes of osteogenesis and chondrogenesis. Pathway analysis was performed to assess molecular pathways expressed in each cell cluster.
Next, we developed a tissue-engineered model of mesenchymal cell differentiation using biochemical and biophysical cues (including polyacrylamide hydrogel stiffness) to direct differentiation into two lineages: high stiffness for osteogenic, low stiffness for adipogenic. Cells were cultured and collected at multiple time points until terminal differentiation, and for each time point scRNA-seq was performed to measure gene expression. Additionally, a publicly available dataset containing a similar single-cell time course of chondrogenic differentiation was integrated along with our experimentally generated lineages, resulting in a MTL atlas containing three mesenchymal lineages: osteogenic, adipogenic, chondrogenic. With the MTL gene expression data, clustering and differential expression analysis were used to identify lineage-specific marker genes of early and late differentiation. Then, archetype analysis based on nonnegative matrix factorization was applied to define temporal gene expression signatures corresponding to lineage-specific mesenchymal differentiation states.
The differentiation state signatures learned from MTL were used to directly estimate the composition of differentiation states within each OS PDX and OS tumor as a means of quantitatively assessing intratumor heterogeneity. Further, we applied these signatures to a dataset of bulk RNA-seq data in pediatric OS (TARGET-OS study) and performed survival analysis (multiple Cox regression) with respect to the estimated archetype compositions.

Results: We observed 15 distinct subpopulations of OS PDX cells, some of which were PDX-specific and others which were shared across multiple PDX suggesting some conserved states among different OS PDXs. We observed significant heterogeneity in osteoblastic markers (RUNX2, CDH11, SPP1) expression across the PDXs. All PDXs exhibited a subpopulation of proliferative cells expressing cell cycle markers MKI67 and TOP2A. In one PDX classified as chondroblastic (SA98), we identified expression of several fibroblastic and chondroblastic marker genes indicative of a mixed phenotype.
In our experimentally generated mesenchymal differentiation atlas (MTL), we observed a clear trifurcation into three distinct lineages: osteogenic, adipogenic, chondrogenic. Early time points comprised largely undifferentiated and proliferative cells, marked by cell cycle markers and MSC-specific markers. YAP/TAZ regulated genes were enriched in cluster MSC-H versus MSC-L. Cells at later time points were committed to one of the three lineages, with corresponding markers of early and late differentiation specific to each lineage.
Archetype analysis determined 12 distinct gene signatures that corresponded to lineage-specific and time-ordered expression dynamics observed in the MTL. We used these signatures to estimate the proportion of each differentiation state within OS PDX and OS tumor gene expression. Importantly, the quantification of osteogenic and chondrogenic fractions of each tumor agreed closely with pathologist labels of predominant cell of origin, suggesting that gene expression analysis with reference to normal mesenchymal differentiation can quantify tumor heterogeneity in agreement with histological assessment. Finally, estimated compositions in bulk OS gene expression were correlated with survival, indicating a more well-differentiated signature was associated with improved outcome.

Conclusion: In our study, we apply a novel paradigm to study OS tumor heterogeneity by constructing a multi-lineage single-cell mesenchymal differentiation landscape. With this framework, multiple differentiation states corresponding to each lineage were able to be resolved corresponding to gene expression dynamics in normal mesenchymal differentiation. Furthermore, novel gene signatures were defined corresponding to a decomposition of temporal- and lineage-specific differentiation states.
Importantly, the signatures learned from the mesenchymal landscape were shown to be effective in estimating the cell lineages and differentiation states within OS PDX and tumor data. Our findings indicate that OS tumors exhibit a high degree of intratumor heterogeneity as well as highly variable compositions of cell subpopulations among different OS tumors. Critically, the compositions estimated from gene expression agreed closely with pathologist labels of predominant cell type, validating the accuracy of the signatures and further allowing quantitative estimation of OS heterogeneity. Compared to a single histological label, the proposed method offers greater information by quantifying multiple subpopulations and can reveal mixtures of multiple cell types. Lastly, we demonstrated the differentiation signatures could be applied to estimate compositions of bulk tumor data, wherein a well-differentiated signature was correlated with improved survival outcome. This approach has potential to inform treatment strategy, as dominant cell populations might influence outcomes or require concurrent treatment strategies.