(P 424) CROSS-COHORT INTEGRATION AS A STRATEGY TO CLASSIFY IMMUNE CHECKPOINT BLOCKADE RESPONSE IN SARCOMA

Objective: Soft tissue sarcomas (STSs) are a diverse class of aggressive malignancies. Although the efficacy of standard-of-care treatment is limited, recent clinical studies demonstrate that some patients respond to immune checkpoint blockade (ICB). Despite this advance, molecular determinants of response are not yet well characterized. The definition of sarcoma immune classes (SICs) comprising “hot” versus “cold” immune microenvironments captures one aspect of the response, with hot tumors more likely to benefit from ICB. While this classification achieves some specificity, improvement is needed. The rarity of STS, a high diversity of tumor histologies, and heterogeneity among patients are all obstacles toward better classification. We aim to overcome these through the integration of both molecular and clinical response data across multiple available STS cohorts. Our broad goal is to enable the prospective classification of STS patients into those most similar to ICB-responders versus non-responders.

Methods: We sequenced a pan-sarcoma transcriptome dataset (n=98 prospective cases), and further collected external cohorts (n=552) including those with known ICB response and SIC annotations. Data deconvolution applied to each dataset identified gene expression programs representing tumor cell states, immune composition, and normal tissue components. Programs were integrated across all cohorts into a network (mosaicMPI), generating communities of programs corresponding to biological themes; these were further annotated based on pathway enrichment (g:Profiler2). Within each community, annotations (labels) from clinical and molecular data in each external cohort (e.g., ICB response, sex, molecular tumor subtypes, SIC, etc.) were statistically transferred to our in-house cohort. Using an orthogonal approach (immunedeconv), we estimated immune cell composition at the sample level. This enabled direct comparison with (and validation of) the estimated immune composition based on community usage and label transfer results in each patient.

Results: mosaicMPI integration identified 19 program communities representing immune responses, extracellular matrix organization, cell cycle, tumor metabolic processes, normal tissues (muscle), and other components of the STS biological landscape. Our community-based label-transfer of SIC and ICB response annotations between cohorts revealed that some patients had a high concordance of hot (SIC-E) tumors with ICB response, while others did not. This indicates our integration approach can be used to combine multiple sources of relevant ICB response and immune state variables at the single patient level in a cohort without those annotations. Patients enriched in ICB response status had high usage of programs annotated for T/B cell activation and interferon production, whereas those enriched in non-responder status had usage of programs annotated for cell migration. These label-transfer results were strongly supported by immune composition data – immune-high/ICB-responders showed statistically higher abundance of CD8 and CD4 T-cells, dendritic cells, monocytes, and macrophages compared to immune-low/non-responders.

Conclusion: Our strategy can integrate multiple RNA-seq cohorts, each distinguished by treatment status, clinical response, and variable composition of tumor subtypes. This allows the statistical transfer of molecular and clinical labels among samples within a cross-cohort community, such that relevant annotations (i.e. SIC, ICB response) can be projected onto patients without these labels. Our orthogonal validation suggests this transfer is accurate. We noted that SIC and ICB response statuses did not always coincide in individual patients, indicating the influence of other important variables not quantified at the RNA level, for instance, neoantigen quality. This underscores the significance of incorporating multiple annotations in understanding ICB responses. Future work will extend the integration to incorporate these other variables.