Artificial intelligence- and machine learning-based approaches to identify biomarkers of patient response to immunotherapy for non-small cell lung cancer
77. Predicting the tumor microenvironment molecular composition to assess immunotherapy efficacy in non-small cell lung cancer from digital histopathology images
Alex Chen (National Cancer Institute, Bethesda, MD, USA) described a novel artificial intelligence- (AI-) based approach to characterize the composition of the tumor microenvironment (TME) using histopathology images of samples from patients with non-small cell lung cancer (NSCLC) and how to use this characterization as a predictive biomarker for response to immunotherapy. The AI model was trained with Hematoxylin and Eosin (H&E)-stained pathology slides of patient samples and bulk RNA-sequencing signatures transcriptomics data from the TCGA-NSCLC cohort of 865 patients, and the model was validated with samples form a 333-patient cohort of CPTAC-NSCLC. The test cohort consisted of samples from 652 patients with NSCLC; 290 were treated with immune checkpoint inhibitors (ICIs), and 362 were not treated with ICIs. Data from the validation dataset indicated the model accurately predicts the expression of 30 TME molecular signatures from H&E images, with an average Pearson correlation of 0.50. TME signatures were used to classify the TME into two clusters, Immune-Desert and Immune-Inflamed. Overall survival was more favorable among patients with an Immune Inflamed TME compared to an Immune Desert TME (p=0.022), and PD-L1 expression of 50% or higher was also associated with improved survival (p=0.13). Survival benefits were more closely associated with an Immune Inflamed TME and high PD-L1 expression among patients receiving immunotherapy in the first-line setting (p=0.0012 and p=0.0059, respectively), suggesting that the AI-based classification of the TME adds benefit in addition to PD-L1 expression to predict potential responders to ICIs. Similarly, Immune-Inflamed and high-PD-L1 were significantly associated with progression free survival benefits only among patients receiving ICI in the first-line setting (p=0.0037 and p=0.003, respectively). This observation may be due to patient samples being obtained prior to first-line treatment, thus samples may not be representative of the TME during second-line or later treatments. A supervised learning/machine learning model incorporating the interactions between the different TME signatures was used to predict patient response to immunotherapy, and this model was more accurate in predicting OS and PFS for patients receiving ICI as first-line treatment. These findings suggest that AI- and machine learning-based approaches to analyze patient samples could potentially accurately characterize the complex TME of NSCLC, supplement other currently used biomarkers of patient response (e.g. PD-L1 expression), and potentially be used to predict response to immunotherapy in the future, ultimately to promote more personalized medicine. These results with NSCLC need to be validated with additional studies of other clinical cohorts; and future directions include expanding these models to other cancer types.
Understanding the role of tertiary lymphoid structures in the tumor microenvironment of pancreatic cancer
209. Spatial multi-omics reveal humoral immunity niches associated with tertiary lymphoid structures in pancreatic cancer pathologic responders to neoadjuvant immunotherapy.
Dimitrios Sidiropoulos (Sidney Kimmel Cancer Center, Johns Hopkins University, Baltimore, MD, USA) reported a study characterizing tertiary lymphoid structures (TLS) in patients with pancreatic ductal adenocarcinoma (PDAC) who responded to neoadjuvant immunotherapy. TLS are associated with higher survival in most cancers, and the rare survivors of PDAC have high TLS density. Past studies have used the cancer vaccine GVAX to generate TLSs, but this was not associated with significant survival benefit, likely due to the immunosuppressive tumor microenvironment (TME) associated with PDAC. In order to understand the role of TLS in PDAC and its spatial relationships within the TME, tumors from 26 patients with PDAC who received neoadjuvant treatment with GVAX, nivolumab, and the CD137 agonist urelumab were profiled. A machine learning-enabled H&E classifier identified TLSs and other components of the TME. TLSs were embedded in multiple PDAC intratumoral domains, including within the tumor stroma, within fat, and within regions containing a variety of tumor elements, including endocrine and exocrine cells. A 40-marker TLS panel to evaluate TLS maturity was developed and used to characterize and compare the TLS of pathologic responders to immunotherapy vs. non-responders to immunotherapy. TLS of non-responders often had aggregates with diffuse T and B cell mixtures, while TLS of responders featured clustered B cells expressing germinal centers and B cell trafficking markers. Spatial transcriptomics mapped TLS transcriptional programs within the broader TME, and an unsupervised machine learning algorithm was used to mine continuous gene expression programs throughout the TME. A transcriptional pattern enriched in immune markers correlated with both TLSs and with germinal centers of adjacent lymph nodes; and it was used as a TLS maturation score. This TLS maturation score correlated with positive survival outcomes across a variety of cancers, including PDAC. TLS that bordered non-immune cells like stromal cells and cancer cells were associated with gene expression programs associated with humoral immunity, immunoglobulin expression, and plasma cell markers. B cell receptor sequence alignment indicated BCRs were polyclonal and often expanded in TLS and nearby stromal tissue. TLS of responders to neoadjuvant therapy expressed higher levels of IgG, while TLS of non-responders expressed higher levels of IgA. TLS of responders also exhibited increased density of collagen hybridizing peptide (CHP), which may facilitate extracellular matrix remodeling and the trafficking of lymphocytes and other immune cells throughout the tumor. These findings further support the role of TLS in promoting favorable clinical outcomes in PDAC and other cancers and provide new insights into how TLS promote anti-tumor immunity in the TME during immunotherapy treatment.
Single-cell and spatial analyses of cellular interactions associated with response to immune checkpoint blockade for head and neck cancer
543. Integrated single-cell and spatial analysis of immune cell interactions in response to immune checkpoint blockade
Athena Golfinos-Owens (University of Wisconsin-Madison, Madison ,WI, USA) presented an integrated analysis of single cell and spatial transcriptomic data to identify specific myeloid cell to T cell interactions during immune checkpoint blockade (ICB) response. The goal of this study was to characterize myeloid cell interactions with T cells in the tumor microenvironment (TME) and how the heterogeneity of these interactions contribute to response or non-response to immune checkpoint blockade. Datasets of treatment naïve patients with head and neck cancer (HNC), ICB-treated patients with HNC, and mouse models of HNC were included in the analyses. Single-cell RNA sequencing (scRNA-seq) data sets identified multiple myeloid subsets in ICB responders and non-responders. An in silico cell-cell communication inference method identified ligand-receptor/cell-cell interactions that were enriched in patients who responded to ICB. CXCL16-CXCR6 was the most enriched interaction in responders pre-ICB treatment and CXCL9/10-CXCR3 was the most enriched interaction in responders post-ICB treatment. CXCL9-CXCR3 was the only interaction enriched in both pre-treatment and post-treatment responders. Ligand-receptor interactions were further analyzed to identify specific interactions that would be spatially constrained through cell-cell interactions and those that occur in CD45+ (immune positive) cellular neighborhoods that contain immune cells. CXCL9-CXCR3 and CXCL16-CXCR6 were identified as two ligand-receptor interactions that were enriched in these cellular niches in ICB responders. Tumors from ICB responders were also enriched in CXCL9+ CXCR3+ immune positive spots compared to tumors from ICB non-responders. Single cell image-based spatial profiling was performed to further analyze ligand-receptor interactions in cellular neighborhoods where all immune cells were present. Within these immune-rich cellular neighborhoods, ICB responders exhibited increased levels of interaction between three different myeloid cell populations (dendritic cells, macrophages, and monocytes) to CD4 T cells, and CXCL9/10-CXCR3 and CXCL16-CXCR6 ligand-interactions were the most prevalent chemokine interactions contributing to these cell-cell interactions. Conversely, in non-responders, no interactions between myeloid cells and CD4 T cells were observed. These findings are further supported by a previous study indicating that CXCL9 expression is a stronger predictor of response to ICB than PD-L1. Results from this study have identified cellular neighborhoods, myeloid cell-T cell interactions, and ligand-receptor interactions, specifically CXCL9/10-CXCR3 and CXCL16-CXCR6, as potential biomarkers and/or targets for response to ICB for HNC. Additional work is needed to validate these interactions and elucidate the mechanisms that lead to improved response to ICB for these patients.