SA3 - Toward Improved Therapy for Oral Cancer: Dissecting the Tumor Microenvironment at the Single-Cell Level With Machine Learning Predicts Novel Therapeutic Targets in OSCC

Abstract: Of “The Patients We Serve”, oral squamous cell carcinoma (OSCC) is one of the most devastating and difficult diseases to treat. It presents a significant therapeutic challenge: treating both the tumor itself and the microenvironment in which it develops. Clinically, this manifests as high post-surgical recurrence rates – 28% in early stages and up to 60% in late stage. It is likely that the tumor-promoting microenvironment that remains, even after histologically negative margins, contributes to this high risk of recurrence in OSCC.¹ Recent insights propose that cancer emerges only within a permissive microenvironment, where altered cell-cell communications and disrupted gene regulatory networks pave the way for tumor progression.² For this reason, we set out to analyze the transcriptional activity of the tumor microenvironment of OSCC leveraging computational biology.

To explore these dynamics, we analyzed single-cell RNA sequencing data from 37 human oral epithelial tissue samples – including HPV (-), non-tumoral normal tissue (NL) and primary cancer (CA). Data was downloaded from GEO, accession GSE181919. Standard single cell quality control was conducted, and transcriptomes were normalized/scaled via the SCTransform protocol from Seurat (V5.2) in R. Samples were integrated with Harmony (V4.1) and cells clustered and displayed with a UMAP projection via Seurat. FindAllMarkers Wilcoxon test was used to identify marker genes (adj P < .05) for the 30 identified clusters/cell-types, across T cells, B cells, macrophages, epithelial cells, and fibroblasts. Our novel machine learning platform, TrokaChatML, demonstrated exceptional performance (R² = 0.996, RMSE = 0.165, STDEV = 2.495, Range = 13.18) in predicting perturbed cell-cell communications. Communications were filtered for an adjusted P < .05 via the Wilcoxon test. TrokaChatML tensor decomposition discovered latent patterns in the communication network, providing coordinated “factors” of communication upregulated in CA across cell-types and ligand-receptor pairs, with a rank selected via Bayesian Information Criterion. Finally, TrokaChatML simulated signal propagation of perturbed communication patterns through a multilayered biological knowledge graph using a random walk with restart algorithm. This predicts which intracellular genes and pathways are most perturbed in the microenvironment due to the changes in cell-cell communication, as well as predict which drugs could be used to target these changes.

TrokaChatML decomposed the identified perturbed interactions into 17 distinct signaling patterns, with 12 patterns specifically upregulated in cancerous tissues. One notable finding was enhanced RARRES2/POSTN signaling in cancer-associated fibroblasts – pathways linked to cellular senescence and chronic inflammation. This altered signaling likely contributes to the creation of a pro-tumorigenic stroma, setting the stage for both OSCC development and recurrence. Furthermore, our analysis uncovered a significant expansion of regulatory T cells (Tregs) in the malignant microenvironment. Another identified pattern included heightened CXCL16/CD86 signaling from macrophages that interact with CXCR6/CTLA4 on Tregs, thereby recruiting more Tregs to the microenvironment while suppressing effector T cell activity and facilitating immune evasion. These mechanistic insights led TrokaChatML to predict actionable targets. In particular, anti-CTLA4 therapies such as Ipilimumab and Erfronrilimab emerged as promising candidates for modulating the immune landscape and reducing the risk of recurrence.

Primary OSCC treatment has shown minimal improvement in 5-year survival. Further study of disease mechanisms and potential interventions is essential to improve outcomes. By dissecting the tumor microenvironment at single-cell resolution and applying advanced computational modeling, this study provides a robust framework for personalized, targeted oral cancer therapy. Our findings underscore the urgent need to address aberrant cell signaling that sustains a malignancy-supportive stroma. In working towards personalized and preventative medicine, machine-learning tools like TrokaChatML are integral to translating complex cellular data into actionable clinical strategies.

1. Blatt, S. et al. J Clin Med 11, 7061 (2022).
2. Huang, S., Soto, A. M. & Sonnenschein, C. PLoS Biol 23, e3003052 (2025).