Causal modeling of TCGA, NSCLC, and HNSCC data to identify network drivers of tumor immune subtypes.

Abstract

68 Background: Immune checkpoint inhibitors have achieved unprecedented success in several cancer types, yet only a subset of patients derives clinical benefit. Better understanding of tumor-immune interactions is imperative to improve clinical outcomes. Bayesian causal machine learning was applied on real world data to elucidate the molecular drivers of immune subtypes of tumors. Methods: Using a Reverse Engineering Forward Simulation (REFS) platform, ensembles of causal models were built on genomics, transcriptomics, and clinical data from 681 non-small cell lung carcinoma (NSCLC), 328 lung adeno (LUAD) and 353 lung squamous cell (LUSC), and 413 head and neck squamous cell carcinoma (HNSCC) patients from TCGA. The outcomes modeled were six tumor immune subtypes (Thorsson et al., 2018): wound-healing, IFNγ dominant, inflammatory, lymphocyte-depleted, immunologically quiet, and TGFβ dominant. Causal drivers of immune subtypes were identified from average causal effect (ACE) of the variables, as computed from the counterfactual simulations. ACE was defined as median of posterior distribution of odds ratio (1 vs 0 for discrete; 95th vs. 5th %ile for continuous variables). Results: The models showed impressive k-fold cross validation predictive performance (AUC ~ 0.8-0.9) for the most prevalent immune subtypes in TCGA: wound-healing and IFNg dominant in both LUAD and HNSCC, as well as inflammatory in LUAD. The potential causal drivers of the tumor immune subtypes and their ACE are listed in Table. The findings suggest that macrophage activation and polarization, which is driven in part by metabolic reprogramming, is a primary driver of tumor immune subtypes. Conclusions: Bayesian causal modeling revealed literature-supported hypotheses regarding predictors of response (CXCL13) and resistance (STK11 mutation, mTOR pathway) to PD-(L)1 blockade therapy. The additional target pathways such as AKT1/mTOR may be actionable for altering immunogenicity of tumors. [Table: see text]