Abstract 136: Modeling of prognostication and immune profiling, based on genomic analysis in the tumor microenvironment of pancreatic adenocarcinoma via machine learning

Abstract

The extensive tumor microenvironment (TME) in pancreatic adenocarcinoma (PAD) modulates cancer progression and impact prognosis. Although gene analysis has enhanced understanding of cancer biology, few models exist to model prognosis in association with mRNA expression in the TME. Clinical outcomes data and mRNA-seq of 156 and 64 patients (pts) were obtained from TCGA and Bailey at el. [1] for testing and validation, respectively. Expressivity of 191 genes enriched in cellular and structural elements of TME and clinical data were analyzed by multivariate nonlinear regression aided by machine learning for confined optimization with model-data error minimization. Statistics including Kaplan-Meier (KM), Cox Hazard (CH), and correlation analysis was used. Most pts (85.89% and 85.94%, respectively) were in stage II, and pts in stage I/III/IV were excluded. Prognostication was modeled with higher risk score (RS) representing worse prognosis: RS = -7.6526 x (Age-5.5679) + 0.0813 x (P/G0.3677) + 0.7069, where P/G is a ratio of genes associated with poor to good prognosis (Table 1). Based on RS, pts were clustered into 2 groups (high and low RS) with 2 KM curves showing p<0.0001 and p=0.014 in test and validation sets. Immune profiling of high and low RS groups in both test and validation sets shows that in low RS group, genes related to both immune activation (IA) and inhibition (II) (Table 2) are highly co-expressed, implying that co-expression of IA and II contributes to PAD’s poor prognosis even in pts with immune system activation. In high RS group, genes related to cancer stem cells (CD44 and EPCAM) significantly contributed to poor prognosis. Machine learning-assisted modeling of RS and gene analysis suggest that IA genes are suppressed by co-expression of high degree of II, contributing to poor prognosis in PAD. RS enables prognostication of pts encountered in the clinic when genomic profiles are provided. [1] Nature 531, 47-52 (2016). Table 1genes associated with good and poor prognosis out of 191 genes (identification via KM and CH with p<0.05)Good prognosisFCRL3, LILRA4, IL3RA, IL10, CCL22, DOK3, CXCR4, PDGFA, ICOSLG, TNFRSF4Poor prognosisTNFSF10, CD44 Table 2gene groups of immune activation (IA) and immune inhibition (II)IA gene groupscytotoxic T, B, NK, T-helper 1 cells, IFN, cytolytic activity, T cell co-stimulation, and antigen presentationII gene groupsregulatory T cells, desmoplasia, immunosuppressive chemokines, immune checkpoints, angiogenesis, cancer stem cells, epithelial-mesenchymal transition, and neutrophils Note: This abstract was not presented at the meeting. Citation Format: Sunyoung S. Lee, Seok Joon Kwon, Ahmed Elkhanany, Andrew Baird, Seongwon Lee, Jillian Dolan, Stuart Baird, Shinyoung Park, Renuka Iyer. Modeling of prognostication and immune profiling, based on genomic analysis in the tumor microenvironment of pancreatic adenocarcinoma via machine learning [abstract]. In: Proceedings of the American Association for Cancer Research Annual Meeting 2019; 2019 Mar 29-Apr 3; Atlanta, GA. Philadelphia (PA): AACR; Cancer Res 2019;79(13 Suppl):Abstract nr 136.