Identifying pan-cancer transcriptomic determinants of perineural and lymphovascular invasion using machine learning.

Abstract

3621 Background: Tumor invasion of nerves, blood vessels, and lymphatics are a primary means of local recurrence and escape from the local microenvironment, resulting in metastases and poor clinical outcomes. However, the genetic drivers that are most pertinent to these malignant processes are not well understood, and few therapeutics successfully target perineural invasion (PNI) and lympho-vascular invasion (LVI). Identifying genetic drivers and biomarkers can be valuable for therapeutic targeting and prognostication. Methods: We analyzed surgical pathology reports and bulk RNA-seq data of 1,624 patients across 12 cancer types from The Cancer Genome Atlas (TCGA). Differential gene expression analysis between patients with and without PNI/LVI was performed using DEseq2 in Python while adjusting for age, sex, race, and cancer type. Genes with an adjusted p-value < 0.001 were then used to derive parsimonious signatures using random forest classifier and recursive feature selection algorithms. Results: To assess whether these invasive histological phenotypes have clinical ramifications, we examined outcomes data and found that patients with PNI or LVI have reduced overall (OS) and disease-free survival (DFS) ( p < 0.05) relative to those without. In addition, patients with both PNI and LVI have the lowest DFS from our pan-cancer analysis, suggesting that each may have non-redundant contributions to poor outcomes. From the differential gene expression analysis, we identified a set of 621 and 606 genes that were highly associated with PNI and LVI, respectively (padj < 0.001). Many of these genes such as TEKT5 (padj = 3.18 x 10−64), which is canonically associated with ciliary and flagellar microtubules, and SCRIB (padj = 1.60 x 10−21), which helps establish apico-basal cell polarity, have not been described previously in relevance to PNI and LVI, and warrant further scientific and clinical investigation. These genes were ultimately condensed into a signature that optimizes for both model simplicity and goodness of fit with up to 90% accuracy as determined by trials on both a logistic regression and neural network model. Conclusions: We concluded from a pan-cancer analysis that PNI and LVI are associated with poor outcomes, and we were able to robustly identify sets of genes that characterize each invasive mechanism for further functional investigation.