Modeling RAS Phenotype in Colorectal Cancer Uncovers Novel Molecular Traits of RAS Dependency and Improves Prediction of Response to Targeted Agents in Patients

Abstract

Purpose KRAS wild-type status is an imperfect predictor of sensitivity to anti-EGFR monoclonal antibodies in colorectal cancer (CRC), motivating efforts to identify novel molecular aberrations driving RAS. This study aimed to build a quantitative readout of RAS pathway activity to: (1) uncover molecular surrogates of RAS activity specific to CRC; (2) improve the prediction of cetuximab response in patients; (3) suggest new treatment strategies. Methods A model of RAS pathway activity was trained in a large CRC dataset and validated in three independent CRC patient datasets. Novel molecular traits were inferred from the TCGA CRC data. The ability of the RAS model to predict resistance to cetuximab was tested in mouse xenografts and three independent patient cohorts. Drug sensitivity correlations between our model and large cell line compendiums were performed. Results The performance of the RAS model was remarkably robust across 3 validation datasets. (1) Our model confirmed the heterogeneity of the RAS phenotype in KRAS wild-type patients, and suggests novel molecular traits driving its phenotype (e.g. MED12 loss, GBXW7 mutation, MAP2K4 mutation). (2) It improved the prediction of response and progression free survival (HR=2.0; p<.01) to cetuximab compared to KRAS mutation (xenograft and patient cohorts). (3) Our model consistently predicted sensitivity to MEK inhibitors (p<.01) in 2 cell panel screens. Conclusions Modeling the RAS phenotype in CRC allows for the robust interrogation of RAS pathway activity across cell lines, xenografts, and patient cohorts. It demonstrates clinical utility in predicting response to anti-EGFR agents and MEK inhibitors.

A Composite Gene Expression Signature Optimizes Prediction of Colorectal Cancer Metastasis and Outcome

“…Thus, ΔPC1.EMT captures both epithelial and CSC features, which are supported by a recent report demonstrating that in breast cancer, CDH1 and CD24 were highly enriched in the epithelial CSCs ( ALDH1 -positive), while their expression was down-regulated in the mesenchymal CSCs ( CD44 + CD24 -)(33). ERBB3 , a member of the EGFR family (34), was also identified as one of the genes whose contribution was increased in ΔPC1.EMT (Figure 3C). In agreement with thi s , we observed that ΔPC1.EMT, but not EMT, was associated with activation of the RAS/MAPK pathway, evidenced by its positive correlation with various RAS signature scores (Supplementary Table S10).…”

“…In agreement with thi s , we observed that ΔPC1.EMT, but not EMT, was associated with activation of the RAS/MAPK pathway, evidenced by its positive correlation with various RAS signature scores (Supplementary Table S10). Thus, we speculated that ΔPC1.EMT-associated poor prognosis might, in part, result from RAS/MAPK activation-mediated drug resistance (34) in epithelial-like CRC.…”

Heterogeneity in the colorectal primary tumor and the synchronous resected liver metastases prior to and after treatment with an anti-EGFR monoclonal antibody

“…Changes in PIK3CA mutation status during treatment with anti-EGFR drugs have previously been described, although a clear correlation with resistance to anti-EGFR mAbs was not identified (17). Mutations in FBXW7 have been recently reported as potential markers of resistance to anti-EGFR mAbs (18). However, the disappearance of the FBXW7 mutation from cells following response to cetuximab-based therapy argues against this hypothesis.…”

A machine learning classifier trained on cancer transcriptomes detects NF1 inactivation signal in glioblastoma

“…Machine learning has been applied to gene expression in a variety of studies with various goals [37][38][39][40][41]. In a similar study, Guinney et al trained a classifier to model RAS activity in colorectal cancer and demonstrated its clinical utility by predicting response to MEK inhibitors and anti-EGFR based treatments [18]. With a wealth of signal embedded in gene expression and a rapidly growing library of datasets, the performance of machine learning models is likely to rapidly improve.…”

“…We chose a penalized regression model because it is simple to train and has easily interpretable outputs including importance scores for each gene (feature weights) associated with the downstream consequences of NF1 loss of function and a probability for each sample that NF1 is lost. An elastic net logistic regression model has also been successfully implemented in similar studies [18][19][20].…”