SA105 - Predicting Response to Chemotherapy in Colorectal Cancer Using a Clinical-Genomic Machine Learning Model

Purpose/Background: For first-line chemotherapy regimens in colorectal cancer, no established biomarkers exist to predict response to treatment. Next-generation sequencing has enabled large-scale characterization of colorectal tumors using gene mutation panels. We sought to use machine learning to develop a combined clinical-genomic risk model, which could predict patients’ response to chemotherapy and be used as a prognostic clinical decision support tool.

Methods/Interventions: We used data from the American Association for Cancer Research Project GENIE (Genomics Evidence Neoplasia Information Exchange). We included all patients with colorectal cancer who received FOLFOX. Using the least absolute shrinkage and selection operator (LASSO) model, we identified a set of gene mutations, copy-number alterations, and gene fusions which were potentially predictive of 3-year overall survival. We applied a variety of machine learning modeling approaches, including elastic-net, random forest, gradient boosting, and deep neural networks to this genomic dataset. Each model was assessed using 5-fold cross-validation and area under the receiver operating characteristic curve (AUROC). The highest performing model was used to create a genomic risk score (GRS), which was evaluated as an independent prognostic risk factor along with clinical variables using multivariable logistic regression. Finally, a combined clinical-genomic risk model was developed and assessed using AUROC. Kaplan-Meier curves were created for patients with various quartiles of predicted risk. Shapley additive explanation (SHAP) values were used to evaluate the relative importance of genomic variables.

Results/Outcomes: The dataset included 1111 patients from three institutions. 725 variables were considered in total, including 21 clinical variables, 224 gene mutations, 224 copy number alterations, and 256 gene fusions. Of the modeling techniques tested, random forest showed the best performance and was used to generate a GRS for each patient. The GRS showed strong association with overall survival with an odds ratio of 17.0 (95% CI 14.9 – 19.1), p-value < 0.001. A combined clinical-genomic model showed an AUROC of 0.74 (95% CI 0.69 – 0.80) for predicting 3-year overall survival. A Kaplan-Meier curve separating patients into quartiles of predicted risk showed clear separation between groups (Figure 1A). Analysis of variable importance showed that copy number alteration of the genes RARA, MYC, and ASXL1 and mutations of SDHA, EP300, and CDKN2A had the highest predictive importance for overall survival (Figure 1B).

Conclusion/Discussion: This study developed and validated a genomic risk score to predict overall survival after initiation of first-line chemotherapy for patients with colorectal cancer, using a large, multi-center dataset. With further optimization and validation, this model has the potential to be used as a prognostic decision support tool. This may be particularly useful for predicting response to total neoadjuvant therapy for patients with locally advanced rectal cancer.