Predictive Polygenic Score for Outcome after First-Line Oxaliplatin-Based Chemotherapy in Colorectal Cancer Patients Using Supervised Principal Component Analysis

Oxaliplatin is a platinum drug often given in combination with other anticancer drugs to treat colorectal cancer. None of the previously reported candidate genetic polymorphism with oxaliplatin treatment outcomes has been robustly validated (1). Applying a machine-learning approach on 592-gene-sequencing data, a recent study derived a 67-gene signature predictive of oxaliplatin-based treatment efficacy in metastatic colorectal cancer (mCRC; ref. 2). However, no germline genetic variants are currently available as robust predictive markers, i.e., differential association with outcome according to the type of chemotherapy (e.g., with oxaliplatin versus without oxaliplatin) for clinical practice. To efficiently extract the information value of multiple genetic variants showing significant association, we used a supervised principal components approach (3) to construct polygenic hazard scores (PHS) for differential outcomes in colorectal cancer patients receiving oxaliplatin-containing chemotherapy versus others.

Details of all studies and inclusion criteria are found in Supplementary Methods. Two phase III trials served as discovery cohorts, including 3,098 resected stage III colon cancer (rCC) patients of N0147, 96% of patients received oxaliplatin-containing regime (4), and 506 mCRC patients of N9741, 79% of patients received oxaliplatin-containing regime (ref. 5; Supplementary Fig. S1). Both studies were led by the North Central Cancer Treatment Group (NCCTG). NCCTG is now part of the Alliance for Clinical Trials in Oncology. All participants provided written informed consent and the studies were approved by the institutional review board of all participating centers in accordance with Declaration of Helsinki.

All analyses were conducted separately for rCC and mCRC. First, genome-wide association analysis for overall survival (OS) was conducted for each SNP. Cox models included an interaction term of SNP with treatment type (predictive effect). Models were adjusted for age, sex, grade (only for NCCTG N0147), KRAS mutation, and the first three principal components (PC) of all tested SNPs. Manhattan plots were used to illustrate the P values for 4,695,046 imputed/genotyped SNP interactions with type of treatment across chromosomes (Supplementary Fig. S2). Predictive SNPs with genome-wide significance (P < 5 × 10^–8) were used for principal component analyses (PCA) to derive PHSs separately for rCC and mCRC. The PHSs (composed of selected PCs), treatment type, and interaction terms of each PHS with treatment type were added to clinical models [including age, sex, grade (only for patients rCC), and KRAS mutation] to predict outcome.

We attempted validation of the prediction models based on the discovery cohort using patients of the population-based DACHS study (6, 7), separately for 549 patients with rCC and 437 patients with mCRC (Supplementary Fig. S1). The study was approved by ethics committees of the Medical Faculty of the University of Heidelberg and the State Medical Boards of Baden-Wuerttemberg and Rhineland-Palatinate. The prediction error curve for OS comparing the clinical models and prediction models was evaluated using the loss function approach (8) in the R package “pec”.

The data from NCCTG N0147 and NCCTG N9741 analyzed in this study are available from dbGaP (phs001290.v1.p1). The data from DACHS study are available on reasonable application from the corresponding author.

Selected characteristics of patients from participating studies according to the type of chemotherapy are presented in Supplementary Table S1. In the discovery cohorts, GWAS yielded 15 significant SNPs for rCC and 13 SNPs for mCRC. PCA of these predictive SNPs revealed a linear combination of the first two PCs as sufficient for an appropriate presentation of their information value, with corresponding weights given in Table 1. The PHSs, that is, first two PCs of the selected SNPs, were used to construct interaction terms with treatment type. In the discovery cohorts, PHS showed significant interaction with treatment, as expected, because the SNPs were selected for interactions using the same dataset.

The prediction models (model parameters in Table 2) were not validated in DACHS. The prediction error of the 5-year OS remained unchanged after adding the PHS terms to the clinical models (0.173 in resected stage III colon patients; 0.145 in mCRC patients) (Fig. 1).

In DACHS, we also failed to replicate the two PHSs and the interaction with treatment in Cox models for OS (Table 2). Unadjusted Kaplan–Meier curves also showed no difference in OS for patients according to PHSs (Supplementary Fig. S3).

We used an unconventional method to construct predictive PHSs for oxaliplatin treatment outcome for rCC and for mCRC based on genome-wide interaction analysis. However, neither the prediction models nor the PHSs was validated in the independent replication cohort. We cannot rule out that the findings in the discovery cohort were simply due to chance or type I error. The meager discriminative ability of PHS compared to clinical factors may have remained undetected in the replication cohort due to limited power.

The principal components approach may have limitations when applied to highly correlated SNP data. Therefore, we constructed the classical polygenic score using thresholding and pruning for comparison; however, the prediction did not improve.

Limitations of the study should be considered when interpreting the results. Discovery cohorts, NCCTG N0147 and NCCTG N9741, were not originally designed to evaluate the efficacy of oxaliplatin. Patients were defined as having received oxaliplatin-based treatment if they were assigned regimens including oxaliplatin. Otherwise, they were considered to have received the other treatment without oxaliplatin. For that reason, only 4% of the patients from the NCCTG N0147 study did not receive the oxaliplatin-based regime, which limits the statistical power to detect interaction with treatment. The patients who did not receive oxaliplatin showed longer OS compared to the patients who received oxaliplatin (Supplementary Table S1). In addition, we were unable to examine tumor response to oxaliplatin using metrics based on tumor size, like early tumor shrinkage or depth of response, which have been shown to be clinical outcomes in metastatic colorectal cancer. OS is, however, one of the widely used end-points in prospective studies to predict long-term survival.

Strengths of this study include assessing predictive markers (interaction with treatment). Considering that oxaliplatin is not used alone but in association with other chemotherapeutic drugs, prognostic markers from patients who received a combination of oxaliplatin and the other drugs provide limited information on oxaliplatin specifically as a treatment option. Another strength of this study was testing a novel approach to generate PHSs that efficiently combined information from multiple predictive SNPs, thus avoiding the necessity of multiple interaction terms in prediction models for estimating the likelihood of benefit from a particular treatment versus another. Lastly, we accounted for all relevant clinical variables to address the research question on integrating PHS to improve existing clinical models. All analyses were stratified for tumor stage to account for heterogeneity in patient characteristics and corresponding different clinical outcome.

In conclusion, despite efficiently combining multiple SNPs to predict individual benefit from oxaliplatin-containing treatment the predictive PHSs failed to replicate.

Q. Shi reports personal fees from Yiviva Inc, grants from Celgene, personal fees from Regeneron Pharmaceuticals, Inc, Hoosier Cancer Research Network, and Chugai Pharmaceutical Co., Ltd; grants from Bristol Myers Squibb/Celgene, Bristol Myers Squibb, Roche/Genentech, and Janssen; and grants from Novartis outside the submitted work. R.M. Goldberg reports personal fees from Taiho, Astra Zeneca, Bayer, G1 Therapeutics, Compass Therapeutics, UpToDate, Eisai, Sorrento Therapeutics, GSK, Merck, Advanced Chemotherapy Technologies, and Innovative Cellular Therapeutics outside the submitted work. H. Brenner reports grants from German Federal Ministry of Education and Research during the conduct of the study. A.T. Chan reports grants and personal fees from Pfizer Inc., personal fees from Bayer Pharma AG and Boehringer Ingelheim, and grants from Zoe Ltd. outside the submitted work. No disclosures were reported by the other authors.

H.A. Park: Formal analysis, validation, visualization, methodology, writing–original draft, project administration, writing–review and editing. D. Edelmann: Methodology, writing–review and editing. F. Canzian: Writing–review and editing. T.A. Harrison: Data curation, writing–review and editing. X. Hua: Data curation, writing–review and editing. Q. Shi: Data curation, writing–review and editing. A. Silverman: Data curation. M. Schneider: Writing–review and editing. R.M. Goldberg: Data curation, writing–review and editing. S.R. Alberts: Data curation, writing–review and editing. M. Hoffmeister: Data curation, funding acquisition. H. Brenner: Data curation, funding acquisition, writing–review and editing. A.T. Chan: Data curation. U. Peters: Data curation, writing–review and editing. P.A. Newcomb: Data curation, writing–review and editing. J. Chang-Claude: Conceptualization, data curation, supervision, funding acquisition, methodology, writing–review and editing.

N0147 was conducted as a collaborative trial of the North Central Cancer Treatment Group (NCCTG), Mayo Clinic, and was supported in part by Public Health Service grants CA25224, CA37404, CA35103, CA35113, CA35272, CA114740, CA32102, CA14028, CA49957, CA21115, CA31946, CA180868, and CA189867 from the National Cancer Institute, Department of Health and Human Services. NCCTG N9741 was supported by National Institutes of Health grants CA25224, CA32102, CA38926, CA21115, and CA77202. The Alliance for Clinical Trials in Oncology was supported by the National Cancer Institute of the National Institutes of Health grant U10CA180821. The DACHS study was supported by grants from the German Research Council (BR1704/6-1z, BR1704/6-3, BR1704/6-4, BR1704/6-6, CH117/1-1, BR1704/17-1, and HO5117/2-1) and the German Federal Ministry of Education and Research (01KH0404, 01ER0814, 01ER0815, and 01GL1712). GECCO was supported by the NCI/NIH, USA Department of Health and Human Services, grant R01CA176272.

The publication costs of this article were defrayed in part by the payment of publication fees. Therefore, and solely to indicate this fact, this article is hereby marked “advertisement” in accordance with 18 USC section 1734.