Abstract Background: Associations between candidate germline genetic variants and treatment outcome of oxaliplatin, a drug commonly used for patients with colorectal cancer, have been reported but not robustly established. This study aimed to construct polygenic hazard scores (PHSs) as predictive markers for oxaliplatin treatment outcome by using a supervised principal component approach (PCA). Methods: Genome-wide association analysis for overall survival, including interaction terms (SNP*treatment type) was carried out using two phase III trials, 3,098 resected stage III colon cancer (rCC) patients of NCCTG N0147 and 506 metastatic colorectal cancer (mCRC) patients of NCCTG N9741, separately. SNPs showing interaction with genome-wide significance (P < 5 × 10–8) were selected for PCA to derive a PHS. PHS interaction with treatment was included in Cox regression models to predict outcome. Replication of prediction models was performed in an independent cohort, DACHS. Results: The two PHSs based on the first two principal components of selected SNPs (15SNPs for rCC and 13SNPs for mCRC) were used to construct interaction terms with treatment type and included in models adjusted for clinical covariables. However, in the DACHS study, the addition of the two PHS terms to clinical models did not improve the prediction error in either patients with rCC or mCRC. PHS interaction was also not replicated. Conclusions: The PHSs derived using principal components efficiently combined multiple predictive SNPs for estimating likelihood of benefit from oxaliplatin versus other treatment but could not be replicated. Impact: These results highlight the potential but also challenges in generating evidence for a predictive polygenic score for oxaliplatin efficacy.