Abstract Genome-wide association studies (GWAS) have identified a wealth of germline risk variants, but the clinical impact of these findings has been modest. In response, GWAS results have been used to create polygenic risk scores (PRS) that summarize an individual's genetic susceptibility profile. We assembled PRS for 16 cancers in the UK Biobank (n=413753, 52 to 4170 incident cases) to quantify the extent to which common, low-penetrance risk variants improve upon risk discrimination provided by family history and modifiable exposures. PRS were derived by abstracting variants with minor allele frequency ≥0.01 and P<5 × 10−8 from the literature and selecting independent (LD r2<0.3) variants with the smallest p-value. To account for uncertainty in the reported effect sizes (β), we combined risk alleles in the PRS using inverse variance weights: β/SE(β). Cancer-specific PRS, modifiable risk factors, and family history were modelled using Cox regression accounting for competing risks. All models were well-calibrated and each PRS was associated with the target cancer (P<0.01). Risk discrimination was assessed using Harrell's C-index between 1 and 5 years of follow-up time. An interesting initial observation was that replacing self-reported family history of cancer with the PRS improved prediction accuracy for prostate (C=0.763, ΔC=0.047), breast (C=0.618, ΔC=0.060), and colorectal (C=0.708, ΔC=0.029) cancers. Incorporating the PRS in addition to all risk factors available in our data resulted in substantial gains in risk discrimination for several cancer sites: testes (CPRS=0.766, ΔC=0.138), thyroid (CPRS=0.692, ΔC=0.099), leukemia (CPRS=0.756, ΔC=0.061), breast (CPRS=0.631, ΔC=0.060), prostate (CPRS=0.768, ΔC=0.051), melanoma (CPRS=0.664, ΔC=0.042), and colon/rectum (CPRS=0.716, ΔC=0.030). Next, we assessed risk stratification by examining 5-year absolute risk trajectories. Stratifying by percentiles of PRS (high: ≥80%, average: >20% to <80%, low: ≤20%) and modifiable risk factor profiles revealed significantly diverging risk trajectories (P<10−10) for 15 out of 16 cancers. There was also evidence of PRS-environment interaction, with larger than additive risk differences observed for melanoma (P=3.3 × 10−122), post-menopausal breast (P=1.3 × 10−21), colorectal (P=1.3 × 10−208), lung (P=1.1 × 10−37), bladder (P=1.5 × 10−50), and kidney (P=5.5 × 10−29) cancers. Lastly, we quantified the proportion of cancer risk at the population level that is attributed to genetic vs. modifiable risk factors. High genetic risk (PRS≥80th percentile) explained between 4.0% (lung) and 30.3% (testicular) of new cases. For many cancers the attributable fraction (AF) for PRS exceeded the AF for modifiable risk factors or family history: thyroid (AFPRS=0.268, P=1.7 × 10−9), prostate (AFPRS=0.232, P=5.5 × 10−158), colorectal (AFPRS=0.167, P=9.2 × 10−50), breast (AFPRS=0.166, P=2.6 × 10−85), and melanoma (AFPRS=0.139, P=1.3 × 10−23). PRS was the only significant contributor to cancer risk, other than age and sex, for testicular cancer (AFPRS=0.303, P=4.5 × 10−4), leukemia (AFPRS=0.269, P=4.5 × 10−4), and lung cancer in never smokers (AFPRS=0.077, P=0.045). In summary, we provide evidence from an independent validation cohort that supports the potential clinical utility of PRS in complementing conventional cancer risk factors to improve and refine risk prediction. Citation Format: Linda Kachuri, Rebecca E. Graff, Karl Smith-Byrne, Travis J. Meyers, Sara R. Rashkin, Elad Ziv, John S. Witte, Mattias Johansson. Pan-cancer analysis of polygenic risk scores reveals improvement in risk prediction and stratification [abstract]. In: Proceedings of the Annual Meeting of the American Association for Cancer Research 2020; 2020 Apr 27-28 and Jun 22-24. Philadelphia (PA): AACR; Cancer Res 2020;80(16 Suppl):Abstract nr 2319.