Synapse loss is an early event in late-onset Alzheimer's disease (LOAD). In this study we have assessed the capacity of a polygenic risk score (PRS) restricted to synapse-encoding loci to predict LOAD. We used summary statistics from the IGAP genome-wide association meta-analysis of 74,046 subjects for model construction and tested the "Synaptic PRS" in two independent datasets of controls and pathologically-confirmed LOAD. The mean Synaptic PRS was 2.3-fold higher in LOAD compared to controls (p<0.0001) with a predictive accuracy of 72% in the target dataset (n=439) and 73% in the validation dataset (n=136), a 5-6% improvement compared to the APOE locus (p<0.00001). The model comprises 8 variants from 4 previously-identified (BIN1, PTK2B, PICALM, APOE) and 2 novel (DLG2, MINK1) LOAD loci involved in glutamate signaling (p=0.01) or APP catabolism/tau binding (p=0.005). As the simplest PRS model with good predictive accuracy to predict LOAD, we conclude that synapse-encoding genes are enriched for LOAD risk-modifying loci.The Synaptic PRS could be used to identify individuals at risk of LOAD before symptom onset.