Chimaeric human–murine adeno-associated virus (AAV) capsids are described that transduce human primary hepatocytes more efficiently than currently used AAV vectors; the novel vectors may be good clinical candidates. One of the most promising gene-therapy vectors currently in clinical trials is the rAAV8 (recombinant adeno-associated viral serotype 8) virus. Yet the performance of this, and some other AAV vectors, often fails to live up to expectations raised by the results of animal studies. Here Mark Kay et al. use a mouse model with a chimaeric murine/human liver to compare rAAV transduction efficiency in human and mouse cells. They find that the rAAV2 vector transduces both species equally well, but that AAV8 is about 20-fold less efficient at transducing human cells than mouse cells. The data correlate well with clinical data and may go some way towards explaining why rAAV8 underperforms in human trials. The authors suggest that the use of a human primary cell xenotransplant model may more accurately predict potential transduction efficiency in humans than the commonly used mouse and non-human primate models. Recombinant adeno-associated viral (rAAV) vectors have shown early promise in clinical trials1,2,3. The therapeutic transgene cassette can be packaged in different AAV capsid pseudotypes, each having a unique transduction profile. At present, rAAV capsid serotype selection for a specific clinical trial is based on effectiveness in animal models. However, preclinical animal studies are not always predictive of human outcome4,5,6,7,8. Here, in an attempt to further our understanding of these discrepancies, we used a chimaeric human–murine liver model to compare directly the relative efficiency of rAAV transduction in human versus mouse hepatocytes in vivo. As predicted from preclinical and clinical studies4,5,8, rAAV2 vectors functionally transduced mouse and human hepatocytes at equivalent but relatively low levels. However, rAAV8 vectors, which are very effective in many animal models, transduced human hepatocytes rather poorly—approximately 20 times less efficiently than mouse hepatocytes. In light of the limitations of the rAAV vectors currently used in clinical studies, we used the same murine chimaeric liver model to perform serial selection using a human-specific replication-competent viral library composed of DNA-shuffled AAV capsids. One chimaeric capsid composed of five different parental AAV capsids was found to transduce human primary hepatocytes at high efficiency in vitro and in vivo, and provided species-selected transduction in primary liver, cultured cells and a hepatocellular carcinoma xenograft model. This vector is an ideal clinical candidate and a reagent for gene modification of human xenotransplants in mouse models of human diseases. More importantly, our results suggest that humanized murine models may represent a more precise approach for both selecting and evaluating clinically relevant rAAV serotypes for gene therapeutic applications.