Avian Influenza viruses (AIVs) infect a broad host range despite having a natural reservoir in wild aquatic birds. Whilst most strains stay within their host species, some break the species barrier through genetic adaptations. We are most concerned about zoonotic cases, where a human becomes infected. Despite these events being rare, they are associated with high mortality and introduce the risk of onward human-to-human transmission of AIV. As a novel pathogen within the human population, this could have pandemic potential. Using genetic composition features for 8 AIV proteins drawn from viral sequence data, we employ machine-learning algorithms to classify AIV cases as zoonotic or not. These genetic features encode host ´signatures´ which can indicate zoonosis and include frequency measures such as dipeptide composition and amino acid physiochemical properties. We consistently find XGBoost to outperform all other algorithms. We optimise parameters for ten classification models: one for each of the 8 proteins and two combined models. Following this, we show that a multi-model approach gives the best performing prediction for AIV zoonosis. We have identified all 8 proteins as having a role in predicting zoonotic transmission. Of particular importance is the PB2 and HA proteins, with specific amino acid physiochemical properties such as charge, secondary structure and hydrophobicity amongst the most indicative features in our combined models. Our alignment-free computational study can identify AIV cases still within avian hosts which are genetically closest to zoonotic AIV cases, thereby identifying the cases most likely to cross the species barrier. In a resource limited environment, our model could be used to quickly identify high priority cases for further investigation.