Computational prediction of eye irritation and corrosion potential of chemicals is one of the key strategies for animal-free evaluation of ocular toxicity. Over the years, the National Toxicology Program (NTP) Interagency Center for the Evaluation of Alternative Toxicological Methods (NICEATM) has compiled and curated a database of in vivo eye irritation studies from scientific literature and provided by stakeholders. The database contains around 800 annotated records of over 500 unique substances with their eye irritation categories according to GHS and US EPA hazard classifications. We developed a set of in silico models for EPA hazard classification categories at 100% and 10% potency thresholds (by mass or volume content) for the chemical substances in the eye irritation database, many of which are formulations and mixtures. Conventional models (based on chemical structure of the largest component of the test substance) achieve validated balanced accuracy in the range of 67-77% and 84-89% for the 100% and 10% potency thresholds, respectively. Comparatively, the mixture-based models, which account for all components in the substance by weighted feature averaging, showed higher accuracy of 69-78% and 85-91% for the respective potency thresholds. We also noted a strong trend between the pH feature metric calculated for each substance and its activity category. Namely, across all the models, calculated pH of inactive substances is on average 0.8 pH-units away from the neutral pH, while for active substances, it is >3 pH-units away. This pH dependency is especially important for complex substances that are influenced by multiple components. In the future, these in silico models can benefit from additional high quality in vivo data sources (e.g., European Chemicals Agency dossiers) and by including additional variable inputs such as in vitro eye irritation test method results. This work was funded with US federal funds from the NIEHS/NIH/HHS under Contract HHSN273201500010C.