Active learning has been successfully used in the chemistry to improve the performance of the learner including the out-of-sample generalisation monitoring. The standard query functions utilise the model characteristics such as model uncertainty and related information quantities. While focusing on epistemic uncertainty, the learner utility function often omits the aleatoric uncertainty or exploration of the data manifold structure. In this paper we propose two novel query functions which incorporate the structural information about the chemical diversity of the data. We investigate the performance in comparison to various active learning strategies and under the distributional shifted datasets.