Composition-property correlations are fundamental to understand cement-based materials behavior and optimize their formulation. Modelling based on fundamental material component constitutes a reliable tool to establish these correlations with the advantage of better exploring formulation space when compared with the often adopted experimental trial-and-error approaches. In this context, Machine Learning (ML) and Micromechanics-Based (MB) methods have been concurrently used for property prediction from material composition. Here, we show that these techniques can be allies for establishing composition-property correlations. We focus on predictions of Ordinary Portland Cement pastes elastic properties, but the outlined strategy can be extended to other cement systems. Various microstructures representations are considered in MB estimates, including multiscale representations and representations with ellipsoidal inclusions. In contrast, ML predictions do not need any a priori assumption on material microstructure. Predictions using ML and MB yield similar accuracy when compared against test datasets (but ML performed much better regarding the error estimated in training datasets). Working as allies, ML can be deployed to evaluate the (lack of) knowledge over the multi-dimensional parametric domains, and micromechanics provides a theoretical background for property data curation and is a tool to make up for missing data in databases.