Learning continuous and data-driven molecular descriptors by translating equivalent chemical representations