A prototype collection of knowledge on ligands in metal complexes, termed a ligand knowledge base (LKB), has been developed. This contribution describes the design of DFT-calculated descriptors for monodentate phosphorus(III) donor ligands in a range of representative complexes. Using the resulting data, a ligand space is mapped and predictive models are derived for metal complexes. Important characteristics, including chemical, computational and statistical robustness for the generation and exploitation of such an LKB are described. Chemical robustness ensures transferability of the descriptors, as well as comprehensive sampling of ligand space. To make the calculations amenable to automation in an e-science setting, a reliable, well-defined computational approach has been sought from which the descriptors can be readily extracted. The LKB has been explored with multivariate statistical methods. Principal component analysis (PCA) is used for the mapping of chemical space, projecting multiple descriptors into scatter plots which illustrate the clustering of chemically similar ligands. Interpretation of the resulting principal components in terms of established steric and electronic properties and the importance of its statistical robustness to variations in the ligand set are discussed. Multiple linear regression (MLR) models have been derived, demonstrating the versatility of the descriptors for modeling varied experimentally determined parameters (bond lengths, reaction enthalpies and bond-stretching frequencies). The importance of re-sampling methods for testing the robustness of predictions is highlighted. A strategy for the construction of a robust LKB suitable for the modeling of ligand and complex behavior is outlined based on these observations.