Modernisation of computational chemistry and cheminformatics with eScience techniques : applications to chemical property prediction

The prediction of macroscopic chemical properties is recognised as a valuable ability. Statistical models such as Quantitative Structure-Property Relationships (QSPRs) are only as good as the data and assumptions upon which they are based. In this thesis, a number of difficulties in handling large data sets, including trust, relevancy and interpretation are made clear and an approach to modernising chemical data storage is presented. The use of Semantic Web technologies to increase the availability and usefulness of the chemical information is demonstrated, placing emphasis on issues of provenance and quality as well as machine-readability. A method of allowing computers to understand scientific units is a significant part of this, solving many potential problems in data re-use. Existing methods of encoding scientific units are found to be of limited scope and a new representation is formulated to make full use of and reinforce the intrinsic value of units with their measurements. The new system for chemical data management is then used to assemble a unique data set combining data from many sources in order to gain new understanding of the problem of melting point prediction. Solid state data forms the cornerstone of this approach. Attempts to model melting point in a high throughput manner highlight the issues involved in solid state simulation and the limitations of simple descriptors. Further experimental data and modelling reveals the importance of the interrelation of enthalpy and entropy, and the stringent requirements for accuracy in order to achieve useful errors of melting point prediction.