A Report on Fourth Joint Sheffield Conference on
Chemoinformatics June 18-20, University of Sheffield, UK

Johann Gasteiger of the University of Erlangen-Nürnberg showed how modeling of chemical reactions can help in drug discovery. For example, in lead discovery and lead optimization, an estimate of synthetic accessibility can be useful. Gasteiger’s team has devised a scoring method that rapidly evaluates synthetic accessibility of structures based on structural complexity, similarity to available starting materials, and assessment of strategic bonds where a structure can be decomposed to obtain simpler fragments^[5]. These individual components are combined to give an overall score of synthetic accessibility by an additive scheme. The system is called SYLVIA.

Modeling metabolism is also possible. To this end, XENIA, an in-house CYP450 database, has been developed at the University of Erlangen-Nürnberg. MetaboGen systematically generates all metabolites of a drug, applying a set of the most important phase I reactions. ISOCYP is a web service for prediction of the predominant P450 isoform^[6]. Molecular Networks also supplies the biochemical pathways database, BioPath.

Markus Wagener of Organon presented a novel, rule-based method, SyGMa (Systematic Generation of Metabolites) that predicts potential metabolites of a given parent structure. The method is based on reaction rules derived from metabolic reactions that occur in man, reported in Elsevier MDL’s Metabolite database. The database was filtered (to remove assumed metabolites, incomplete and large structures etc.) to give 7,307 biotransformations as a training set. Reaction templates were encoded as SMIRKS and reaction probabilities were calculated based on training set statistics. The predicted metabolites are ranked according to the empirical probability score. Evaluation of the method demonstrated a significant enrichment of true metabolites at the top of the ranking list. The current rule set covers about 70% of the human in vivo data of the Metabolite database. To gain an understanding of the nature of the reactions, a similarity analysis of the reaction types was performed using difference fingerprints^[7] calculated by subtracting fingerprints generated from atom environments^[8]. SPE^[9] was used to project the reaction space. Wagener gave some examples of SyGMa, including the pathway for buspirone. Predictions from SyGMa are used at Organon to plan experiments aimed at experimental metabolite identification and to suggest labile sites amenable to optimization by medicinal chemistry.

Metabolism was also the topic of a paper by Anton Schwaighofer of Fraunhofer FIRST. His team, idalab of Berlin, and Bayer Schering Pharma (BSP) have jointly developed machine learning tools to predict the metabolic stability of compounds from drug discovery projects at BSP. They used experimental metabolic stability data from four different in vitro assays. They compared a variety of machine learning approaches in terms of performance, difficulty of the model selection procedures, interpretability, and how the "domain of applicability" can be checked. They concluded that Gaussian Process classification has specific benefits. The effort required for model selection is minimal, so fully automatic re-training is possible. Also, the probabilistic output is easy to interpret and shows almost ideal properties. Competing methods achieve similar performance, but need more careful tuning by an expert. The models developed were validated on recent project data at BSP: the best models are highly accurate and are able to identify the domain of applicability correctly. These models are fully integrated in the working environment at BSP and a tool for automatic regular retraining of the models is currently being implemented. A paper has been submitted to J. Chem. Inf. Model.