Integrating background knowledge from internet databases into predictive toxicology models

Abstract

While data integration for data analysis has been investigated extensively in biological applications, it has not yet been so much the focus in computational chemistry and quantitative structure–activity relationship (QSAR) research. With the availability and growing number of chemical databases on the web, such data integration efforts become an intriguing possibility (and, in fact, a necessity). In this paper, we take a first step towards the following vision and scenario for predictive toxicology applications. Given a new structure to be predicted, the first step would be to gather (integrate) all relevant information from internet databases for the structure itself, and all structures with available information for the endpoint of interest. In a second step, the collected information is combined statistically into a prediction of the new structure. We simulate this scenario with three endpoints (data sets) from the DSSTox database and collect information from three public chemical databases: PubChem, ChemBank and Sigma-Aldrich. In the experiments, we investigate whether the addition of background knowledge from the three databases can improve predictive performance (over using chemical structure alone) in a statistically significant way. For this purpose, we define groups of features (belonging together from an application point of view) from the three databases, and perform a variant of forward selection to include these feature groups in a prediction model. Our experiments show that the integration of background knowledge from internet databases can significantly improve prediction performance, especially for regression tasks.

Keywords:

• (Q)SAR
• data integration
• internet databases
• cheminformatics
• machine learning

Notes

Notes

1. It is interesting to note that data integration has not received attention in prediction toxicology before, although it could have been done technically 4.

2. This is done for simplicity. Note that (a) FreeTreeMiner works in an unsupervised fashion, and (b) we are just interested in the relative performance of the substructures alone and the substructures plus the background knowledge. Substructures alone are used as a starting point here–the better the substructures, the harder it will be for the background knowledge to improve upon them.