Interpretation, Identification and Reuse of Models. Theory and algorithms with applications in predictive toxicology.

Abstract

This thesis is concerned with developing methodologies that enable existing models to be effectively reused. Results of this thesis are presented in the framework of Quantitative Structural-Activity Relationship (QSAR) models, but their application is much more general. QSAR models relate chemical structures with their biological, chemical or environmental activity. There are many applications that offer an environment to build and store predictive models. Unfortunately, they do not provide advanced functionalities that allow for efficient model selection and for interpretation of model predictions for new data. This thesis aims to address these issues and proposes methodologies for dealing with three research problems: model governance (management), model identification (selection), and interpretation of model predictions. The combination of these methodologies can be employed to build more efficient systems for model reuse in QSAR modelling and other areas. The first part of this study investigates toxicity data and model formats and reviews some of the existing toxicity systems in the context of model development and reuse. Based on the findings of this review and the principles of data governance, a novel concept of model governance is defined. Model governance comprises model representation and model governance processes. These processes are designed and presented in the context of model management. As an application, minimum information requirements and an XML representation for QSAR models are proposed. Once a collection of validated, accepted and well annotated models is available within a model governance framework, they can be applied for new data. It may happen that there is more than one model available for the same endpoint. Which one to chose? The second part of this thesis proposes a theoretical framework and algorithms that enable automated identification of the most reliable model for new data from the collection of existing models. The main idea is based on partitioning of the search space into groups and assigning a single model to each group. The construction of this partitioning is difficult because it is a bi-criteria problem. The main contribution in this part is the application of Pareto points for the search space partition. The proposed methodology is applied to three endpoints in chemoinformatics and predictive toxicology. After having identified a model for the new data, we would like to know how the model obtained its prediction and how trustworthy it is. An interpretation of model predictions is straightforward for linear models thanks to the availability of model parameters and their statistical significance. For non linear models this information can be hidden inside the model structure. This thesis proposes an approach for interpretation of a random forest classification model. This approach allows for the determination of the influence (called feature contribution) of each variable on the model prediction for an individual data. In this part, there are three methods proposed that allow analysis of feature contributions. Such analysis might lead to the discovery of new patterns that represent a standard behaviour of the model and allow additional assessment of the model reliability for new data. The application of these methods to two standard benchmark datasets from the UCI machine learning repository shows a great potential of this methodology. The algorithm for calculating feature contributions has been implemented and is available as an R package called rfFC.