no / Predictive toxicology is concerned with the development of models that are able to predict the toxicity of chemicals. A reliable prediction of toxic effects of chemicals in living systems is highly desirable in cosmetics, drug design or food protection to speed up the process of chemical compound discovery while reducing the need for lab tests. There is an extensive literature associated with the best practice of model generation and data integration but management and automated identification of relevant models from available collections of models is still an open problem. Currently, the decision on which model should be used for a new chemical compound is left to users. This paper intends to initiate the discussion on automated model identification. We present an algorithm, based on Pareto optimality, which mines model collections and identifies a model that offers a reliable prediction for a new chemical compound. The performance of this new approach is verified for two endpoints: IGC50 and LogP. The results show a great potential for automated model identification methods in predictive toxicology.

Toxic compounds, such as pesticides, are routinely tested against a range of aquatic, avian and mammalian species as part of the registration process. The need for reducing dependence on animal testing has led to an increasing interest in alternative methods such as in silico modelling. The QSAR (Quantitative Structure Activity Relationship)-based models are already in use for predicting physicochemical properties, environmental fate, eco-toxicological effects, and specific biological endpoints for a wide range of chemicals. Data plays an important role in modelling QSARs and also in result analysis for toxicity testing processes. This research addresses number of issues in predictive toxicology. One issue is the problem of data quality. Although large amount of toxicity data is available from online sources, this data may contain some unreliable samples and may be defined as of low quality. Its presentation also might not be consistent throughout different sources and that makes the access, interpretation and comparison of the information difficult. To address this issue we started with detailed investigation and experimental work on DEMETRA data. The DEMETRA datasets have been produced by the EC-funded project DEMETRA. Based on the investigation, experiments and the results obtained, the author identified a number of data quality criteria in order to provide a solution for data evaluation in toxicology domain. An algorithm has also been proposed to assess data quality before modelling. Another issue considered in the thesis was the missing values in datasets for toxicology domain. Least Square Method for a paired dataset and Serial Correlation for single version dataset provided the solution for the problem in two different situations. A procedural algorithm using these two methods has been proposed in order to overcome the problem of missing values. Another issue we paid attention to in this thesis was modelling of multi-class data sets in which the severe imbalance class samples distribution exists. The imbalanced data affect the performance of classifiers during the classification process. We have shown that as long as we understand how class members are constructed in dimensional space in each cluster we can reform the distribution and provide more knowledge domain for the classifier.

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. / BBSRC and Syngenta (International Research Centre at Jealott’s Hill, Bracknell, UK).

This thesis develops an Artificial Intelligence (AI) approach intended for accurate patient stratification and precise diagnostics/prognostics in clinical and preclinical applications. The rapid advance in high throughput technologies and bioinformatics tools is still far from linking precisely the genome-phenotype interactions with the biological mechanisms that underlie pathophysiological conditions. In practice, the incomplete knowledge on individual heterogeneity in complex diseases keeps forcing clinicians to settle for surrogate endpoints and therapies based on a generic one-size-fits-all approach. The working hypothesis is that AI can add new tools to elaborate and integrate together in new features or structures the rich information now available from high-throughput omics and bioimaging data, and that such re- structured information can be applied through predictive models for the precision medicine paradigm, thus favoring the creation of safer tailored treatments for specific patient subgroups. The computational techniques in this thesis are based on the combination of dimensionality reduction methods with Deep Learning (DL) architectures to learn meaningful transformations between the input and the predictive endpoint space. The rationale is that such transformations can introduce intermediate spaces offering more succinct representations, where data from different sources are summarized. The research goal was attacked at increasing levels of complexity, starting from single input modalities (omics and bioimaging of different types and scales), to their multimodal integration. The approach also deals with the key challenges for machine learning (ML) on biomedical data, i.e. reproducibility, stability, and interpretability of the models. Along this path, the thesis contribution is thus the development of a set of specialized AI models and a core framework of three tools of general applicability: i. A Data Analysis Plan (DAP) for model selection and evaluation of classifiers on omics and imaging data to avoid selection bias. ii. The histolab Python package that standardizes the reproducible pre-processing of Whole Slide Images (WSIs), supported by automated testing and easily integrable in DL pipelines for Digital Pathology. iii. Unsupervised and dimensionality reduction techniques based on the UMAP and TDA frameworks for patient subtyping. The framework has been successfully applied on public as well as original data in precision oncology and predictive toxicology. In the clinical setting, this thesis has developed1: 1. (DAPPER) A deep learning framework for evaluation of predictive models in Digital Pathology that controls for selection bias through properly designed data partitioning schemes. 2. (RADLER) A unified deep learning framework that combines radiomics fea- tures and imaging on PET-CT images for prognostic biomarker development in head and neck squamous cell carcinoma. The mixed deep learning/radiomics approach is more accurate than using only one feature type. 3. An ML framework for automated quantification tumor infiltrating lymphocytes (TILs) in onco-immunology, validated on original pathology Neuroblastoma data of the Bambino Gesu’ Children’s Hospital, with high agreement with trained pathologists. The network-based INF pipeline, which applies machine learning models over the combination of multiple omics layers, also providing compact biomarker signatures. INF was validated on three TCGA oncogenomic datasets. In the preclinical setting the framework has been applied for: 1. Deep and machine learning algorithms to predict DILI status from gene expression (GE) data derived from cancer cell lines on the CMap Drug Safety dataset. 2. (ML4TOX) Deep Learning and Support Vector Machine models to predict potential endocrine disruption of environmental chemicals on the CERAPP dataset. 3. (PathologAI) A deep learning pipeline combining generative and convolutional models for preclinical digital pathology. Developed as an internal project within the FDA/NCTR AIRForce initiative and applied to predict necrosis on images from the TG-GATEs project, PathologAI aims to improve accuracy and reduce labor in the identification of lesions in predictive toxicology. Furthermore, GE microarray data were integrated with histology features in a unified multi-modal scheme combining imaging and omics data. The solutions were developed in collaboration with domain experts and considered promising for application.

Toxic compounds, such as pesticides, are routinely tested against a range of aquatic, avian and mammalian species as part of the registration process. The need for reducing dependence on animal testing has led to an increasing interest in alternative methods such as in silico modelling. The QSAR (Quantitative Structure Activity Relationship)-based models are already in use for predicting physicochemical properties, environmental fate, eco-toxicological effects, and specific biological endpoints for a wide range of chemicals. Data plays an important role in modelling QSARs and also in result analysis for toxicity testing processes. This research addresses number of issues in predictive toxicology. One issue is the problem of data quality. Although large amount of toxicity data is available from online sources, this data may contain some unreliable samples and may be defined as of low quality. Its presentation also might not be consistent throughout different sources and that makes the access, interpretation and comparison of the information difficult. To address this issue we started with detailed investigation and experimental work on DEMETRA data. The DEMETRA datasets have been produced by the EC-funded project DEMETRA. Based on the investigation, experiments and the results obtained, the author identified a number of data quality criteria in order to provide a solution for data evaluation in toxicology domain. An algorithm has also been proposed to assess data quality before modelling. Another issue considered in the thesis was the missing values in datasets for toxicology domain. Least Square Method for a paired dataset and Serial Correlation for single version dataset provided the solution for the problem in two different situations. A procedural algorithm using these two methods has been proposed in order to overcome the problem of missing values. Another issue we paid attention to in this thesis was modelling of multi-class data sets in which the severe imbalance class samples distribution exists. The imbalanced data affect the performance of classifiers during the classification process. We have shown that as long as we understand how class members are constructed in dimensional space in each cluster we can reform the distribution and provide more knowledge domain for the classifier.