REPOSITORY > RESULTS

Abstract

QSAR methods are a cornerstone of contemporary cheminformatics, yet conventional global models remain limited in their ability to provide reliable predictions for complex chemical spaces and endpoints. This dissertation develops, implements and applies a framework for localized QSAR modelling that attempts to address these limitations.
The aiQSAR methodology is established with a complete workflow for QSAR localization that includes molecular descriptor and fingerprint calculation, similarity-based selection of local groups, descriptor filtering and application of multiple machine learning methods for consensus predictions. The Applicability Domain Measure assigns a quantitative reliability score to each target. This approach is validated within a community-wide computational toxicology initiative with external benchmarking and across five acute rat oral toxicity endpoints, demonstrating that aiQSAR methodology achieves predictive performance that is competitive with, and for some challenges superior to, counterpart models.
This methodology is implemented into an open-source software designed to make local QSAR modelling accessible to both expert and non-specialist users. aiQSAR integrates 38 pre-built endpoints through a dual interface with a streamlined Basic mode for routine predictions and an Advanced mode enabling full control over local group selection, similarity parameters, consensus strategies and with a model exporting functionality. Multiple output formats are supported, including concise PDF summaries, detailed text reports suitable for downstream computational analysis and single-compound reports that facilitate expert inspection of local group chemicals and the associated similarity metrics.
The developed methodology is applied within a novel context, for prediction of chemically induced differential gene expression. Transcriptomic profiles of treated samples for a large panel of genes are quantified using the TempO-Seq RNA sequencing platform and associated bioinformatics pipeline. Distinct patterns of differential expression are revealed through exploratory data analysis, indicating that almost all genes exhibit the differential expression following chemical treatment in only a single direction, either upregulation or downregulation, for all modulating chemicals. Finally, aiQSAR modelling results demonstrate that for a large subset of genes, differential expression is predictably encoded in chemical structure and that local QSAR models can capture these relationships.
Taken together, the theoretical development, software implementation and transcriptomics case study demonstrate that localization in QSAR is a robust and versatile strategy. The aiQSAR methodology and software provide an immediately usable platform for chemical risk assessment, drug discovery and mechanistic toxicology, while establishing a foundation for future integration of QSAR with transcriptomics in emerging scientific areas of computational, predictive and developmental biology.