National Technical University of Athens (NTUA)
Development of procedures for the design of chemical structures with artificial intelligence and machine learning methodologies
Abstract
dc:descriptionThis Dissertation focuses on the development of advanced Machine Learning (ML) and Deep Learning (DL) methodologies within the field of cheminformatics, to tackle key challenges in the design of novel chemical compounds. Particular emphasis is given to drug discovery, addressing not only therapeutic efficacy but also safety considerations, in alignment with the principles of the emerging Safe-by-Design (SbD) framework. Drug discovery and development is a complex, time-consuming and costly process, with low clinical success rates. Therefore, developing predictive and reliable computational approaches is essential to identify promising candidates and eliminate unsafe compounds early in the discovery process. The research focuses on four main objectives: (i) prediction of ADMET (Absorption, Distribution, Metabolism, Excretion, Toxicity) properties, (ii) assessment of skin toxicity within the SbD framework, (iii) prediction of physicochemical properties, and (iv) de novo design of bioactive molecules using both modern and traditional computer-aided drug design (CADD) methods. The first part of this work focuses on molecular property prediction through the development of a specialized neural network architecture, the Atom-Attention Message Passing Neural Network (AA-MPNN). Molecules are represented as molecular graphs, enabling the model to capture detailed structural information through message passing and atom-level attention. To overcome the challenge of data scarcity in molecular property prediction, a self-supervised learning technique was integrated into the workflow. More specifically, contrastive learning was proposed to enhance molecular representation learning and, consequently, improve downstream molecular property prediction tasks. A pretrained AA-MPNN model was derived by minimizing the contrastive loss between positive molecular pairs. This pretrained model was subsequently fine-tuned to predict two key ADMET properties, blood-brain barrier (BBB) permeability and Caco-2 cell line permeability. The resulting pretrained models achieved superior performance compared to non-pretrained AA-MPNN and conventional ML baselines. In addition, the attention mechanism improved interpretability by identifying atomic contributions that affect drug permeability. The second part of this study presents the development of Quantitative Structure-Property Relationships (QSPR) models within a computational safety assessment framework. Multi-view molecular representations combining molecular fingerprints and molecular descriptors were integrated into a Fully Connected Neural Network (FCNN) to predict skin sensitization, skin irritation/corrosion, and acute dermal toxicity. The FCNN with multi-view molecular representations achieved superior performance compared to single-view models. SHapley Additive exPlanations (SHAP) analysis was performed to provide mechanistic insights, and highlight structural features and descriptors associated with skin toxicity effects. In addition, two QSPR models were developed for predicting BBB permeability and hydration free energy, using traditional ML models. These models complemented a suite of additional models of physicochemical and toxicity properties. Among the evaluated algorithms, k-nearest neighbors (kNN) demonstrated strong predictive performance, computational efficiency, and interpretability by providing nearest neighbors in the predictive chemical space. In this study, BBB permeability served as a common endpoint for both an advanced DL model incorporating self-supervised learning and conventional ML model, enabling a direct comparison of their predictive capabilities. The property prediction models developed in this Thesis were validated on external datasets consisting of known drugs with experimentally defined physicochemical or toxicity properties. To enhance accessibility, all validated models were implemented as web-based tools, allowing their broader application from scientific community. Furthermore, all datasets used for model development were made available in accordance with the FAIR (Finable, Accessible, Interoperable, Reusable) data principles. The final part of this Dissertation included an overview of deep generative models (DGM) used in de novo drug design and the development of a DGM combined with Reinforcement Learning (RL) for generating compounds with improved binding affinity against a target protein. Using a serine/threonine-protein kinase as a case study, a hybrid reward function was developed that integrates ML-based binding affinities with periodic physics-based molecular docking evaluations. The fine-tuned RL model efficiently explored the chemical space, producing compounds with improved binding affinity scores. Overall, this Dissertation contributes to advancing Artificial Intelligence (AI)-driven drug discovery through the development of interpretable and accessible predictive and generative modeling frameworks. The work demonstrates how data-driven approaches can accelerate the identification, optimization, and design of compounds with favorable pharmacological and safety profiles.
Degree
thesis:*- Grantor dc:publisher
- National Technical University of Athens (NTUA)
- Year dc:date
- 2025
Author and committee
dc:creator, dc:contributor.*- Authors dc:creator
-
- Koutroumpa, Nikoletta-Maria
- Κουτρούμπα, Νικολέττα-Μαρία
Subjects
dc:subject × 25- Μηχανική μάθηση
- Βαθιά μάθηση
- Σχεδιασμός φαρμάκων
- Χημειοπληροφορική
- Μοντέλα πρόβλεψης
- Παραγωγικά μοντέλα
- Μοριακή πρόσδεση
- Machine learning
- Deep learning
- Drug discovery
- Cheminformatics
- Predictive modeling
- Generative modeling
- Molecular docking
- Επιστήμη Ηλεκτρονικών Υπολογιστών και Πληροφορική
- Φυσικές Επιστήμες
- Τεχνητή νοημοσύνη
- Επιστήμη Χημικού Μηχανικού
- Επιστήμες Μηχανικού και Τεχνολογία
- Χημική μηχανική
- Computer and Information Sciences
- Natural Sciences
- Artificial Intelligence
- Chemical Engineering
- Engineering and Technology
Rights
- Language dc:language
- eng
Identifiers
dc:identifier.*- Identifier
- 10.12681/eadd/60840