Equivariant graph neural networks in drug discover: from property prediction to molecule generation
Advancements in machine learning have revolutionized various aspects of drug discovery, from molecular property prediction to de novo design. This thesis presents a comprehensive exploration of equivariant graph neural networks (EGNNs) and their applications in critical areas of computational drug d...
| Autor: | |
|---|---|
| Tipo de recurso: | tesis doctoral |
| Estado: | Versión publicada |
| Fecha de publicación: | 2024 |
| País: | España |
| Institución: | CBUC, CESCA |
| Repositorio: | TDR. Tesis Doctorales en Red |
| OAI Identifier: | oai:www.tdx.cat:10803/692718 |
| Acceso en línea: | http://hdl.handle.net/10803/692718 |
| Access Level: | acceso abierto |
| Palabra clave: | Equivariant graph neural networks Drug discovery Generative chemistry Diffusion Structure-based drug design Xarxes neuronals de gràfics equivariants Descobriment de drogues Química generativa Difusió Disseny de fàrmacs basat en l'estructura 615 |
| Sumario: | Advancements in machine learning have revolutionized various aspects of drug discovery, from molecular property prediction to de novo design. This thesis presents a comprehensive exploration of equivariant graph neural networks (EGNNs) and their applications in critical areas of computational drug discovery, specifically focusing on property prediction, molecular representation learning, and structure-based de novo ligand design. Toxicity prediction is a critical early-stage filter in drug development. Traditional methods often rely on text-based or 2D representations or simplified 3D models, potentially missing crucial structural information. We hypothesized that full 3D molecular structures in form of ensembles of high-quality conformers could significantly improve toxicity predictions. Next, we explored the de novo molecule generation resulting in the development of an diffusion-based equivariant graph neural network, EQGAT-diff. This framework represents a significant advancement in molecular generation, incorporating both Gaussian and discrete state space diffusion techniques. A key innovation in EQGAT-diff is its timestep-dependent loss weighting scheme, which dramatically improves both training and inference efficiency. Building on the insights and capabilities of EQGAT-diff, we tackled the challenge of structure-based drug design. The result is PILOT, an equivariant diffusion model for de novo 3D ligand generation conditioned on protein pockets. Its multi-objective trajectory-based importance sampling allows for the simultaneous optimization of multiple properties, including binding affinity, synthetic accessibility or toxicity. Finally, we propose PoliGenX, a model based on EQGAT-diff that is conditioned not only on a protein's pocket, but also on parallely learned latent embeddings of seed molecules. We envision the usage of PoliGenX in hit expansion campaigns. We show that it successfully preserves shape similarity with seed molecules, while still being able to chemically diversify them in a de novo fashion. |
|---|