Building Accurate and Interpretable Pipelines for Computational Drug Discovery
Na minha lista:
| Publicado no: | ProQuest Dissertations and Theses (2025) |
|---|---|
| Autor principal: | |
| Publicado em: |
ProQuest Dissertations & Theses
|
| Assuntos: | |
| Acesso em linha: | Citation/Abstract Full Text - PDF |
| Tags: |
Sem tags, seja o primeiro a adicionar uma tag!
|
| Resumo: | Modern small-molecule drug discovery relies on understanding how a protein and potential drug interact and developing methods to screen the best potential drugs from those unlikely to bind the target of interest. As traditional drug-screening methods are time- and resource-intensive, computational methods aim to expedite the process using mathematical models of protein-ligand binding to systematically select the most likely binders for in vitro validation.Much of computational drug discovery focuses on developing and applying scoring functions, which asses the interactions between a small-molecule drug and its protein target’s binding pocket, and assign a numerical score to the complex’s compatibility. Many recent scoring functions use neural networks, which are particularly interesting because they effectively leverage the vast experimental data available today. However, many neural networks can be “black boxes” that obscure the patterns behind their predictions. Additionally, drug discovery pipelines using these scoring functions often overlook biological realities, such as the dynamic nature of proteins and the in vivo modifications that regulate their function.This dissertation presents an alternative to opaque high-throughput methods for computational drug discovery. We present two tools that improve the interpretability of drug discovery pipelines and demonstrate these tools’ utility by applying them to the protein target LARP1. The first tool is an interpretable scoring function that adapts the Contextual Explanation Network (CEN) architecture to protein-ligand binding prediction, providing a biochemical explanation with each affinity prediction. The second tool is a program for training tailored machine-learning models to streamline the analysis of ensemble docking, which better approximates the dynamic nature of proteins than standard single-conformation docking methods. Finally, we use these two tools to conduct a screen of potential chemical probes against the cancer target LARP1, using molecular dynamics (MD) simulations to generate an ensemble of disease-relevant LARP1 conformations for docking. The two tools introduced here illustrate the utility of interpretable methods in drug discovery, especially when used together with simulations that account for protein-target dynamics. |
|---|---|
| ISBN: | 9798314850749 |
| Fonte: | ProQuest Dissertations & Theses Global |