Machine learning-driven hit discovery, validation, and phytochemical recipe enrichment
File(s)
Author(s)
Rita, Luís
Type
Thesis
Abstract
Bioactive phytochemicals show promise for disease prevention and treatment, but their interactions with drugs and disease pathways remain poorly understood. We address this problem by employing machine learning to identify therapeutic phytochemicals, predict their putative pharmacodynamic (PD) interactions with drugs, and discover phytochemically enriched recipes by using existing drugs as a reference. First, we developed a network-based machine learning approach to identify bioactive phytochemicals in extra virgin olive oil (EVOO) that could influence the protein network associated with Alzheimer's disease (AD). The model achieved a balanced classification accuracy of 70.3 ± 2.6%, effectively distinguishing late-stage experimental AD drugs from other clinically approved drugs. The EVOO analysis highlighted phytochemicals such as quercetin, genistein, and luteolin with high predicted activity against AD, providing novel insights into potential mechanisms for prevention or treatment. Second, we utilised a novel network machine learning approach based on the DreamLab DRUGS network propagation method to predict synergies in PD drug-drug and phytochemical-drug interactions. The model was calibrated using an in vitro PD drug-drug interactions dataset, achieving a Pearson correlation score of 68.9%. Experimental validation was conducted in cancer cell lines. The network machine learning model demonstrated consistency with predicted outcomes in 7/11 tested drug-drug combinations and 5/9 food-drug combinations, indicating its effectiveness in identifying clinically relevant PD interactions. Third, we applied large language models (LLMs) to optimize ingredient substitutions in recipes, aiming to enhance phytochemical content. The application of LLMs in ingredient substitution improved Hit@1 accuracy from a baseline of 34.5 ± 0.1% to 38.0 ± 0.3% on the original GISMo dataset, and from 40.2 ± 0.4% to 54.5 ± 0.3% on a refined dataset. This work demonstrates that machine learning can significantly advance the identification of bioactive phytochemicals, predict their interactions with drugs and disease processes, and optimize dietary choices. These findings establish a computational foundation for integrating nutrition and pharmacology, though clinical validation remains essential.
Version
Open Access
Date Issued
2024-10-15
Date Awarded
01/10/2025
License URL
Advisor
Veselkov, Kirill
Bronstein, Michael
Sponsor
Fundacao para a Ciencia e a Tecnologia
Grant Number
2021.05460.BD
Publisher Department
Department of Surgery & Cancer
Publisher Institution
Imperial College London
Qualification Level
Doctoral
Qualification Name
Doctor of Philosophy (PhD)