Protein-Ligand Binding Prediction through Diffusion Models

Date of Award


Document Type


Degree Name

Master of Science in Machine Learning


Machine Learning

First Advisor

Dr. Mohammad Yaqub

Second Advisor

Dr. Martin Takac


Structure-Based Drug Design (SBDD) relies on accurately predicting protein-ligand binding affinity. While diffusion models like DecompDiff show promise, existing methods often lack transparency in utilizing information about protein-ligand bindings and the properties of atoms' edges and bonds. This research investigates a novel approach within SBDD that addresses the lack of transparency in utilizing information about protein-ligand bindings and atom properties. We introduce a weighted sum parameter lambda $\lambda$ within DecompDiff's atom update layers. These layers process information from both the atom's edge features and the atom's bond features. By dynamically adjusting lambda, we explore the optimal weighting for each information source to improve and explain the binding prediction. Six experiments were conducted with different $\lambda$ values. Our findings show that $\lambda$ = 0.4 provides the best performance, assigning a weight of 0.4 to edge features and 1.6 to atom features. Furthermore, the second-best performance came from using only bond information (removing atom edge information entirely). This research offers an explanation and understanding of finding the optimal weighting combination of atoms' information passed to the model. This approach has the potential to significantly enhance the development of highly accurate and robust SBDD models.


Thesis submitted to the Deanship of Graduate and Postdoctoral Studies

In partial fulfilment of the requirements for the M.Sc degree in Machine Learning

Advisors: Mohammad Yaqub, Dr. Martin Takac

Online access available for MBZUAI patrons