Abstract: Protein-ligand interactions play a crucial role in drug development. A ligand docks in the pocket area of a protein to form a docked complex with minimum free energy. Protein-ligand binding pose prediction is to determine the docked position of the ligand within the pocket. Pose prediction is computationally very hard as it involves exploring an astronomically large number of three-dimensional geometric transformations and minimising an unknown free energy equation for the docked complex. Pose prediction has progressed significantly with recent graph neural network (GNN) methods. However, they still struggle to achieve accuracy levels that could be in practical use in the virtual screening of potential drug ligands. This thesis introduces KTransPose to improve pose prediction accuracy by using two strategies: a transformation-invariant loss and an iterative refinement technique. The proposed strategies help various GNN-based pose prediction methods improve their performance. Compared to the state-of-the-art pose prediction methods, KTransPose achieves superior Root Mean Squared Distance (RMSD) values of 3.30˚A on well-known benchmark datasets such as PDBbind-2020 and CASF-2016. The implementation and codebase for this study are available at https://github.com/alam13/KTransPose.git.

Student: Md. Khorshed Alam (Std No. 0421052080)

Supervisor: Dr. Mohammed Eunus Ali

Date and Time: 19-Apr-2025 (3:00 PM - 3:45 PM)

Venue: Graduate Seminar Room