Hamza Hentabli

PhD

Postdoc

AI Algorithm Development Artificial Intelligence in Medicine Artificial Neural Networks Assistant Professor AutoDock Vina AutoML Bioinformatics C++ CNN Computational Biology Computational chemistry Computational Drug Discovery Deep Learning graph NN High Performance Computing HPC/cluster computing Machine Learning Molecular Docking Molecular Dynamics Simulations Python Python (Flask Small Molecules Structural Bioinformatics Structure-Based Drug Design Teaching and mentorship Transformers Virtual Screening

Peer-reviewed publication(s) 1st author peer-reviewed publication(s) Conference Talks(s)

I was a core contributor to multiple CACHE Challenge campaigns, where I designed and executed large-scale virtual screening workflows under realistic drug discovery constraints. I collected and curated bioactivity data from ChEMBL to build QSAR and IC50 prediction models, which were used to support target-specific prioritization. I ran docking and post-docking analyses and developed graph-based and transformer-based models to distinguish active compounds from decoys. These models were applied to ultra-large chemical libraries, including multi-billion compound spaces, to reduce false positives. Our integrated IC50 prediction and interaction-based triage significantly improved hit selection compared to docking alone in blinded CACHE benchmarks.

I was a lead contributor to the HitTPT project, developing a graph transformer model to distinguish active compounds from decoys in structure-based virtual screening. I introduced a new interaction fingerprint that explicitly encodes protein–ligand contacts, geometric relationships, and chemical context directly from docked complexes. This interaction representation was used across multiple stages of the virtual screening pipeline, including post-docking triage and hit re-ranking. I curated and prepared interaction-level datasets across multiple targets to train and validate the model. The resulting approach consistently achieved top performance, significantly reducing false positives compared to docking and standard ML methods.

I contributed to Project by developing a deep learning graph transformer model for binding free energy prediction from protein–ligand complexes. I designed graph representations that capture atomic interactions, geometry, and chemical context from 3D structures. The model was trained and evaluated on curated benchmark datasets of protein–ligand complexes. I conducted systematic benchmarking against seven established methods, including both machine learning and physics-based approaches. The model demonstrated competitive or improved performance, highlighting the strengths of interaction-aware graph transformers for binding affinity estimation.