[Paper Review] Structure-based drug discovery with deep learning

Rıza Özçelik, Derek van Tilborg+2 more|arXiv (Cornell University)|Dec 26, 2022

Computational Drug Discovery Methods4 citations

TL;DR

This review synthesizes deep learning approaches for structure-based drug discovery (SBDD), emphasizing protein structure integration to predict drug-target interactions, detect binding sites, and enable de novo molecular design. It highlights geometric deep learning and diffusion models as transformative methods, while identifying data scarcity and model interpretability as key challenges for real-world translation.

ABSTRACT

Artificial intelligence (AI) in the form of deep learning bears promise for drug discovery and chemical biology, $ extit{e.g.}$, to predict protein structure and molecular bioactivity, plan organic synthesis, and design molecules $ extit{de novo}$. While most of the deep learning efforts in drug discovery have focused on ligand-based approaches, structure-based drug discovery has the potential to tackle unsolved challenges, such as affinity prediction for unexplored protein targets, binding-mechanism elucidation, and the rationalization of related chemical kinetic properties. Advances in deep learning methodologies and the availability of accurate predictions for protein tertiary structure advocate for a $ extit{renaissance}$ in structure-based approaches for drug discovery guided by AI. This review summarizes the most prominent algorithmic concepts in structure-based deep learning for drug discovery, and forecasts opportunities, applications, and challenges ahead.

Motivation & Objective

• To review state-of-the-art deep learning methods that integrate protein structural information into drug discovery.
• To identify key challenges in data availability, model generalization, and interpretability for structure-based AI applications.
• To highlight emerging techniques such as geometric deep learning and diffusion models for molecular generation and interaction prediction.
• To assess the current gap between in silico predictions and prospective experimental validation in SBDD.
• To advocate for explainable AI and improved data curation to bridge the gap between computational models and experimental medicinal chemistry.

Proposed method

• Utilizes deep learning architectures such as graph neural networks (GNNs), convolutional neural networks (CNNs), and geometric deep learning to model 3D molecular structures and protein-ligand interactions.
• Applies geometric deep learning to encode roto-translational invariance and symmetry in 3D molecular representations, improving generalization across diverse protein targets.
• Employs generative deep learning, particularly diffusion models, to enable de novo molecular design conditioned on target protein structures.
• Integrates protein tertiary structures from accurate predictors like AlphaFold to enhance binding affinity and interaction prediction.
• Leverages end-to-end learning from raw molecular representations (e.g., SMILES, 3D coordinates) without manual feature engineering.
• Proposes data curation, bias correction, and multi-modal learning to address imbalanced and low-quality 3D datasets in drug-target interaction prediction.

Experimental results

Research questions

• RQ1 How can deep learning effectively model complex 3D protein-ligand interactions for drug-target affinity prediction?
• RQ2 What role do geometric deep learning and symmetry-aware architectures play in improving generalization across diverse protein targets?
• RQ3 To what extent can diffusion-based generative models enable de novo design of bioactive molecules with desired binding modes?
• RQ4 What are the main data-related barriers limiting the performance and generalizability of structure-based deep learning in drug discovery?
• RQ5 How can explainable AI methods enhance trust and facilitate experimental validation of deep learning predictions in SBDD?

Key findings

• Geometric deep learning and symmetry-aware architectures significantly improve the modeling of 3D molecular systems by encoding roto-translational invariance, reducing search space and enhancing generalization.
• Diffusion models have achieved state-of-the-art performance in molecular generation and show strong promise for de novo drug design with structural constraints.
• Despite progress, no deep learning model for structure-based drug discovery has yet been validated in a prospective experimental setting, highlighting a critical gap in real-world application.
• High-quality 3D co-crystallized protein-ligand complexes remain scarce, leading to biased datasets that limit model generalizability and performance on novel targets.
• Current models often under-report non- and weak-binding molecules, exacerbating data imbalance and reducing predictive reliability for selectivity and activity cliff detection.
• Explainable AI techniques are essential to decode black-box predictions and uncover structure-activity relationships, enabling integration with experimental medicinal chemistry workflows.