Structure-based drug discovery with deep learning

Riza Özçelik, Derek van Tilborg, José Jiménez-Luna, Francesca GrisoniarXiv 2022 · OpenReview

TL;DR

This review explores the burgeoning field of structure-based drug discovery (SBDD) powered by deep learning (DL). It highlights how DL is revolutionizing SBDD by enabling more accurate predictions of drug-target interactions, identification of binding sites, and de novo design of novel drug candidates, moving beyond traditional ligand-based approaches.

What Problem Does This Paper Address?

Traditional drug discovery is a lengthy, expensive, and often inefficient process. While ligand-based drug discovery, which relies on the properties of known active molecules, has seen significant advancements with deep learning, structure-based drug discovery (SBDD) has lagged behind. SBDD leverages the three-dimensional structure of target proteins, offering the potential to address critical challenges like predicting affinity for novel targets, understanding binding mechanisms, and rationalizing drug behavior. However, the complexity of protein structures and the scarcity of high-quality 3D structural data have historically made SBDD more challenging. This paper addresses the need for a comprehensive overview of how deep learning is now enabling a renaissance in SBDD, overcoming these limitations and unlocking new possibilities.

Methods

The paper categorizes deep learning approaches in SBDD based on the protein representation used:

  • Sequence-based: Utilizes the primary amino acid sequence. This often involves recurrent neural networks (RNNs) or transformer architectures to capture sequential dependencies.
  • 3D Structure-based: Leverages the full three-dimensional atomic coordinates of the protein. This can involve:
  • Voxelized representations: Treating the protein structure as a 3D grid, suitable for convolutional neural networks (CNNs).
  • Graph-based representations: Representing atoms as nodes and bonds/interactions as edges, processed by graph neural networks (GNNs). This approach is particularly adept at capturing local atomic environments and relationships.
  • Surface-based representations: Focusing on the protein's solvent-accessible surface, often using geometric deep learning techniques.

The paper also discusses various neural network architectures commonly employed, including RNNs, CNNs, and GNNs, and how they are adapted to process these different protein representations.

Datasets

The paper references several key datasets crucial for SBDD research:

  • Protein Data Bank (PDB): A foundational repository for structural data of biological macromolecules.
  • scPDB: Contains druggable binding sites and ligands extracted from the PDB.
  • BioLip: A curated database of ligand-protein interactions.
  • PDBbind: A comprehensive database of protein-ligand complexes with associated binding affinities.
  • AlphaFold Protein Structure Database & AlphaFill: These provide predicted protein structures and ligand/cofactor transplants, significantly expanding the availability of structural data.
  • Binding MOAD: A database of X-ray crystal structures with bound ligands and experimental binding affinities.
  • Directory of Useful Decoys (DUD-E): Used for benchmarking molecular docking programs.
  • BindingDB: A collection of protein-ligand binding affinities.
  • BigBind: Links protein structures to ChEMBL assay data.
  • KIBA & Davis: Datasets focused on bioactivity measurements against kinases.

The paper emphasizes that while sequence and experimental assay data are abundant, high-quality 3D structural data remains a bottleneck, a gap being addressed by advancements like AlphaFold.

Key Insights and Ideas

  • Renaissance in SBDD: Advances in deep learning, coupled with breakthroughs in protein structure prediction (e.g., AlphaFold), are driving a significant resurgence in structure-based drug discovery.
  • Beyond Ligand-Based: SBDD, powered by DL, offers a higher potential to tackle currently unsolved challenges in drug discovery, such as predicting affinity for novel targets and understanding complex binding mechanisms, which ligand-based approaches struggle with.
  • Protein Representation is Key: The choice of how to represent protein information (sequence, 3D structure, or surface) profoundly impacts the suitability of deep learning architectures and the type of chemical information captured.
  • Geometric Deep Learning's Promise: Geometric deep learning, which leverages symmetry and topology, is poised to further enhance DL's ability to model complex molecular systems and interactions, especially in 3D.
  • Data Scarcity Remains a Challenge: Despite progress, the availability of high-quality 3D structural data, especially for disordered or membrane proteins, and data on non-binding or weakly binding molecules, remains a significant limitation.

Key Results

The paper doesn't present specific quantitative results from a single study but rather synthesizes findings across numerous research efforts. It highlights that deep learning approaches have demonstrated success in:

  • Drug-Target Interaction (DTI) Prediction: Accurately predicting binding affinity and interactions between proteins and ligands.
  • Binding Site Detection: Identifying "druggable" cavities within protein structures.
  • De Novo Design: Generating novel molecular structures with desired properties, conditioned on a target protein's structure.

While specific performance metrics are not detailed in this review, the paper indicates that these DL-based SBDD methods are increasingly competitive with, and in some cases outperform, traditional computational approaches.

Limitations and Future Work

  • Data Limitations: The scarcity of high-quality 3D structural data, especially for challenging protein targets, and the under-reporting of non-binding molecules are significant limitations. This often leads to biased datasets and can hinder model generalizability.
  • Model Interpretability: The "black-box" nature of some deep learning models can be a barrier to experimentalist adoption. Future work should focus on developing more explainable AI (XAI) methods to understand underlying structure-activity relationships and binding modes.
  • Prospective Validation: A notable gap is the lack of prospective validation for many SBDD deep learning approaches. Future work should prioritize experimental validation to build trust and demonstrate real-world applicability.
  • Dynamic Information: Current SBDD algorithms are often agnostic to protein conformational flexibility and dynamics, which are crucial for drug binding. Incorporating dynamic information into DL models is a key area for future research.
  • Bridging Data Modalities: Effectively integrating diverse data types (e.g., sequence, structural, assay data) remains an active research area.

Summary updated

Abstract

Artificial intelligence (AI) in the form of deep learning bears promise for drug discovery and chemical biology, $\textit{e.g.}$, to predict protein structure and molecular bioactivity, plan organic synthesis, and design molecules $\textit{de novo}$. 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 $\textit{renaissance}$ 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.
BibTeX
@misc{ozcelik2022structure,
  author = {Riza Özçelik and Derek van Tilborg and José Jiménez-Luna and Francesca Grisoni},
  title = {{Structure-based drug discovery with deep learning}},
  year = {2022},
  eprint = {2212.13295},
  archivePrefix = {arXiv},
  url = {https://arxiv.org/abs/2212.13295},
}
Figure