Molecular representations in AI-driven drug discovery: a review and practical guide

Laurianne David, Amol Thakkar, Rocío Mercado, Ola EngkvistJournal of Cheminformatics 2020 · DOI

TL;DR

This review provides a comprehensive overview of molecular representations crucial for AI-driven drug discovery, bridging the gap between chemical structures and computational analysis. It guides researchers in selecting appropriate representations for various AI applications, from small molecule property prediction to macromolecule analysis.

What Problem Does This Paper Address?

The core challenge addressed is the need for effective ways to represent complex molecular structures in a format that computers can understand and process for AI applications in drug discovery. Historically, molecules were visualized and understood by chemists through 2D diagrams. However, for computational analysis, especially with the rise of AI and high-throughput screening, these structures need to be translated into machine-readable formats. This translation is complex because molecules have intricate properties like 3D structure, stereochemistry, and various bonding types, which are difficult to capture comprehensively in a single representation. The paper aims to demystify the vast landscape of molecular representations, offering a practical guide for researchers entering this interdisciplinary field.

Methods

This paper is a review, not an experimental study, so it doesn't introduce novel methods. Instead, it categorizes and explains existing methods for representing molecules. The primary focus is on:

  • Molecular Graphs: The fundamental concept of representing atoms as nodes and bonds as edges. This includes discussions on adjacency matrices and node/edge feature matrices for encoding atom and bond properties.
  • Linear Notations: Compact string-based representations derived from molecular graphs, such as SMILES (Simplified Molecular Input Line Entry System) and InChI (International Chemistry Identifier). These are crucial for efficient storage and processing.
  • Chemical Descriptors: Bit strings or vectors that encode physicochemical, structural, or topological properties rather than the exact structure. Examples include MACCS Keys and hashed fingerprints (like ECFPs).
  • Representations for Chemical Reactions: Methods to encode the transformation of molecules, including reaction SMILES, SMIRKS, and Condensed Reaction Graphs.
  • Representations for Macromolecules: Specialized notations for larger biological molecules like peptides, proteins, and glycans, often combining sequence and structural information (e.g., HELM, WURCS).
  • Graphical Representations: 2D and 3D visualizations used for direct interpretation and analysis.

The paper emphasizes that the choice of representation is task-dependent and often involves trade-offs between compactness, expressiveness, and computational cost.

Datasets

Not discussed in the paper.

Key Insights and Ideas

  • No Single "Best" Representation: The paper strongly emphasizes that the optimal molecular representation is highly context-dependent, varying with the specific AI task (e.g., property prediction, de novo design, reaction prediction) and the type of molecule (small molecule, macromolecule).
  • Graph Theory as the Foundation: Molecular graphs form the conceptual bedrock for many modern representations, providing a structured way to encode connectivity.
  • The Trade-off Between Compactness and Information: Linear notations like SMILES offer excellent compactness and ease of use but might lose some nuanced structural information compared to more detailed graph-based or descriptor-based methods.
  • Evolution of Notations: The paper traces the historical development of chemical notations, from early line formulas to sophisticated, machine-readable formats, driven by advancements in computing power and the increasing complexity of chemical data.
  • AI's Growing Reliance on Representations: The review highlights how AI, particularly deep learning, is increasingly leveraging these molecular representations for tasks like property prediction, molecule generation, and reaction outcome prediction.

Key Results

As a review paper, it doesn't present experimental results. Instead, it summarizes the state-of-the-art in molecular representations for AI in drug discovery. Key findings include:

  • Ubiquity of Graph Representations: Molecular graphs and their matrix-based encodings are foundational for many AI applications, especially with the rise of graph neural networks.
  • Dominance of SMILES: SMILES has emerged as a highly popular and versatile linear notation due to its balance of compactness and expressiveness, though its canonicalization can be complex.
  • Importance of InChI: The InChI and its hashed counterpart, InChIKey, are highlighted as open-source, standardized identifiers crucial for data sharing and searching.
  • Emergence of Macromolecule Representations: Specialized notations are essential for handling the complexity of biopolymers and other macromolecules, bridging bioinformatics and cheminformatics.
  • AI's Impact on Representation Choice: The paper notes that AI is not only using these representations but also influencing their development and application, for instance, through representation learning techniques.

Limitations and Future Work

  • Limitations of Graph Representations: The paper acknowledges that graph models struggle with certain chemical structures, such as those with delocalized bonds (e.g., organometallic compounds) or highly dynamic bonding.
  • Challenges with Specific Notations: Some notations have limitations, like SMILES not inherently encoding stereochemistry (though isomeric SMILES addresses this) or the complexity of certain reaction notations.
  • Data Availability and Standardization: While not explicitly a limitation of representations themselves, the paper implies that the effectiveness of AI applications is tied to the availability of standardized, high-quality datasets encoded in these representations.
  • Future Directions: The paper suggests that the ongoing development of AI will continue to drive innovation in molecular representations, with a focus on more efficient, expressive, and universally applicable formats, particularly for complex molecules and reactions. The integration of different representation types and the development of robust methods for handling stereochemistry and dynamic molecular behavior are likely areas for future research.

Summary updated

Abstract

The technological advances of the past century, marked by the computer revolution and the advent of high-throughput screening technologies in drug discovery, opened the path to the computational analysis and visualization of bioactive molecules. For this purpose, it became necessary to represent molecules in a syntax that would be readable by computers and understandable by scientists of various fields. A large number of chemical representations have been developed over the years, their numerosity being due to the fast development of computers and the complexity of producing a representation that encompasses all structural and chemical characteristics. We present here some of the most popular electronic molecular and macromolecular representations used in drug discovery, many of which are based on graph representations. Furthermore, we describe applications of these representations in AI-driven drug discovery. Our aim is to provide a brief guide on structural representations that are essential to the practice of AI in drug discovery. This review serves as a guide for researchers who have little experience with the handling of chemical representations and plan to work on applications at the interface of these fields.
BibTeX
@article{david2020molecular,
  author = {Laurianne David and Amol Thakkar and Rocío Mercado and Ola Engkvist},
  title = {{Molecular representations in AI-driven drug discovery: a review and practical guide}},
  journal = {Journal of Cheminformatics},
  year = {2020},
  doi = {10.1186/s13321-020-00460-5},
}
Figure