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

Analysis

Why This Paper Matters

This review addresses a critical bottleneck in AI-driven drug discovery: the need for standardized, computer-readable molecular representations. As the abstract notes, the complexity of capturing all structural and chemical characteristics in a single syntax has led to a proliferation of representations, creating confusion for newcomers. By cataloging the most popular representations—many of which are graph-based—the paper provides a clear entry point for researchers from AI backgrounds who lack cheminformatics expertise.

The paper's significance is underscored by its 586 citations, indicating it has become a foundational reference in the field. It bridges two communities: machine learning practitioners who need to understand chemical input formats, and cheminformaticians who want to apply modern AI techniques. This cross-disciplinary guidance is essential for progress in areas like virtual screening, de novo drug design, and property prediction.

Technical Contributions

• Comprehensive survey of representations: Covers SMILES, InChI, molecular graphs, fingerprints (e.g., Morgan, MACCS), and 3D conformers, explaining their strengths and weaknesses.
• Focus on graph representations: Highlights how molecular graphs naturally encode atoms and bonds, making them suitable for graph neural networks (GNNs), a dominant AI paradigm in drug discovery.
• Practical guidance: Offers recommendations on which representation to use for common tasks (e.g., fingerprint for similarity search, graph for property prediction), reducing trial-and-error for newcomers.
• Application mapping: Links each representation to specific AI-driven tasks (e.g., variational autoencoders for SMILES, GNNs for molecular property prediction), showing how representation choice impacts model performance.

Results

As a review, the paper does not present new experimental results. Its value lies in the synthesis of existing knowledge: it organizes decades of cheminformatics research into a digestible format for AI practitioners. The paper's impact is measured by its citation count (586) and its role in enabling subsequent work—many papers citing it likely used its guidance to select representations for their own AI models.

Significance

This paper has democratized access to chemical representation knowledge for the AI community. By lowering the barrier to entry, it has accelerated the integration of machine learning into drug discovery pipelines. Its practical, non-technical tone makes it accessible to graduate students and industry researchers alike. The focus on graph representations also foreshadowed the rise of GNNs in molecular modeling, which have since become a standard tool. As AI continues to transform drug discovery, this guide remains a key reference for ensuring that representation choices are informed and appropriate.