RediscMol: Benchmarking Molecular Generation Models in Biological Properties.

Area of Science:

• Computational chemistry
• Drug discovery
• Machine learning

Background:

• Deep learning models are increasingly used for generating novel molecules with specific properties.
• Current evaluation metrics for these models are insufficient, especially for biological applications.
• A need exists for robust benchmarks to assess generative models in realistic drug design contexts.

Purpose of the Study:

• To introduce the RediscMol benchmark for evaluating molecular generative models.
• To assess the performance of eight representative generative models using novel metrics.
• To provide a framework for advancing generative models in drug discovery.

Main Methods:

• Constructed the RediscMol benchmark using active molecules from kinase and GPCR datasets.
• Introduced rediscovery- and similarity-based metrics for model evaluation.
• Evaluated eight generative models: CharRNN, VAE, Reinvent, AAE, ORGAN, RNNAttn, TransVAE, and GraphAF.

Main Results:

• Performance rankings differ from previous evaluations based on standard metrics.
• CharRNN, VAE, and Reinvent demonstrated superior ability in reproducing known active molecules.
• RNNAttn, TransVAE, and GraphAF showed limitations in reproducing known actives, despite strong distribution-learning performance.

Conclusions:

• The RediscMol benchmark provides a more biologically relevant evaluation of generative models.
• Specific models show varying strengths in rediscovering known active compounds.
• The proposed evaluation framework can guide the development of generative models for practical drug design.