Generative Models Should at Least Be Able to Design Molecules That Dock Well: A New Benchmark.

Area of Science:

• Computational chemistry
• Drug discovery
• Artificial intelligence in medicine

Background:

• Designing effective drug compounds is crucial for drug discovery.
• Progress measurement is hindered by a lack of realistic benchmarks and high validation costs.
• Computational docking is a key method for assessing molecular binding to proteins.

Purpose of the Study:

• To introduce a novel benchmark for evaluating computational models in drug discovery.
• To address the challenge of measuring progress in designing molecules with specific properties.
• To facilitate the development of models for *de novo* drug design.

Main Methods:

• Proposing a benchmark centered on molecular docking using the SMINA software.
• Generating drug-like molecules predicted to have high binding scores.
• Including simpler scoring functions for broader applicability.
• Releasing the benchmark as an accessible software package.

Main Results:

• Current graph-based generative models demonstrate limitations when trained on realistic datasets.
• These models often fail to propose molecules with high docking scores.
• The benchmark highlights challenges in current *de novo* drug design approaches.

Conclusions:

• The proposed benchmark provides a realistic evaluation for computational drug design models.
• Existing generative models require improvement to effectively design high-affinity drug candidates.
• This work aims to advance automated generation of promising drug candidates.