Revisiting Target-Aware de novo Molecular Generation with TarPass: Between Rational Design and Texas Sharpshooter.

Area of Science:

• Computational chemistry
• Drug discovery
• Artificial intelligence in medicine

Background:

• Target-aware molecular generation models aim to accelerate drug discovery by designing novel compounds.
• The effectiveness of these models in truly utilizing target information versus the Texas Sharpshooter fallacy is debated.
• A standardized benchmark is needed to rigorously evaluate model performance.

Purpose of the Study:

• To introduce TarPass, a benchmark for evaluating target-aware de novo molecular generation models.
• To assess the performance of 15 representative models across different paradigms.
• To identify limitations in current models' ability to capture target-specific constraints.

Main Methods:

• Developed TarPass benchmark with 18 targets, annotated interactions, and validated active compounds.
• Evaluated 15 models (non-3D, 3D in situ, optimization-based) on protein-ligand interactions (PLIs), plausibility, and drug-likeness.
• Proposed a multi-tier virtual screening workflow for post-processing generated molecules.

Main Results:

• 3D in situ models showed a modest advantage in predicted PLIs but often performed similarly to random sampling.
• Non-3D models generated more drug-like and synthesizable molecules but lacked target specificity.
• Optimization-based methods improved single properties but often compromised others, like Lipinski's rules compliance.

Conclusions:

• Current target-aware molecular generation models have limitations in capturing fine-grained target-specific constraints.
• The TarPass benchmark provides a standardized framework for future structure-based drug design model development.
• A multi-tier virtual screening workflow can enhance molecule enrichment for improved PLIs and plausibility.