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

Target-aware molecular generation models hold promise for drug discovery, but it remains unclear whether they genuinely exploit target information or merely resemble the Texas Sharpshooter fallacy by retrospectively rationalizing outputs. To address this, we introduce TarPass, a benchmark comprising a curated dataset of 18 well-studied targets with expert-annotated key interactions and experimentally validated active compounds, enabling fair evaluation of target-aware de novo molecular generation models. We assessed 15 representative models across three paradigms: non-3D, 3D in situ, and optimization-based, considering protein-ligand interactions (PLIs), molecular plausibility, and drug-likeness. Results show that 3D in situ models have a modest average advantage in predicted PLIs. However, many fail to outperform random sampling. Non-3D models, benefiting from broader pretraining, generate more drug-like and synthesizable molecules but exhibit weaker target specificity. Optimization-based methods effectively redirect outputs toward favorable chemical regions for single properties, often at the expense of others, for example by reducing compliance with Lipinski's rules. Integrating these insights, we propose a multi-tier virtual screening workflow for target-aware molecular generation as a post-processing strategy to enrich molecules with improved PLIs and plausibility. Overall, this study highlights the limitations of current models in capturing fine-grained target-specific constraints and provides a standardized framework for future structure-based drug design.