Structure-Based Generation of 3D Small-Molecule Drugs: Are We There Yet?

Structure-based drug design (SBDD) plays a crucial role in preclinical discovery. Recently, structure-based generative algorithms have been developed to streamline the SBDD process, by generating novel, drug-like molecule designs based on the binding pocket structure of target protein. However, there is no effective metric to evaluate the chemical plausibility of molecules designed by these algorithms, which can limit further applications. In this study, we introduce two new metrics for assessing the chemical plausibility of generated molecules. Along with additional analyses, we demonstrate that these algorithms can generate chemically implausible structures with certain property distributions that differ from those of known drug-like molecules. We also compare results with high-throughput virtual screening hits for three protein targets (the c-SRC kinase, the Smoothened receptor, and the Dopamine D1 receptor). These metrics and analysis methods described here offer valuable tools for model developers and users to assess the chemical plausibility and drug-likeness of generated molecules, ultimately enhancing the use of structure-based generation in drug discovery.