Quality Matters: Deep Learning-Based Analysis of Protein-Ligand Interactions with Focus on Avoiding Bias

• DOI: 10.1101/2023.11.13.566916
• Publication Date: 2023-11-17T02:00:20Z
• Authors: Manuel S. Sellner; Markus A. Lill; Martin Smieško
• Keywords: Protein ligand; Affinities

Abstract The efficient and accurate prediction of protein-ligand binding affinities is an extremely appealing yet still unresolved goal in computational pharmacy. In recent years, many scientists have taken advantage the remarkable progress deep learning applied it to address this issue. Despite all advances field, there increasing evidence that typically validation these methods not suitable for medicinal chemistry applications. This work assesses importance dataset quality proper splitting techniques demonstrated on example PDBbind dataset. We also introduce a new tool analysis complexes, called po-sco. Po-sco allows extraction interaction information with much higher detail comprehensibility than tools available date. trained transformer-based model generate fingerprints can be utilized downstream predictions, such as affinity. When using po-sco, generated predictions were superior those based commonly used PLIP ProLIF tools. demonstrate more important number data points suboptimal lead significant overestimation performance.