Abstract
Three-dimensional descriptors are often used to search for new biologically active compounds, in both ligand- and structure-based approaches, capturing the spatial orientation of molecules. They frequently constitute an input for machine learning-based predictions of compound activity or quantitative structure–activity relationship modeling; however, the distribution of their values and the accuracy of depicting compound orientations might have an impact on the power of the obtained predictive models. In this study, we analyzed the distribution of three-dimensional descriptors calculated for docking poses of active and inactive compounds for all aminergic G protein-coupled receptors with available crystal structures, focusing on the variation in conformations for different receptors and crystals. We demonstrated that the consistency in compound orientation in the binding site is rather not correlated with the affinity itself, but is more influenced by other factors, such as the number of rotatable bonds and crystal structure used for docking studies. The visualizations of the descriptors distributions were prepared and made available online at http://chem.gmum.net/vischem_stability , which enables the investigation of chemical structures referring to particular data points depicted in the figures. Moreover, the performed analysis can assist in choosing crystal structure for docking studies, helping in selection of conditions providing the best discrimination between active and inactive compounds in machine learning-based experiments.
Introduction
Numerical descriptors are fundamental to computer-aided drug design, serving as the primary input for machine learning (ML) models. While 2D descriptors (such as molecular weight or atom counts) are invariant, 3D descriptors capture essential spatial information regarding a molecule's orientation within a receptor's binding site. However, the inherent flexibility of ligands and the variety of available protein crystal structures introduce significant variance in docking poses. If the 3D descriptors calculated from these poses vary wildly, the resulting ML models become "noisy" and less predictive. This study investigates how this "stability" of 3D descriptors fluctuates across the aminergic G protein-coupled receptor (GPCR) family and its subsequent impact on predictive modeling.
Methodology
The researchers developed a systematic pipeline to evaluate descriptor sensitivity across 14 aminergic GPCR targets, including serotonin (5-HT), dopamine (D), and adrenergic receptors. Ligands were categorized as actives ($K_i < 100\text{ nM}$) or inactives ($K_i > 1000\text{ nM}$) based on ChEMBL data. Using the Glide docking engine, up to 10 poses were generated per compound. For each pose, 214 3D descriptors were calculated using the Mordred library, covering categories like CPSA, geometrical indices, and MoRSE descriptors. The stability of these descriptors was quantified by calculating the standard deviation across poses. High-dimensional descriptor spaces were then visualized using LargeVis to assess the spatial separation of active and inactive compounds.
Results and Discussion
A key finding of the study is that biological affinity does not necessarily correlate with docking stability. In several receptors, such as 5-HT$_{2C}$, inactive compounds exhibited higher pose consistency than actives. Instead, descriptor variance was strongly driven by molecular flexibility; a high correlation (Pearson coefficient up to 0.718) was found between the number of rotatable bonds and descriptor fluctuations. More flexible molecules lead to "fuzzier" data for ML models.
Furthermore, the choice of crystal structure significantly influenced descriptor distributions. Certain protein conformations provided superior discrimination between active and inactive compounds in the descriptor space. The robustness of specific descriptor types also varied; 3D-MoRSE descriptors proved highly stable at short distances, whereas geometrical shape indices were prone to high fluctuations, making them potentially riskier for ML training.
The study also highlighted that increasing the number of docking poses considered (from 3 to 10) led to higher average standard deviations in descriptors. This suggests that ML models relying solely on the "top-1" pose may overlook significant underlying data instability. To aid researchers, the authors launched an interactive visualization tool (chem.gmum.net/vischem_stability) that allows for manual inspection of how specific chemical structures map to data points in the descriptor space.
Conclusion
This research underscores the importance of considering docking pose and crystal structure variability when utilizing 3D descriptors for ML-based drug discovery. By identifying stable descriptors and optimal crystal structures, researchers can improve the signal-to-noise ratio in their predictive models. While currently focused on aminergic GPCRs and the Glide docking engine, this methodology provides a framework for a "pre-flight check" in structure-based ML experiments across other protein families, helping to select conditions that provide the best discrimination between active and inactive compounds.
Summary updated