Annotation of the results obtained in 2019
We have developed a predictive model based on graph convolutions and message passing neural network of neural network architecture for predicting the binding constants of protein-ligand complexes. In doing so, we used multi-task machine learning, where the target variables are dissociation constant (Kd), inhibition constant (Ki) and half the maximum inhibitory concentration (IC50). Carefully trained on the PDBbind dataset, the model achieves a Pearson correlation coefficient of 0.87 and an RMSE value of 1.05 in pK units, surpassing state-of-ther-art models, including the newly developed Kdeep 3D convolutional neural network model. The developed model, graphDelta, estimates the bioactivity of a molecule by the three-dimensional conformation of the protein-ligand complex, and can be used as a scoring function of the resulting conformations obtained using molecular docking. We used graphDelta for the library of chemical compounds from the Zinc database, filtered using standard filters for the physicochemical properties of the molecules, and selected the most promising molecules for subsequent experimental verification.   The resulting predictive model for assessing the bioactivity of low molecular weight compounds was published in the prestigious journal ACS Omega (Karlov et al graph Delta: MPNN scoring function for the affinity prediction of protein-ligand complexes ACS Omega (2020). The resulting program scripts and libraries were successfully used in the analysis of determinants of binding of human cysteinyl receptors CLTR1, CLTR2, also belonging to the GPCR class - the results of these works were published in the prestigious journal Nature Communications (Gusach A, Luginina A et al., Structural basis of ligand selectivity and disease mutations in cysteinyl leukotriene receptors Nature Communications (2019))

 

Publications

---

Annotation of the results obtained in 2018
Angiotensin II is a regulator of a number of important processes of the cardiovascular system, including blood pressure. It interacts with the first type of renin-angiotensin system receptor (AT1R). Blocking this receptor with drugs based on low molecular weight AT1R ligands, such as losartan, is used to treat cardiovascular diseases. Currently, the ligands used in pharmaceuticals completely block the functional activity of the receptor (antagonists). However, the maximum positive effect could be reached with biased ligands, that selectively block one of the signalling pathways. Thus, finding new high-affinity ligands to AT1R is an urgent and important problem. This project aims to develop a methodology for assessing the binding of low molecular weight compounds to the AT1R receptor by means of structural bioinformatics and machine learning algorithms. To this end, at the first stage, a library of low-molecular compounds was assembled with experimentally established binding constants with AT1R. The resulting library is a basis for the compilation of a training set for solving the problem of classification - so, depending on the value of the compound binding constant in the library, they are divided into positive and negative examples. Then, using molecular modeling methods (homology modeling, structural optimization, molecular docking, and other ICM-Pro software modules), structural models of molecular complexes of low molecular weight compounds and AT1R were obtained, as well as analysis of AT1R interaction interfaces with agonists, antagonists and biased agonists for determining key amino acid residues. Further, for the numerical description of the intermolecular interactions of low-molecular compounds with the receptor, a software library was developed for calculating molecular descriptors. The molecular descriptor consists of four comprehensive SIF, SPLIF, RDF and PHAP3PT3D vectors, thus reflecting the intermolecular interaction between the low molecular weight compound and the receptor in a high-dimensional Euclidean space. At the next stage, a statistical analysis of the distributions of the molecular descriptors obtained, data augmentation, as well as the formulation and solution of the machine learning problem in a given equal space will be carried out. The obtained results and the developed libraries were actively used to study the structural and functional properties of AT1R and other GPCRs. Several abstracts for conferences and two scientific publications (Q1) were prepared and published (https://doi.org/10.1016/j.cell.2018.12.011 and https://doi.org/10.1016/j.sbi. 2019.02.010). Also, some results were published in the media with proper RSF funding information: https://mipt.ru/news/sozdan_tochnyy_vychislitelnyy_metod_dlya_stabilizatsii_lekarstvennykh_misheney , https://mipt.ru/news/vtoroy_kannabinoidnyy_retseptor_vydal_svoyu_strukturu.

 

Publications

---