ИСТИНА

The problem of computational methods for the affinity prediction of protein-protein interactions remains relevant. We have developed a new model for computational estimation affinity based on gradient boosting using the LGBM software package [1]. Model training was performed on 591 crystallographic structures of protein complexes selected from PDBbind [2]. Selection criteria: 1) the resolution of the crystallographic structures must be less than 3 angstroms; 2) Debye-Waller factor for atoms of the crystal structure should not exceed 80 along the median. The complexes are divided into affinity classes according to experimentally known interaction constants. Computationally, the protein-protein interaction in each complex was represented as a matrix of atomic group contacts, according to the typing of heavy atoms in the CHARMM force field [3]. Next, the pairwise distances between the two protein structures of the complex were calculated. Distances were calculated using the MDAnalysis software package [4]. The set of test data was 10% of the entire sample (60 complexes). The AUC of the resulting model was 0.992. Test data metrics in the table below. The obtained predictive score also inversely correlates with the experimentally measured Kd (spearman coefficient = - 0.678, p-value < 0.05). The resulting model works on various protein targets, as well as in a wide range of interaction constants.