Multitarget drug discovery for amyotrophic lateral sclerosis by integrating graph neural networks and computational biophysical approachesOA

OBJECTIVE To construct a computational framework that integrates heterogeneous biomolecular interaction networks and graph neural networks(GNN)for screening of candidate smallmolecule compounds targeting multiple pathogenic nodes in amyotrophic lateral sclerosis(ALS).METHODS Multi-source data was integrated that involved 236001 protein-protein interaction pairs(confidence score≥0.7)from the STRING database,294719 drug-target interaction pairs from DrugBank and BindingDB,and 64180 drug-drug interaction pairs from DrugComb to construct a heteroge⁃neous interaction network containing 510973 edges.A GNN model equipped with a multi-relational message-passing mechanism and orthogonal subspace decomposition was established.Using a set of 114 ALS core pathogenic genes curated from the DisGeNET database as seed nodes,the mean asso⁃ciation score between each small molecule and the disease gene module was calculated across the entire network(screening threshold≥0.5).Molecular docking of top-ranked compounds against the superoxide dismutase 1(SOD1)protein(AlphaFold structure AF-P00441-F1-v6)was performed using AutoDock Vina,followed by 100 ns molecular dynamics simulations with GROningen MAchine for Chemical Simulations(GROMACS).Binding free energies were calculated using the Molecular Mechanics Poisson-Boltzmann Surface Area(MM-PBSA)method to evaluate the conformational stability and binding modes of the complexes.RESULTS A total of 378 high-priority candidate compounds were screened by the GNN model.Molecular docking revealed that the binding free energies of nilo⁃tinib,irinotecan and lonafarnib exhibited were-8.6,-8.4 and-8.3 kcal·mol^(-1)with SOD1,respectively.All the compounds were bound to flexible regions surrounding the active site and exerted allosteric stabilizing effects via hydrogen bond networks involving such key residues as arginine-144 and histidine-64.The 100 ns molecular dynamics simulations demonstrated that the root-mean-square deviation of the nilo⁃tinib-SOD1 complex remained stable within 0.1-0.3 nm throughout the trajectory,whereas the lona⁃farnib-SOD1 and irinotecan-SOD1 complexes underwent conformational rearrangement after approxi⁃mately 40 ns and subsequently entered a new equilibrium state.The radius of gyration of all the three complexes were maintained at approximately 1.45 nm,and the solvent-accessible surface areas were stabilized at approximately 81-84 nm^(2).MM-PBSA calculations yielded total binding free energies of-210.717,-140.039,and-155.303 kJ·mol^(-1)for the complexes of nilotinib,irinotecan,and lonafarnib with SOD1,respectively,with van der Waals interactions identified as the dominant driving force.Subsequent docking validation found that these three compounds also exhibited favorable binding potential against 15 core ALS-associated proteins,including TAR DNA binding protein 43(TDP-43),FUS RNA binding protein(FUS),and chromosome 9 open reading frame 72(C9orf72)(binding energies≤-7.4 kcal·mol^(-1)),confirming their multi-target regulatory properties.CONCLUSION This study establishes a computa⁃tional drug repurposing framework for ALS that integrates heterogeneous network analysis with GNN modeling.Through multi-level computational simulations,nilotinib,irinotecan and lonafarnib are identi⁃fied as lead compounds possessing both allosteric stabilizing effects on SOD1 and multi-target regula⁃tory potential,thereby providing reliable computational evidence and candidate molecules for the devel⁃opment of therapeutic strategies against ALS.