Computational Analysis Of Molecular Properties By Modeling Its 3D Topology

Molecules are the fundamental units of matter in the natural world.The determination of their properties plays a crucial role in drug discovery.The rapid and efficient assessment of molecular properties aids in the swift identification of effective drug molecules.This article focuses on the modeling and analysis of the 3D conformational topology of molecules to investigate the properties of both small and large molecules.The main research objectives are as follows:The first objective is to employ the MT-ToxGNN,a combination of message passing networks and multitask deep networks,to rapidly and accurately assess the quantitative toxicity of small molecules.This approach decomposes the 3D topological structure of small molecules into two sub-topological structures:Atom-Bond Graph(AB-G)and Atom-Noncovalent Graph(AN-G).In AB-G,atoms are represented as nodes and covalent bonds as edges,while in AN-G,atoms are also nodes,but noncovalent interactions are represented as edges.These two sub-topological structures are then processed separately by a bipartite graph neural network to predict multiple toxicities of the molecules.The next objective involves using vdWGraph,based on the message passing network of ToxGNN,to predict multiple properties of small molecules.This method addresses the limitation of ToxGNN requiring detailed 3D structural information by leveraging large-scale pre-training.It directly predicts molecular properties from the coarse-grained 2D planar structural information of molecules.Furthermore,vdWGraph extends the single toxicity prediction of ToxGNN to ten different properties and achieves competitive results in the prediction of multiple properties.Lastly,the deep model DGCddG,based on graph convolutional networks,is utilized to predict the free energy change of protein macromolecules under mutation.This method expands the scope from studying small molecule properties to protein macromolecule properties.The protein complex interface is abstracted as a virtual molecule,where residues are defined as nodes and residue-residue interactions(contacts)are defined as edges.DGCddG employs a multi-layer graph convolutional network to extract contextual information for prediction purposes.The methods proposed in this article provide a more biologically meaningful modeling of molecular structures from a 3D perspective.They also leverage the excellent capabilities of graph neural networks in handling topological structures.The promising experimental results further validate the effectiveness of the approach for molecular structure modeling.Overall,this research provides important insights for drug discovery and offers valuable avenues for future exploration.