Prediction Of Molecular Coupling Constants And Structure Reconstruction Based On Deep Learning

The Scalar Coupling Constant（SCC）plays a key role in the analysis of the three-dimensional structure of organic matter,but various methods based on density functional theory to accurately calculate the coupling constant are very time-consuming,and the prediction accuracy of traditional machine learning method has not yet reached practical needs.At the same time,the operation process of analyzing the three-dimensional structure of unknown molecules by coupling constants of NMR spectroscopy is still in semi-manual stage,requiring a lot of time for experts in the field of chemistry,which limits the application of high-throughput and high-speed screening of potential drug molecular structures from chemical space.In view of this,this paper studies the method based on deep learning to predict the coupling constant from the molecular structure,and reversely reconstruct the three-dimensional structure of the molecule from the coupling constant and other properties.The specific work is divided into the following three parts:（1）A Graph Embedded Local Attention Encoder（GELAE）is proposed based on the molecular topology,using bond length,bond angle,and dihedral angle to describe the molecular structure,and the improved classification loss function is used to train the model.A 3JHH coupling type is predicted.A series of comparative experiments were carried out using different structural representations,different attention modules and different losses.The results show that compared with the traditional chemical bond vector representation,the rotation and translation invariant structure representation proposed in this paper can improve the prediction accuracy of SCC.By embedding local attention in the graph,it is verified that the mean absolute error（MAE）of the validation set is reduced from 0.1603 Hz to 0.1067 Hz;using the classification-based loss function,the MAE of the predicted SCC can be reduced to 0.0963 HZ.（2）In order to use a network to predict all 8 coupling types at the same time,this paper further represents the molecular point cloud composed of atomic features as a sequence of atomic pair features.The characteristics of each atom pair include the type,charge,position coordinates,coupling type,and distance of the two atoms.The feature sequence of atom pairs is input into the Transformer-like encoder,where the self-attention mechanism can learn the positional relationship and interaction between atom pairs,thereby modeling the coupling effect of atom pairs.For the output,using classification to solve regression problem,and adopting the linear interpolation of the two output neurons with the highest responses as the coupling constant prediction.The experimental results show that this model can predict multiple coupling types with high accuracy,and has good stability and generalization ability.However,the prediction accuracy of 1JHC and 1JHN types still needs to be improved.（3）At the end of this article,we will realize the automatic reconstruction of the three-dimensional structure of molecules based on the deep attention model.The problem is abstracted as the reconstruction of a three-dimensional structure from a two-dimensional incidence matrix.The interaction between N atoms in the molecule can form an N×N symmetric matrix A,where matrix element A_ij is the coupling constant of atom i and atom j.Experiments show that the reconstructed molecular structure shows good consistency with the target molecular structure.