Research On Molecular Property Prediction Methods Based On Small Sample

In recent years,molecular property prediction has become increasingly important in the medical field in the context of novel coronaviruses,and the problem has received extensive attention in the field of deep learning.Most of the current molecular property prediction methods are based on top of a large sample of molecules with known properties.Macroscopically,molecules with known properties are rare in life,especially new ones,and it becomes a challenge to effectively predict the properties of molecules with small sample size;microscopically,the information contained in different molecular representations can vary,so it is not easy to maximize the extraction of molecular information with a small number of samples.Therefore,molecular property prediction based on few-shot learning has gained a wider application prospect.In this paper,three molecular property prediction methods based on few-shot learning are proposed.(1)We propose Few-Shot Graph and SMILES Learning for Molecular Property Prediction(Few-GS)method.In this method,we first sample a batch of tasks for training and use Rdkit to convert them into molecular graphs and molecular sequences respectively,and feed the molecular graphs and molecular sequences into GNN and Mol2 Vec to obtain feature vectors respectively.Then,a self-attentive mechanism is used to assign weights to the molecular map feature vectors and molecular sequence feature vectors,and then connect them to calculate the overall molecular feature vectors.Finally,the molecular feature vectors are fed into the few-shot learning model,while the spatial adaptive module is used in the test set to eliminate the bias between tasks and predict the molecular properties through multiple gradient updates of the conditioning parameters.Experiments on the Tox21 and Sider datasets show that the Few-GS method outperforms various existing graph string-based molecular property prediction methods.(2)We propose Few-Shot Local Graph Learning for Molecular Property Prediction(Few-LG)method.In this method,we first transform the molecular string into a complete molecular map,and then divides a large molecular graph into several local structures,i.e.,functional groups that determine the main properties of molecules,according to the newly defined functional group partitioning rules,and since local structures of different sizes require different levels of extraction to achieve better Since different sizes of local structures need different levels of extraction to achieve better results,a multilayer GNN combined with an attention mechanism is used to calculate the feature vectors of the molecular map,and the molecular representations of each class are summed and averaged to obtain the center vector of that class.Finally,the test samples are fed into a few-shot model to predict molecular properties by the proximity principle.Experiments on the Chembl and TRIANGLES datasets show that the Few-LG method outperforms various existing local graph-based molecular property prediction methods.(3)We propose Few-Shot Graph Learning for Molecular Property Prediction(FewGra)method.In this method,we first transform the molecular sequence into a complete molecular graph,then transforms the molecular graph into a molecular tree,and enhances the message transfer between the molecular tree and the molecular graph to improve the accuracy of molecular feature extraction,and secondly inputs the vectors obtained from the molecular graph and the molecular tree into a size fusion gate to obtain a complete global structure of the molecular representation vector of the complete global structure.Then,it is fed into a few-shot learning model,and the framework parameters are updated by multiple gradient update iterations to strengthen the whole learning model.Experiments on datasets such as Tox21 and Sider show that the Few-Gra method outperforms various existing methods for predicting molecular properties based on global molecular maps..