Construction Of A Deep Learning-based Scoring Function For Reverse Docking And A Comprehensive Pesticide-likeness Prediction Web Platform

Pesticides are critical to food security.However,with the emergence of pesticide resistance and people’s attention to pesticide safety,the demand to develop new green pesticides is becoming more urgent.However,the discovery of new pesticide molecules often relies on the chemical synthesis of a large number of compounds and the determination based on biological experiments,which usually consumes a huge workforce and material resources,and lengthen the research and development cycle of pesticides.This traditional research and development of pesticide is inefficient and slow.But computer-aided drug design methods have been proven to be a means to facilitate drug discovery.How to improve the efficiency of modern pesticide discovery by utilizing computer-aided drug design methods is an open problem.The discovery of new pesticide targets is an important link in modern pesticide discovery.In computer-aided drug design,reverse docking is a method to find protein targets that may interact with specific small molecules.Therefore,it can be used to explore new targets of existing pesticides,explain the molecular mechanism of pesticides,and assist in the study of the toxic mechanism of pesticides.However,scoring functions of molecular docking software limit the performance of reverse docking methods:conventional scoring functions may not accurately identify the impact of changes in protein residue structure on binding energy.Therefore,we intend to design a new deeplearning-based scoring function for reverse docking.We first collected protein-ligand complex structures with experimentally determined binding energy data from the PDBbind database and computed molecular descriptors for them by nine molecular descriptor software.In addition to the parameters of small molecule compounds,the computed descriptors also include a large number of parameters related to protein structure,which helps to describe the structural characteristics of proteins in more detail.Next,we built a neural network model using the TensorFlow.framework.By repeatedly tuning to find the optimal hyperparameters and training the model,we finally determined the appropriate model parameters and trained to obtain a usable model.Evaluated by the popular CASF2016 benchmark in the field of scoring function research,our neural network model achieves a Pearson correlation coefficient R=0.786 and a standard deviation SD=1.34 for scoring ability.And a Spearman correlation coefficient SP=0.681,the Kendall correlation coefficient τ=0.586,and the prediction index PI=0.713 for ranking ability.These indicators all prove that our model has a strong predictive ability.Evaluating the pesticide-likeness of molecules is an effective method to understand the laws of pesticide molecules.Pesticide-likeness is a means of estimating the probability of compound molecules to become pesticide molecules based on easily quantifiable characteristics of molecules.Therefore,pesticide-likeness may play an important role in the preliminary screening of a large number of compound structures and the optimization of lead compounds.At present,many different tools and algorithms can be used for pesticide-likeness prediction.However,these numerous pesticide-likeness prediction tools and algorithms from different researchers still lack systematic aggregation and sorting,which is difficult for general users to use multiple tools to perform a comprehensive analysis of the pesticide-likeness on small molecules.Therefore,we collected and organized various pesticide-likeness prediction algorithms and tools reported in the literature.More than ten peer-reviewed algorithms and tools for pesticide-like property prediction were collected,covering characteristics on basic physicochemical properties,molecular flexibility,lipophilicity,toxicity prediction,pesticide-likeness scoring,and many other property aspects.Then we programmatically integrated them into a set of workflows that can be automated to run independently on Linux servers.Next,we used modern Web technology to build it into a one-stop pesticide-likeness prediction platform named CoPLA.This platform has been deployed on the high-performance computing platform of our research team and is publicly available to all researchers(http://chemyang.ccnu.edu.cn/ccb/server/CoPLA/).This platform is currently the first online platform that provides users with pesticide-likeness predictions for free.In conclusion,we try to improve the efficiency of pesticide research and development with computer-aided drug design methods through these two research works.The new deep-learning-based scoring function for reverse docking has a better predictive ability and is expected to be applied to various reverse docking research in the future.We hope that this scoring function can provide new possibilities for researchers to use the reverse docking method to study the molecular mechanism and find new molecular targets.The one-stop pesticide-likeness prediction platform can reduce the difficulty of using pesticidelikeness prediction tools,thereby promoting the widespread use of pesticide-likeness prediction methods in modern pesticide research and development.In general,we hope that these two works can solve the difficulties in practical applications for researchers in related fields and inspire algorithm and tool researchers in these fields to develop newer and better tools to help the research and development of modern new pesticides jointly.