Optimization And Application Of Computer-Aided Drug Design Methods

Computer-aided drug design is one of the key techniques in drug discovery and development,and has promoted the approval and marketing of many drugs.The explosive biomedical data and fast-growing artificial intelligence technologies inspired its new vigor and vitality.However,the theory of some existing methods is not as perfect as we suppose.For example,the scoring functions of docking programs could not evaluate the binding strength of protein-ligand complexes;2D/3D fingerprint-based similarity searching is inefficient in recalling active molecules with novel scaffolds;the accuracy of ADMET evaluation is limited by data deficient and outdated algorithms,etc.Therefore,how to improve the success rate by optimizing the classical computer-aided drug design methods is an important topic in drug discovery.In this thesis,we conducted optimization research towards several key methods and focused on the application of specific drug discovery practices and the developments of cheminformatic platforms.In addition,considering that the management of drug-drug interactions is one of the issues in clinical medication,we developed a professional drug interaction knowledge system named DDInter to assist clinical decision-making and improve patient safety.The main contents and results of this thesis are as follows:（1）We proposed a new method named“energy auxiliary terms learning”based on the DUD-E database and 10 scoring functions of popular docking programs MOE,GOLD and Schrodinger,in which the scoring components are extracted and used as the representation of protein-ligand interactions to train machine learning classification models.Our method of different levels not only outperformed classical scoring functions for the absolute performance and initial enrichment but also yielded comparable performance compared with other advanced machine learning-based methods.We also found that complementarity existed between different scoring functions and docking programs since the consensus score of multiple scoring functions could make up shortcomings and achieve higher screening performance.（2）Here,we reported the computational bioactivity fingerprint（CBFP）based on 832 protein targets for easier scaffold hopping,where the predicted activities in multiple quantitative structure-activity relationship models are integrated to characterize the biological space of a molecule.The scaffold hopping abilities of CBFP representation and nine other chemical descriptors had been systematically benchmarked in 17 different query datasets paired with five screening datasets by ranking the databases in decreasing order of Tanimoto similarity values with the query molecule.The CBFP representation could highly discriminate structural descriptors in terms of its efficacy for improving the rankings of molecules with novel scaffolds.In the prospective validation for the discovery of novel inhibitors of PARP1,35 predicted compounds with diverse structures are tested,25of which show detectable growth-inhibitory activity;beyond this,the most potent（compound 6）has an IC₅₀ of 0.263 n M.These results support the use of CBFP representation as the bioactivity proxy of molecules to explore uncharted chemical space and discover novel compounds.（3）We firstly conducted a comprehensive data retrieval by using different ADMET-related keywords and obtained a high-quality dataset collection of 0.25M entries spanning 53 ADMET-related endpoints.We employed the multi-task graph attention（MGA）framework to develop the classification and regression predictors simultaneously.The MGA models yield better performance relative to the corresponding XGBoost models,and enabled one input with multiple outputs,thus greatly simplifying the calculation process.Finally,we developed a systematic online platform named ADMETlab 2.0（https://admetmesh.scbdd.com/）by integrating these models、physicochemical properties、medicinal chemistry rules、and toxicophore rules.ADMETlab 2.0 is believed to have greater capacity to assist medicinal chemists in accelerating the drug research and development process.（4）In this work,we proposed DDInter to help physicians and pharmacists detect inappropriate medications and manage clinical outcomes.The manually curated platform contains about 0.24M DDI associations connecting 1833 approved drugs by bringing together the DDI information scattering in literature and product labels.It supplies detailed and professional information about each DDI association,including severity level,mechanism description,management of concurrence,alternative medications,etc.Additionally,it integrates various functions including data browsing,retrieval,and interaction checker to support clinical decision-making.Some data visualization tools are embedded to help users to understand and explore the search results.It can also be used for informatics-based DDI investigation and evaluation of other prediction frameworks.We hope that DDInter will prove useful in improving clinical decision-making and patient safety.DDInter is freely available,without registration,at http://ddinter.scbdd.com/.