Research On Antimicrobial Drug Combination Prediction Algorithms Based On Network Models

Antimicrobial drug combinations can combat antimicrobial resistance.Compared to monotherapy,combination therapy(two or more drugs)is used for refractory,persistent,multidrug-resistant infections and microbial infections of unknown etiology.Antimicrobial drug combinations are categorized into three groups:antibiotic combinations,antibiotic-adjuvant combinations,and adjuvant-adjuvant combinations.According to their efficacy,antimicrobial drug combinations also can be categorized: synergistic drug combinations,antagonistic drug combinations,and additive drug combinations.In microbiology laboratories,antimicrobial drug combinations are identified by checkerboard method.However,the method is both time-consuming and expensive.The development of high-throughput screening and computational method provide opportunities for antimicrobial drug combination prediction.In this paper,we constructed databases focusing on antibiotic combinations and antibiotic-adjuvant combinations by collecting and integrating existing drug combinations.Based on the databases,several network models were proposed to predict antibiotic combinations and to identify antibiotic adjuvants and antibiotic subgroups,and to functionally annotate compounds with unknown functions.The research is described below:(1)Antimicrobial drug combination databaseWith the development of high-throughput screening,the number of antimicrobial drug combinations has increased dramatically.Therefore,we need a comprehensive database for collecting,storing,and integrating these data.In this study,we developed the first database focusing on antibiotic combination(ACDB)and antibiotic adjuvant(AADB)through manual collection and web crawling.Specifically,ACDB includes6,175 antibiotic combinations covering 304 compounds and 460 bacterial strains;AADB includes 3,035 antibiotic-adjuvant combinations covering 83 antibiotics,225 adjuvants and 325 bacterial strains.For each compound,we provided its physicochemical properties,chemical structure(2D/3D),targets and chemogenomic data,etc.,which provide a valuable resource for data-driven computational models.ACDB and AADB are available at www.acdb.plus and www.acdb.plus/AADB,respectively.(2)Prediction of antibiotic combinations based on network pharmacologyFrom the perspective of network pharmacology,when a drug acts on targets,the perturbations will propagate along protein-protein interaction(PPI)network and eventually form a drug-action module.We simulated this process using network propagation algorithm and mapped these drug-action modules to PPI network.Subsequently,network proximity algorithm was used to quantify relations between these drug-action modules.We found that drug combinations with smaller network proximity are usually synergistic.In addition,network proximity also can be used to construct the affinity matrix in graph regularization models for predicting synergistic antibiotic combinations.Compared to existing methods,our model shows better prediction performance and good interpretability.(3)Prediction of antibiotic-adjuvant combinations based on network pharmacologyIn the previous study,we proposed a model for synergistic antibiotic combination prediction.The model requires the targets of antibiotic as input.However,for adjuvants,their targets are usually unknown.Therefore,the model cannot be used to predict antibiotic-adjuvant combinations.To overcome the challenge,we used offtarget metabolomics data to predict targets of adjuvants and improve the above model.Taking minocycline as an example,we first constructed a network using the targets of minocycline and chemogenomic data.The key nodes of network are usually more important.Therefore,we used a variety of network algorithms to screen out 19 key proteins.Gene ontology(GO)analysis revealed that these proteins are associated with lipid synthesis and metabolism.In other words,drugs that interfere with lipid biosynthesis may synergize with minocycline.Therefore,we performed a virtual screening for FDA-approved drugs and identified 10 drugs.Six of the 10 drugs have been shown to enhance the antibacterial activity of minocycline.It confirms the validity of our model.(4)Clustering drug-drug interaction network based on network structurePrevious studies have shown that drugs can be divided into several groups based on certain characteristics(e.g.,chemical structure,mechanism of action,phenotype,etc.)and drug-drug interactions are related to their groups.However,these studies focused on drug characteristics.In this study,we focused on drug-drug interaction(DDI)network itself and proposed a method to measure drug similarity in DDI network.The clustered DDI network showed good monochromaticity.However,accuracy of our method is dependent on the quality of dataset itself.If noise exists in the dataset,it will have a large impact on the clustering results.To overcome this challenge,we collected DDI from multiple species and developed an information fusion algorithm.This algorithm can reduce the effect of noise by integrating multispecies DDI.Compared to clustering with mono-species DDI,clustering with multispecies DDI can obtain more robust clustering results.In addition,we proposed two measures(edge purity and edge mutual information)to evaluate the clustering quality of DDI networks.(5)Clustering DDI network based on multi-source drug informationAs mentioned above,drugs can be categorized by a variety of drug information(e.g.,chemical structure,mechanism of action,phenotype,etc.).Different drug information has both commonalities and specificities.It remains a challenge to integrate multi-source drug information and to obtain high-quality clustering results.To overcome the problem,we first collected multiple drug information(chemical structure,pharmacological information,phenotypic information and ATC codes)and calculated the corresponding drug similarity.Subsequently,we proposed an information fusion algorithm to integrate this drug information.Experimental results show that clustering with multiple drug information can obtain the best clustering results.The ablation experiments show that each of the drug information can improve clustering results.