Many studies of antibiotic combinations have been reported in PubMed. To obtain high-quality antibiotic combinations, we manually collected literature-reported antibiotic combinations from thousands of studies. Specifically, the keywords “antibiotic (s),” “fractional inhibitory concentration,” “synerg*,” and “combination (s)” were used to retrieve related literature in PubMed, and we finally obtained 6,175 antibiotic combinations from 6,121 publications (Dec. 2021).

We then filtered out duplicate content and 304 unique compounds were obtained. We normalized them into a standard format (PubChem CID) because they are well-known and easily linked to external databases. Simplified Molecular Input Line Entry System (SMILES) of each compound was obtained from PubChem, and we used SMILES to calculate descriptors (e.g., LogP, molecular weight, Lipinski’s five rules, Table 1). Moreover, we provided an additional Python script that converts SMILES into Protein Data Bank (PDB) format for download and further studies (molecular docking and/or molecular dynamics simulations). Targets of each compound were collected from DrugBank (Wishart et al., 2007).

ACDB is a relational database. It is hosted on a cloud server (CentOS Linux release 8.5.2111), which employs Apache (version 2.4.37) and MySQL (version 8.1) as the web and database server, respectively. The website is built with PHP, HTML and CSS and it can be freely accessed at http://www.acdb.plus.

Since experimental approaches for distinguishing antibiotic combinations are expensive and time-consuming, an increasing number of researchers are using machine learning methods to predict potential antibiotic combinations (Chandrasekaran et al., 2016; Mason et al., 2017; Wu et al., 2021). To make it easier for users, ACDB incorporated a series of features including Lipinski’s five rules, molecular fingerprints (e.g., MACCS Keys, Morgan fingerprint) and chemogenomic data. These features have been successfully applied in previous work to predict potential antibiotic combinations. For example, Yilancioglu et al. (Yilancioglu et al., 2014) investigated the relationships between Lipinski’s five rules and synergistic drug combinations, and they found a significant correlation (r = 0.51, p = 3.6 × 10^–3) between synergistic drug combinations and lipophilicity. Chandrasekaran et al. (Chandrasekaran et al., 2016) used chemogenomic data of Escherichia coli (E. coli) as features to build an AI-based model and subsequently predict new antibiotic combinations (AUC for synergy = 0.79). On the download page, users can obtain the entire chemogenomic data of E. coli in 324 conditions (Nichols et al., 2011). For other bacterial strains, orthologous genes can be obtained from OrtholugeDB (Whiteside et al., 2012). The disadvantage of the method is that chemogenomic data are expensive to obtain. Computational features add another alternative for AI-based models. Mason et al. (Mason et al., 2017) used MACCS keys to build an AI-based classifier and obtained acceptable results (AUROC = 0.74). On the download page, we offered three types of molecular fingerprints of each compound (MACCS keys, Morgan fingerprint and topological pharmacophore fingerprints) for users to choose.

While machine learning models can offer satisfactory outcomes, the mechanisms underlying the synergy effect are still poorly understood. As such, mechanism-driven methods are needed in order to predict antibiotic combinations. Network pharmacology provides a new paradigm to explore intricate relationships between drugs, genes, and diseases (Hopkins, 2008). ACDB provides targets for each compound and several common protein-protein interaction (PPI) networks. Furthermore, Cytoscape (Shannon et al., 2003) and the Python package networkx can be used to draw the PPI network and calculate topological parameters (e.g. degree, betweenness, eigenvector centrality) for each node. Zou et al. (Zou et al., 2012) used these topological parameters to explore the underlying mechanisms of drug combinations. Network-based proximity (Cheng et al., 2019; Lv et al., 2022) can also be used to measure the relationship of two drugs. Based on the ACDB, comprehensive studies of antibiotic combinations at the system level can be undertaken.

Pharmacologically, an antibiotic combination may produce synergy effect, additive effect, and antagonism effect (Cokol et al., 2011). Every antibiotic combination has its own advantages. For synergistic antibiotic combinations, they are frequently used in clinics because they can provide improved efficacy at lower dosages (Yeh et al., 2009). The combination of trimethoprim and sulfamethoxazole, for example, can interfere with folic acid synthesis in a synergistic way (Yeh et al., 2006). For antagonistic antibiotic combinations, they have been shown to slow down the evolution of AMR (Chait et al., 2007; Michel et al., 2008). However, the potency of antibiotic combinations is not immutable and it is affected by metabolic conditions (Cokol et al., 2018), bacterial strains (Chandrasekaran et al., 2016), etc. This is one of the important drivers for development of ACDB. Through the “Organism Search” in ACDB, users can obtain a series of species-specific antibiotic combinations and their efficacy. Undoubtedly, ACDB combined with antimicrobial susceptibility testing can help clinicians tailor treatments based on the pathogen microenvironment and the patient’s condition.