The discovery of small molecule drugs is time-consuming, expensive and labor-intensive. (Dickson and Gagnon, 2004; Paul et al., 2010; Dimasi et al., 2015) It is resource intensive, and involves typical timelines of 10–20 years and costs that range from US$0.5 billion to US$2.6 billion (Paul et al., 2010; Avorn, 2015). In addition to economic and technical reasons, the main reason is that almost half of the candidate drugs failed in clinical trials. Up to 25% of compounds were withdrawn due to metabolic, pharmacokinetic, or toxic problems (Hwang et al., 2016). Drug metabolism can produce metabolites with physicochemical and pharmacological properties, which are significantly different from the physical and pharmacological properties of parent drugs (Kirchmair et al., 2013). As Figure 1 shows, when drugs or other exogenous substances enter the human body, they are largely controlled by three stages of drug metabolism. In the first stage, reactive groups are introduced by oxidation, reduction, or hydrolysis. In the second stage, conjugation reactions with macromolecules occur in vivo. In the third stage, allogeneic and metabolites are removed from liver and intestinal cells. After these three stages, exogenous substances such as drugs may be transformed into non-physiological active substances or toxic metabolites. 70% clinical drugs are removed by the body’s metabolic system, so as part of drug development, it is also necessary to conduct in-depth research on drug metabolism (Grant et al., 2001; Embrechts and Ekins, 2007; Lazar and Birnbaum, 2012; Damsky and Bosenberg, 2012; Di, 2014; Mackenzie et al., 2017; Kang et al., 2018). Understanding the metabolic process of drugs is essential for successful drug discovery and development, and helps to optimize the stability of drugs, so as to optimize the half-life in vivo.

In order to reduce the risk caused by metabolic characteristics of candidate drugs, effective andreliable methods are needed to predict drug metabolism in vitro. Many experimental methods can be used to explore the metabolic process of drugs (Diao et al., 2016; Mackenzie et al., 2017). For example, fast LC-MS scans can be carried out to specifically detect predicted metabolites. However, experimental methods are still highly demanding in terms of equipment, expertise, cost, and time (Kirchmair et al., 2013). Therefore, it has great prospects to develop computational tools for predicting drug metabolism with lower cost and higher throughput than existing tools. Many different methodologies to predict metabolites or sites of metabolism have been reported recently. Various methods in predicting drug metabolism using in silico approaches have been reviewed (Fox and Kriegl, 2006; Gleeson et al., 2011; Zhang et al., 2011; Tan et al., 2017). However, most of these methods are limited to P450 catalytic reactions and represent only unstable sites, rather than predicting the actual metabolites formed.

Metabolic sites (SOMs) and metabolite structure are two main research directions of computer-aided metabolic prediction methods, which can provide decisive support and guidance for experimenters. SOM prediction methods usually have high prediction accuracy. The program MetaSite estimate the possibility of metabolic reactions at an atomic site using protein structure information, GRID-derived MIFs of protein, and ligand and molecular orbital calculations (Gabriele et al., 2005). The program SMARTCyp contains a pre-calculated energy reaction analysis table for density functional theory activation, where a large number of ligand fragments pass through CYP3A4 or CYP2D6 mediated transformation (Rydberg et al., 2010a; Rydberg et al., 2010b). A method called cypscore, in which 2400 CYP-mediated transformations and 850 literature compounds are used as data bases (Hennemann et al., 2010). However, most of these methods are limited to CYP450 catalytic reactions and can only predict unstable sites rather than metabolite structures. Furthermore, predicted SOMs are not identical to identifying the correct bioinformations that will occur at an atomic location, and they do not provide information about the type of reaction that will occur. Therefore, these limitations make it difficult to draw any quantitative conclusions about the metabolic possibilities of a molecule.

So far, only a few computational methods have been developed for predicting the structure ofmetabolites. Existing methods can be divided into two categories: expert rule-based anddescriptor-based. Rule-based approaches use data mining techniques. Large databases with data onmetabolism are used to extract generalized rules to determine the part of a molecule that undergoes metabolic alteration (Cariello et al., 2002). The ligand-based approach relies on the assumption that the metabolic fate of compounds is entirely determined by their chemical structure and properties. These methods include quantum mechanics methods. Descriptor-based methods to obtain an idea of the route of a drug through the metabolic system, the identification of the involved enzymes, and the reaction pathways is necessary (Livingstone, 2010). The program of Bioprint contains a database of most marketed drugs together with reference compounds and data from a wide variety of biological and in-vitro ADME assays, called the Biological fingerprint (Krejsa et al., 2003). Thus, the possible results of new compounds can be calculated by neighborhood relation and QSAR model. In the MetaDrug database (Ekins et al., 2006; Ekins et al., 2005b), metabolic reactions with substrates (including primary and secondary metabolites), xenobiotic reactions, and kinetic data on enzyme inhibition are stored. 317 molecules (parent drug and primary and secondary metabolite) were randomly selected from this database to build kernel-partial least squares models for metabolism rules (Embrechts and Ekins, 2007). Metabolite prediction is usually accomplished through a large set of transformation rules. Given the reactant, all rules are then matched to determine the site of metabolic instability. Expert systems such as METEOR (Testa et al., 2010; Button et al., 2015), META (Klopman et al., 1994; Talafous et al., 1994; Klopman et al., 1997), MetabolExpert (Darvas, 1988), RD-Metabolizer (Meng et al., 2017), MetaDrug (Korolev et al., 2003; Ekins et al., 2005a), and KnowItAll (Stouch et al., 2003) are based on these databases and provide a ranked list of most likely metabolites. In a study described by AstraZeneca (Scott et al., 2007), the substrates and reaction centers of the metabolite database were stored as fingerprints in two databases. Then the query molecule powders are compared with the two databases, and the proposed SOM is ranked by using the number of clicks as a weighted scheme. An approach called SyGMa based on the MDL metabolite database was developed (Ridder and Wagener, 2010). According to the corresponding rules of MDL metabolite data coding, the structure of possible metabolites is predicted, and probability scores are assigned to each metabolite, covering 70% of all known human metabolic reactions.

So far, one of the difficulties in predicting possible metabolites is that this task means identifying the reaction site (SOM) and the type of metabolic reaction correctly. Current methods for predicting metabolite structure tend to have high false-positive rates and can only be used for specific enzymes without covering all the metabolic enzymes involved in human reactions. In view of the above problems, we mainly designed a deep learning algorithm combined with drug metabolism characteristics.

In this work, by combining metabolic reaction template and Deep Learning, we have established a model to predict the main metabolites of drugs ( Figure 2 ). Our method has the following innovations: (1) Data enhancement strategy, which provides chemically reliable examples of negative reactions through the metabolic reaction template library; (2) the implementation and validation of a neural network-based model, which can obtain that some reaction modes are more or less likely to occur than other potential modes.

In summary, we developed a deep learning based drug metabolites prediction algorithm to complement the experimental methods. By generating a broad coverage of metabolic reaction templates, we can generate a large number of potential metabolic reactants, and rank all metabolites by deep neural network algorithm to get the right metabolites ranked high. The accuracy of Top-1, Top-3, Top-6, and Top-10 in 300 external test sets with metabolic reactions is 35, 55, 67, and 78% respectively, which is significantly higher than that of random guess and the rule-based method SyGMa. Nevertheless, our method still has some limitations. It can rank the metabolites, but cannot give the probability of occurrence of metabolic sites. Besides, despite the relatively high prediction accuracy, it still has a high false-positive problem. To sum up, A approach of drug metabolites prediction based Deep learning was developed in this paper, which has certain predictive ability and can be used to provide some guidance information for researchers to improve the metabolic properties of lead compounds.

The raw data supporting the conclusions of this article will be made available by the authors, without undue reservation, to any qualified researcher.

DW conducted the research and wrote the paper. WL validated the model using known drugs and natural products. ZS helped collect the metabolic reaction data. LJ participated in building the metabolic reactions template library. JW tested the trained model. HL and SL designed and performed research, interpreted data, and approved the final manuscript.

This work was supported by the National Key Research and Development Program (2016YFA0502304 to HL); the National Natural Science Foundation of China (grant 81825020 to HL, 81803437 to SL); the National Science & Technology Major Project “Key New Drug Creation and Manufacturing Program,” China (Number: 2018ZX09711002); the Fundamental Research Funds for the Central Universities, Special Program for Applied Research on Super Computation of the NSFC-Guangdong Joint Fund (the second phase) under Grant No.U1501501. SL is also sponsored by Shanghai Sailing Program (No. 18YF1405100). HL is also sponsored by National Program for Special Supports of Eminent Professionals and National Program for Support of Top-notch Young Professionals.

Author WL was employed by company Jiangzhong Pharmaceutical Co., Ltd.

The remaining authors declare that the research was conducted in the absence of any commercial or financial relationships that could be construed as a potential conflict of interest.

The reviewer MZ declared a past co-authorship with one of the authors HL to the handling editor.