Over the past ten years, most research on drug discovery has focused on searching for highly selective molecules acting on a single target [1]. However, the highly selected single-target affinity always leads to the instability of the cellular metabolic system, which seriously affects the normal physiological functions of cells and then produces adverse actions. So, the development of single-target drugs is severely restricted [2]. Meanwhile, some complex diseases, such as drug-induced liver injury (DILI), are regulated by multiple targets and are often difficult to cure by a single-target drug [3]. It has been found that multi-target drugs are essential for treating complex diseases [4]. Thus, the research and development strategy for multi-target drugs will become the main topic of the pharmaceutical industry in the future [5]. Multi-target drugs can be defined as chemical entities that combine the pharmacophores of two or more targets with different mechanisms of action in a single molecule, capable of simultaneously interacting with two or more molecular targets [6]. Polypharmacology, treating complex diseases by modulation of multiple targets with one or more drugs [7], has been widely recognized as a new direction of modern multi-target drugs discovery and focuses on many targets that single drugs can hit [8]. With the development of in silico pharmacology [9], many strategic approaches to studying multi-target drugs have been proposed with great success, including molecular docking [10,11], network pharmacology [12,13], multi-omics-based system biology [14,15], machine learning [16,17], Multi-target Quantitative Structure-Activity Relationship (mt-QSAR) [18,19], perturbation model combined with machine learning (PLMT) [20,21]and pharmacophore modeling [22]. These methods, mainly based on common elements of multi-target ligands or binding strength of ligand-protein, are used to screen multi-target drugs. In addition, the relationship between the action of multiple targets and the drug's overall efficacy is also essential for developing multi-target drug discovery that should be considered.

With the continuous development and application of modern technological methods, machine learning has gradually gained scholars' attention and has been widely used in numerous studies of polypharmacology [23]. Moreover, machine learning also shows distinct advantages in multi-target drug discovery and drug repositioning [24]. For example, through a machine learning technique that uses multiple CPIs, Hiroaki Yabuuchi [25] et al. have successfully identified novel lead compounds for two pharmaceutically essential protein families, G-protein-coupled receptors and protein kinases. Guomeng Xing [26] combined machine learning and deep learning to build an integrated model of the three main targets of protein tyrosine kinases and screened Syk/JAK or Btk/JAK dual-target inhibitors for the treatment of rheumatoid arthritis. The molecular docking method, a computational method to predict the binding strength between organic molecules and biological macromolecules [27], is widely used to discover and design multi-target drugs with the advantages of high efficiency, time-saving, and so on [28]. For example, Yunqi Li [29] et al. found that 2-arylbenzimidazole compounds could act as inhibitors of Epidermal Growth Factor Receptor (EGFR), Vascular Endothelial Growth Factor Receptor-2 (VEGFR-2), and Platelet-derived Growth Factor Receptor (PDGFR) by screening a series of new benzimidazole derivatives using support vector machine (SVM) and molecular docking. Prabhavathi [30] et al. performed virtual screening of phytochemical inhibitors by molecular docking and dynamic simulation and found that panaxadiol could be developed into a novel multi-target inhibitor of EGFR and Human Epidermal Growth Factor Receptor-2 (HER2) with less toxicity. Therefore, machine learning and molecular docking have provided technical support to construct computational models for studying the relationship between the action of multiple targets and the drug's overall efficacy.

DILI has become a worldwide health problem and has increasingly attracted public attention. Besides, DILI ranks as the first cause of acute liver failure in Europe and the USA. As a representative of complex diseases, DILI [31] includes complex pathogenesis, such as direct hepatotoxicity, oxidative stress, mitochondrial dysfunction, immune responses, Etc. Kinds of literature have confirmed that Farnesoid X Receptor (FXR) [32], Liver X Receptor α (LXR-α) [33], Pregnane X Receptor (PXR) [34], Protease-Activated Receptors 1 (PAR-1) [35] and Peroxisome Proliferators-Activated Receptor α (PPAR-α) [36] could all play a key role in treating the DILI. Therefore, based on the above five targets, constructing a relationship model between the action of multiple targets and the drug's efficacy is beneficial to finding multi-target drugs that exert overall efficacy and provide a new method for treating DILI.

Traditional Chinese medicine (TCM) has distinctive characteristics that multi-components could act on multi-targets to treat complex diseases through multi-pathways [37]. Modern research has shown that TCM compounds such as resveratrol [38,39], berberine [40,41] and curcumin [42,43] could regulate various pathological characteristics and have become an essential material of multi-target drugs [44]. Therefore, TCM provides rich material for the development of multi-target drugs.

Based on the above ideas, in this paper, a novel approach is first reported to clarify the relationship between the action of multiple targets and the drug's overall efficacy. As shown in Fig. 1, with the case of DILI, we introduce a method named multi-target based polypharmacology prediction (mTPP), a computational model using virtual screening and machine learning for multi-target drug discovery. We trained the model using four machine learning algorithms, with the binding strength of ingredients with multi-target and the proliferation rate of components against APAP-induced injury L02 cells as input. By comparison, the model based on the GBR algorithm has better accuracy and is suitable for a multi-target-based polypharmacology prediction. Next, the mTPP model was used to predict hepatoprotective ingredients from the Traditional Chinese Medicine Chemistry Database (TCMD). Finally, in vitro cell assay was employed to validate the activity of hepatoprotective ingredients.