In-silico prediction of anti-breast cancer activity of ginger (Zingiber officinale) using machine learning techniques

2.1. Data mining and fingerprint extraction

algorithm. A machine learning algorithm predicted protein KIT and MAPK2 inhibitors in ginger. Ginger compound data is gathered from the Knapsack database core system using the keyword Zingiber officinale [15]. Using the open-source RDkit in Python software base (https://www.rdkit.org/), smile structures from all compounds are extracted to create Klekota-Roth fingerprints. There are 4860 substructures in the Klekota-Roth fingerprint, each with a binary score of one or zero [16], through Dudedocking, a dataset of chemicals that inhibit KIT and MAPK2 proteins was gathered [12]. The website also provides a decoy compound that can be utilized as a non-active compound. Machine learning models would be created using active and inactive substances. Before developing a model, each substructure is obtained using the fingerprint extractor from RDKit.

2.2. Machine learning model development

Scikit-learn, a Python-based program, is used to create machine learning models [17]. Jupyter notebooks are used to write and implement the code [18]. The AUC/ROC score was used to select the most effective model. In this analysis, we use three algorithms that have been often referenced in other works. Random forest (RF), Support vector machine (SVM), and Logistic regression are three algorithms (RF). The three methods are compared, and the model with the highest Score is utilized to forecast the active chemicals in ginger.

2.3. Molecular docking from predicted active compound and interaction analysis

A machine learning-predicted active substance is subjected to molecular docking investigation. To determine the probable binding affinities of each component from ginger, we used the PLANTS 1.1 software [19]. The ligand’s minimal energy conformation in the protein’s binding region is determined using an artificial ant colony [20]. Two empirical scoring functions, PLANTSCHEMPLP and PLANTSPLP, were created for the docking method PLANTS (Protein-Ligand ANT System), which is based on ant colony optimization (ACO). The optimal parameter settings for the search algorithm have been found, and they can be used to Balance pose prediction accuracy with search speed [19]. PRODIGY was used to calculate binding affinity (ΔG) after molecular docking [21]. PRODIGY protein-ligand principles (https://wenmr.science.uu.nl/prodigy/lig) by appropriately modifying the small ligand prediction technique to use atomic interactions rather than residue contacts. PRODIGY-advantages LIGs include their simplicity, generality, and applicability to all protein-ligand complexes [22].

2.4. ADME (Absorption, Distribution, Metabolism, and Excretion) and toxicity analysis

The purpose of ADME analysis is to evaluate how much each chemical resembles a medicine expected to be active by machine learning. To ensure a drug’s pharmacokinetics, an examination called an ADME analysis (Absorption, Distribution, Metabolism, and Excretion) is performed. SwissADME is used for the ADME analysis (http://www.swissadme.ch/) [23]. Using Toxtree, each compound’s toxicity is examined [24]. The toxicological concern threshold is frequently called the toxicity level (TTC). TTC refers to determining an exposure threshold for all substances below which there is no appreciable danger to human health [24]. The Lipinski rule of five is a frequently used metric to measure drug-likeness, which includes (1). a maximum of five hydrogen bond donors (the sum of the bonds between hydrogen and oxygen and nitrogen), (2). ten or fewer (all nitrogen or oxygen atoms) hydrogen bond acceptors; (3). fewer than 500 daltons for the molecular weight (4). No more than 4.15 LogP [25].

Blood-brain barrier (BBB) and human intestinal absorption (HIA) analyses are frequently combined with toxicity analysis to predict the number of compounds that can be absorbed by the gastrointestinal tract (GI) [26]. Toxicity analysis indicates whether these compounds can pass through the blood-brain barrier.

2.5. Analysis of bioactivity prediction

The final technique we employed in this investigation was bioactivity prediction. The web-based PASS-SERVER allows for the prediction of bioactivity. This bioactivity prediction considers pharmacological effects, action mechanisms, toxic and adverse effects, interactions with metabolic transporters and enzymes, impact on gene expression, etc [27].

2.6. Tanimoto similarity for chemical structure and pharmacophore

Potential compounds already mentioned above are then converted to fingerprints. The fingerprint of each compound is used to calculate structure similarity structure with reference ligand control using RDKit in jupyter notebook [18]. The molecular docking result of each compound has to interact with several amino acids of the protein; the active residue that interacts with the ligand compound is also used to calculate the similarity interaction with the reference ligand as a control. This method tries to see if the ligand interacts similarly with the reference ligand using Pyplif-Hippos [28].

3.1. Machine learning model development

Table 1 displays the Score for the MAPK2 protein prediction model. Regarding the final result, the AUC/ROC score (0.997).

It is also displayed by KIT protein, where logistic regression (LR) has the best Score regarding the AUC/ ROC score (0.984) (Table 2). The logistic regression (LR) model is then used to predict potential active compounds from ginger that possibly inhibit MAPK2 and KIT protein.

We discovered ten compounds that a machine learning model suggested could potentially be a KIT protein inhibitor (Table 3), and a total of 64 compounds are predicted as active chemicals because of the MAPK2 protein model (Table 4). Compounds predicted as inhibitors for KIT and MAPK2 protein are subjected to molecular docking to explore possible attachment in the binding pocket of KIT and MAPK2 protein.

Software validation has been performed before using PLANTS 1.1 software for protein-ligand-docking. The PLANTS 1.1 software is used in a benchmarking analysis with data from Dudedocking [12]. Docking and binding affinity scores (ΔG) are calculated using active KIT and MAPK2 protein inhibitors and non-active (decoy) compounds. The docking and binding affinity scores (ΔG) are then projected to generate the AUC/ROC score for the PLANTS 1.1 software results using the predicted docking score and binding affinity of active and decoy compounds. This method was used to test whether the PLANTS 1.1 software could distinguish between active and decoy compounds based on the binding and affinity scores. The AUC/ROC score of retrospective analysis from PLANTS 1.1 software yields excellent results, with AUC ROC scores of 0.992 for KIT protein and 0.986 for MAPK2 protein (scale 0–1, the getting closer to 1 the better result) (Fig. 3).

OPEN IN VIEWER

Fig. 2.

AUC/ROC score for KIT protein (A) and AUC/ROC score for MAPK2 protein.

KIT (PDB: 1T46) and MAPK2 (PDB: 3M2W) were the proteins used in this study. Both proteins have a reference ligand that acts as an inhibitor. Imatinib (DrugBank code: DB00619) is an inhibitor of the KIT protein (PubChem code: 5291). The drug has already docked with the KIT protein (PDB: 1T46). While the reference ligand for the MAPK2 protein is 2 ^′ -(2-fluorophenyl)-1-methyl-6 ^′ ,8 ^′ ,9 ^′ ,11 ^′ -tetrahydrospiro [azetidine-3,10 ^′ -pyrido [3 ^′ ,4 ^′ :4,5] pyrrolo [2,3-f] isoquinolin], -7 ^′ (5 ^′ H) -one (PubChem code: 42646698). PLANTS 1.1 was used to perform 100 repetitions of re-docking reference ligand to validate the accuracy of the PLANTS 1.1 software (Fig. 3). The average docking score of a reference ligand to KIT protein is −142.904, while that of a MAPK2 reference ligand is −101.628.

OPEN IN VIEWER

Fig. 3.

PLANTS 1.1 docking Score between ref_ligand to KIT protein (A) and MAPK2 protein (B) for 100 repetitions using PLANTS 1.1.

Figure 3 shows the binding affinity calculation result after 100 repetitions. The average Root mean square deviation (RMSD) of the reference ligand to KIT protein is 0.7365, while the RMSD of the reference ligand to MAPK2 protein is 1.3256. The RMSD after 100 repetitions of re-docking was stable. The figure shows that the RMSD of docking ligand to reference ligand after 100 repetitions to protein was stable below 2.0 Å [29].

OPEN IN VIEWER

Fig. 4.

Root means square deviation (RMSD) of re-docking between ref_ligand to KIT protein (A) and MAPK2 (B) protein for 100 repetitions using PLANTS 1.1.

Beta_caroten is the only compound predicted as active from a machine learning model for KIT protein with a lower binding affinity score (ΔG) than the reference ligand (ref_ligand). The binding energies score (ΔG) for the reference ligand is −10.4 Kcal/mol, while the Score for beta-carotene is −14.3 Kcal/mol. This finding indicates that beta-carotene may bind better to the KIT protein (Table 3). On the other side, 14 of the 64 compounds predicted as active by the machine learning model for the MAPK2 protein have lower binding energies than the reference ligand. Lycopene, 5-Hydroxy-7,4 ^′ -dimethoxyflavone, [12]-Shogaol, Isogingerenone B, Curcumin, trans-[10]-Shogaol, Gingerenone A, Dihydrocurcumin, Demethoxycurcumin, [8]-Shogaol, [6]-Shogaol, [1]-Paradol, Bisdemethoxycurcumin, and alpha-Zingiberene are among the 14 compounds (Table 4).

After calculating compounds classified as active and inactive KIT and MAPK2 protein inhibitors, we found that a machine-learning model revealed promising results. Compounds grouped as an active protein to inhibit KIT have a binding affinity (ΔG) ranging from −7.1 (Isovanilin) to −14.3 (beta-carotene). No compounds with a lower binding affinity score than −7.0 Kcal/mol are classified as active inhibitors of the KIT protein. A similar accurate prediction result was also shown in the MAPK2 protein. Binding affinity (ΔG) ranges from −7.4 Kcal/mol (3-Octen-2-one) to −10.6 Kcal/mol (Lycopene and 5-Hydroxy-7,4 ^′ -dimethoxyflavone) (Table 4). The compounds 5-Hydroxy-74 ^′ -dimethoxyflavone, lycopene [12] -Shogaol, Isogingerenone B, curcumin, Trans-[10]-Shogaol, Gingerenone A, Dihydrocurcumin, and Demethoxycurcumin were found to be more effective than the reference ligand, despite having a lower binding affinity. While [8]-Shogaol, [6]-Shogaol, [1]-Paradol, Bisdemethoxycurcumin, and alpha-Zingiberene appear to be similar but better than the reference ligand.

However, after ADME analysis, beta carotene is the only ligand with a lower binding affinity than the reference ligand of the KIT protein. Beta caroten has two Lipinski violations and is classified as intermediate or high due to toxic prediction using toxtree (Table 5). Lycopene, [8]-Shogaol, [6]-Shogaol, [1]-Paradol, and alpha-Zingiberene were identified as toxtree low toxicants. Due to Lipinski rules in ADME analysis, each compound also shows no Lipinski violation. The compounds [8]-Shogaol, [6]-Shogaol, and [1]-Paradol have docking scores identical to those of the reference ligand. The binding affinity score for reference ligands is −9.2 Kcal/mol, while the scores for [8]-Shogaol, [6]-Shogaol, and [1]-Paradol are all −9.3 Kcal/mol. Compounds classified as at least intermediate (Class 2) in toxic prediction using toxtree software [24] with no violations of the Lipinski rule [25] are subjected to bioactivity prediction analysis using pass-server (http://www.way2drug.com/passonline) [27]. Lycopene, [12]-Shogaol, [8]-Shogaol, [6]-Shogaol, [1]-Paradol, and alpha-Zingiberene are among the compounds filtered from Lipinski and toxtree (Table 6).

3.4. Analysis of bioactivity prediction

The outcome of bioactivity is then visualized using orange [30]. The heatmap analysis (Fig. 4) reveals a variety of potential activities for each compound. We discovered that each compound chosen has the potential to be an apoptosis agonist [31], which increases apoptosis in cancer cells. Except for [1]-Paradol, all compounds discovered have bioactivity as a proliferative disease treatment, which means that each of the compounds mentioned can both inhibit and aid in the proliferation of cancer cells. All chosen compounds have the potential to be anticarcinogenic. Only [1]-Paradol and alpha-Zingiberene lacked potential antineoplastic activity. Alpha-Zingiberene is also the only compound that did not exhibit TP53 expression enhancer activity, which is responsible for suppressing the cancer development process. The TP53 gene causes cell-cycle arrest, senescence, or apoptosis in response to cellular stressors such as DNA damage, hypoxia, nutritional deficiency, and oncogenic signaling [32]. Approximately half of all human cancers have lost or altered the tumor suppressor p53 [33].

Lycopene was discovered to be antineoplastic in certain cancers, including breast cancer. The anticancer activity is also shown in beta-carotene. Although beta-carotene has two Lipinski violations after ADME analysis, the antineoplastic properties of beta-carotene have been found in the brain, lung, lymphoma, pancreatic, and solid tumors. Beta-carotenes can increase apoptosis, stop the cell cycle at various stages, and prevent cell proliferation [34]. Lycopene is also one potential active compound that has antineoplastic potential in liver cancer, lung cancer, ovarian cancer, and renal cancer. Lycopene has also been identified as an antineoplastic enhancer. Lycopene has also been shown to have anti-cancer properties in the development of gastric cancer [35]. Some of the signal transduction pathways regulated by lycopene include the manipulation of the insulin-like growth factor system, the inhibition of the activity of sex steroid hormones, the modification of significant gene expression, and the alteration of mitochondrial function [36]. It is also found that alpha-zingiberene is the only compound with antineoplastic activity in pancreatic cancer cells (Fig. 5).

OPEN IN VIEWER

Fig. 5.

Bioactivity prediction from compounds selected from ADME and toxic prediction.

Beta-carotene is shared about 6.55% similarity structure with the reference ligand for KIT protein inhibitor Imatinib (PubChem code: 5291), while MAPK2 protein reference ligand 2 ^′ -(2-fluorophenyl)-1-methyl-6 ^′ ,8 ^′ ,9 ^′ ,11 ^′ -tetrahydrospiro [azetidine-3,10 ^′ -pyrido [3 ^′ ,4 ^′ :4,5] pyrrolo [2,3-f] isoquinolin], -7 ^′ (5 ^′ H) -one (PubChem code: 42646698) has the highest similarity with 5-Hydroxy-7,4 ^′ -dimethoxyflavone (9%). None of the compounds selected have a similarity greater than 10%. The model can choose the possible active compound even with a smaller similarity fingerprint of compounds from ginger with an active compound of training and testing data (Table 7).

The Pharmacophore similarity of amino acids interacting with the compound selected as a ligand is also calculated. Beta-carotene only shares about 12.9% pharmacophore similarity with reference ligands for KIT protein. This means there are substantial differences between amino acids from proteins, which are amino acids from proteins that interact with beta-carotene compared to reference ligands. On the other hand, we found Gingerenone A share about 33.3% similarity with reference ligand for MAPK2 protein, Isogingerenone B, and [1]-Paradol also shows higher pharmacophore similarity with MAPK2 reference ligand with a similarity of 27.8% and 22.2% (Table 8). It is clear that even though 5-Hydroxy-7,4 ^′ -dimethoxyflavone ligands have the highest structure similarity (9%) with reference ligands, the compound does not always show the same result in pharmacophore similarity; the compound interaction only shares about 11.1% similarity.