Targets preliminary screening for the fresh natural drug molecule based on Cosine-correlation and similarity-comparison of local network

The CSLN mainly uses the following databases for experiments and verification. We acquired four datasets of the DTI network from the gold standard data [23] coovering nuclear receptors (NR), enzymes (EN), G-protein coupled receptors (GPCR) and ion channels (IC). And it can be downloaded from http://​web.​kuicr.​kyoto-u.​ac.​jp/​supp/​yoshi/​drugtarget/​. Each dataset includes 2 types of information, the observed DTIs and the similarities among drugs. In addition, the detailed statistics of the above four datasets are shown in Table 1.

The specific implementation process of CSLN is shown in Fig. 1, which can be subdivided into the following steps. First, we got the molecular fingerprint through MACCSkeys based on Rdkit [24] for all drug molecules. Second, Tanimoto [25] was used to calculate the similarity between two molecules. Third, the screened target was expressed in combination with related drugs, and its similarity with the fresh drug molecule was calculated with Cosine-correlation [16]. Fourth, compared the average similarity of local networks after the addition of the fresh drug molecule with that before the addition. Finally, the binding score of the target to the fresh drug molecule was obtained by combining the two values and assigning different weights to them through negative feedback adjustment in machine learning. Then the scores were ranked from high to low, and the predicting results of the higher score were more likely to be the potential DTIs.

The drug-target dataset was described as a binary network \(V=\left(D,T,A\right)\). \(D=\left\{d1, d2, ..., dm\right\}\) was a collection of drug nodes, \(T=\left\{t1, t2, ..., tn\right\}\) was the set of target nodes and \(A=\left\{a11, ..., aij, ..., amn\right\}\) was the set of edges between interconnected nodes in the network, where D and T respectively represented two independent sets. If there was a known interaction between the drug di and the target \(tj\), then set \(aij = 1\), otherwise set \(aij = 0\).

Based on RDKit, the characteristics of chemical molecules were expressed in binary. The MaccsKeys fingerprint was put forward by a company whose name is MDL and had a total of 166 features, but the total length of MaccsKeys was 167bits, because bit 0 was a placeholder, and bit 1–166 was a molecular feature bit. Then, the drug molecule di can be expressed as:

The Tanimoto score between the drug molecule \(di\) and \(dj\) is calculated according to the following formula:where:

\(Fdi\) is the elements of molecular fingerprint of \(di\);

\(Fdj\) is the elements of molecular fingerprint of \(dj\);

For example, we calculate the binding score between a target \(ty\) and a fresh drug molecule \(dy\), the drug molecules \(d1, d2, ..., dx\) are which interact with\(ty\).

Here, \(d1, d2, ..., dx\) together denote\(ty\):

Cosine-correlation of \(ty\) and \(dy\)

Then, the mean similarity of drug molecules (\(d1, d2, \dots \dots , dx\)) S1 is calculated by the following formula

And, the mean similarity S2 is calculated according to the following formula when \(dy\) is merged with the drug moleculeswhere

\(TN(a, b)\) represents the similarity between drug molecule a and drug molecule b (Obtained through Tanimoto);

Finally, we can get the binding score of target \(ty\) with drug \(dy\) according to the formulawhere w1 (Global Shared) is the weight value between the Cosine-correlation and the Similarity-comparison scores. And they are obtained through feedback learning in training. The calculation formula of residual error in feedback learning is as followswhere \({y}_{i}\) is the true label of the sample, and \({\widehat{y}}_{i}\) is the predicted value.

In the past decades, triptolide, a very widely studied natural drug molecule, has attracted considerable interest in the organic and medicinal chemistry society owing to its intriguing structural features and promising multiple pharmacological activities. However, its imprecise mechanism of action and severe toxicity have greatly hindered its clinical potential [21]. Therefore, in this study, triptolide was selected as the experimental object to predict its targets by CSLN and TCMSP was used to build the training set and test set. Notably, the environment of new natural drug molecules was simulated in this experiment.

The HL-7702 cell lines and L02 cells were obtained from China Academy of Chinese Medical Sciences (Beijing, China). L02 cells were cultured in Dulbecco’s Modified Eagle’s Medium (DMEM, Sigma, USA) supplemented with 10% fetal bovine serum (Solarbio, Beijing, China), 100 units/ml penicillin and 100 mg/ml streptomycin (Solarbio, Beijing, China). All cell lines were in a humidified environment containing 5% CO₂ at 37 °C.

The indicated cells were lysed in RIPA buffer (solarbio, Beijing, China) at 4 °C. The protein concentration was determined according to BCA Protein Assay Kit (solarbio, Beijing, China), and the equal amounts of cell lysates (30–50 µg of proteins) were subjected to 10% SDS-PAGE. After electrophoresis, proteins were transferred onto PVDF membrane Then the membranes were blocked with 5% BSA at room temperature for 1 h, followed by the incubation with NRH dehydrogenase [quinone] 2(NQO2) (1: 1000; ProteintechGroup, Inc, Beijing, China) and β-actin (1:1000; ProteintechGroup, Inc, Beijing, China) antibody at 4 °C overnight. After the incubation with the primary antibodies, the membranes were rinsed and probed with HRP-conjugated Affinipure Goat Anti-Rabbit IgG(H + L) (1:5000; ProteintechGroup, Inc, Beijing, China) for 1.5 h at room temperature. Then the immune-reactive bands were detected through ChemiDocTM Touch Imaging System (Bio-RAD, USA), and ImageJ software was used to analyze the gray value of the strip.

We mainly compared CSLN with GRGMF [26], the state-of-the-art approach which was published in 2020 and GRGMF has demonstrated superior performance over previously published models in biomedical networks. GRGMF formulated a GMF model which learns the latent factor of each node based on its neighborhood information. And instead of utilizing the similarity matrices derived from external-related databases with predefined metrics, this model could learn the neighborhood information for each node adaptively and further promote the prediction of potential links. And there is no threshold screening for GRGMF because according to the description of the author of GRGMF in the article, it is the result of calculation after decomposition of the whole matrix. If threshold screening is carried out, the information in the whole network will be reduced, so the accuracy will be reduced Here, we mainly compared the AUROC(area under ROC curve) and AUPR(area under the precision-recall) performance of CSLN and GRGMF on gold standard data, and showing the results in Fig. 3. In the results of performance comparison, CSLN's AUROC on EN and IC datasets is superior to GRGMF, meanwhile, for AUPR, the former performs better than the latter in all four data sets.

Further, to demonstrate the reliability of CSLN in targets prediction for fresh natural drug molecules, we did an experiment and took the triptolide (Fig. 4a) as its object. We simulated the environment of a fresh natural drug molecule by deleting its interaction with targets in the database. Meanwhile, from the results, we selected a protein with a high score that had not been found to interact with triptolide in previous relevant work to see whether triptolide could affect its expression with Western-Blot (WB).

We collected data of natural drug molecules from TCMSP (https://​tcmsp-e.​com/​), which includes 6,494 natural drug molecules and 1,718 targets that have interactions with them, and the number of interactions is 54,852. Since triptolide is not a fresh natural drug molecule, the interactions data of it in TCMSP were deleted to simulate the situation of a fresh natural drug molecule and construct a new data-set, and CSLN was used to calculate the binding score between triptolide and the targets in the new dataset. When reconstructing the data-set, we deleted the triptolide known interactions [34] in the data-set, among which 3 targets only interacted with triptolide, to simulated the environment of a fresh natural drug molecule for triptolide. Therefore, 34 interactions and 3 targets were deleted. The detailed statistics of the above two datasets are shown in Table 2.

In this simulated prediction, Leave-One-Out was used to detect the performance of CSLN on the reconstructed data-set, and the data-set was adjusted according to the change of the threshold value to obtain the changing relationship between the threshold value and the performance of CSLN. Finally, the optimal threshold value was selected as 24. The calculation would be skipped when the link degree of the target was less than 24. There were 152 targets (include 10 positive samples) that have been selected. CSLN was used to calculate the binding scores of triptolide with 152 targets and we ranked them according to the score, the top 20 were shown in Table 3.