Neighborhood-based inference and restricted Boltzmann machine for microbe and drug associations prediction

Microbe-drug association

In this article, three different datasets of MDAD (Sun et al., 2018), aBiofilm (Rajput et al., 2018) and DrugVirus (Andersen et al., 2020) were used to test predictive ability of NIRBMMDA. The MDAD dataset presented in the model contains 2,470 associations between 1,373 drugs and 173 microbes that was collected from MDAD database (Sun et al., 2018). Furthermore, the aBiofilm dataset used in the model includes 2,884 associations between 1,720 drugs and 140 microbes collected from aBiofilm database (Rajput et al., 2018). Also, Andersen et al. (2020) built the DrugVirus database for exploration and analysis of broad-spectrum antiviral drugs, in which summarized experimentally verified virus-drug associations. Therefore, the dataset of DrugVirus built here includes 933 associations between 175 drugs and 95 viruses. The statistics of three datasets above are shown in Table 1. Here, we built an adjacency matrix A ∈ R^nd×nm to preserve microbe-drug association information, where nd represents the number of drugs and nm denotes the number of microbes. If drug d_i associated with microbe m_j, the value of entity A_ij is 1. Otherwise, the value is 0.

Drug structural similarity

In this article, SIMCOMP2 search was employed to compute the drug structural similarity (Hattori et al., 2010). SIMCOMP2 search (https://www.genome.jp/tools/simcomp2/), a chemical structure search server, can provide links to the KEGG PATHWAY database that contains manually drawn pathway maps with information about molecular interaction, reaction and relation (Wrzodek, Dräger & Zell, 2011). In SIMCOMP2 search, by mapping drugs of datasets to those in KEGG, we can obtain drug structural similarity with 0.01 of cut off score that filtrate drug structural similarity score of 0.01 or higher (Long et al., 2020a). Then, we defined a matrix DS1 to save drug structural similarity where element $DS1\left(i,j\right)$ denotes the similarity value between drug d_i and drug d_j.

Drug side effect similarity

The drug-side effect association dataset used in this article were downloaded from SIDER (Kuhn et al., 2016). SIDER (http://sideeffects.embl.de/), a side effect resource database, collects information on marketed drugs and their recorded adverse drug reactions. In the dataset, we used N(i) to represent the side effect set associated with drug d_i and employed N(j) to indicate the side effect set of drug d_j. Based on the assumption that the more side effect two drugs share, the more similar between the two drugs. If two drugs do not have the same side effects, the score of side effect similarity between the two drugs is equal to 0. We applied Jaccard score to compute drug side effect similarity, which described as Eq. (2) (Gottlieb et al., 2011). After that, the matrix DS2 was defined to save the drugs side effect similarity and the entity DS2(i, j) denotes the side effect similarity between drug d_i and drug d_j. (2)${\displaystyle DS2=\text{Jaccard}\text{score}=\left|\frac{N_i\cap N_j}{N_i\cup N_j}\right|.}$

Microbe sequence similarity

In the model, three different datasets for known microbe-drug associations were used. For 95 viruses in the DrugVirus dataset, we downloaded their complete genome sequences from the National Center for Biotechnology Information (NCBI, https://www.ncbi.nlm.nih.gov/) based on FASTA format. Then, we employed MAFFT, a multiple sequence alignment program, to align the genome sequence of viruses (Katoh et al., 2002). Since the MAFFT introduces the approximate distance calculation algorithm and the fast Fourier alignment algorithm, its performs well in accuracy of alignments compared with other multiple sequence alignment software including TCoffee version 2 and CLUSTAL W (Katoh et al., 2005). Based on aligned genome sequence of virus, we further used BioEdit to derive the virus sequence similarity matrix. BioEdit, a gratis sequence analysis tool, can compute sequence similarity matrix by using the function of sequence identify matrix (Tippmann, 2004). Specially, for microbes in MDAD dataset or aBiofilm dataset, because of the lack of complete genome sequences in NCBI for nearly all microbes, we downloaded another FASTA format of whole genome shotgun sequence of microbe in NCBI. Then, microbe sequence similarity can further be calculated based on MAFFT and BioEdit. According to the idea that more common sequence two microbes share, the more similar between the two microbes. Therefore, the value of microbe sequence similarity score is equal to 0 when the two microbes have no common sequence. Here, the matrix MS was defined as microbe sequence similarity matrix and MS(m_i, m_j) represented the sequence similarity value between microbe m_i and microbe m_j.

Gaussian interaction profile kernel similarity for drugs and microbes

According the former study (Van Laarhoven, Nabuurs & Marchiori, 2011), we computed Gaussian interaction profile kernel similarity for drugs and microbes based on the known microbe-drug association matrix A. First, we used IV(d_i) to denotes the i − th row vector of matrix A and IV(m_j) to represent the j − th column vector of matrix A. Then, Gaussian interaction profile kernel similarity for drugs and microbes can be computed by using Eqs. (3) and (4), respectively. Here, $∥IV\left(d_i\right)-IV\left(d_j\right)∥$ can be considered as the Euclidean distance for $IV\left(d_i\right)$ and $IV\left(d_j\right)$. Similarly, $∥IV\left(m_i\right)-IV\left(m_j\right)∥$ can represent Euclidean distance for $IV\left(m_i\right)$ and $IV\left(m_j\right)$. (3)${\displaystyle GD\left(d_i,d_j\right)=exp\left(-{\beta }_d{∥IV\left(d_i\right)-IV\left(d_j\right)∥}^2\right)}$ (4)${\displaystyle GM\left(m_i,m_j\right)=exp\left(-{\beta }_m{∥IV\left(m_i\right)-IV\left(m_j\right)∥}^2\right)}$

where ${∥\cdot ∥}^2$ is 2-norm, the β_d and β_m are normalized kernel bandwidth and are defined as follows: (5)${\displaystyle {\beta }_d={\beta }_{\mathrm{d}}^{\prime }/\left(\frac{1}{nd}\sum _{i=1}^{nd}{∥IV\left(d_i\right)∥}^2\right)}$ (6)${\displaystyle {\beta }_m={\beta }_m^{\prime }/\left(\frac{1}{nm}\sum _{i=1}^{nm}{∥IV\left(m_i\right)∥}^2\right)}$

where ${∥\cdot ∥}^2$ is 2-norm, ${\beta }_d^{\prime }$ and ${\beta }_m^{\prime }$ are the original bandwidths, them are set as 1.

Integrated similarity for drugs and microbes

To derive the integrated drug similarity, we combined the drug side effect similarity, drug structural similarity and Gaussian interaction profile kernel similarity of drug. If drug d_i and drug d_j have drug structural similarity or drug side effect similarity, the integrated drug similarity is equals to the average of drug structural similarity and drug side effect similarity. Otherwise, the integrated drug similarity is Gaussian interaction profile kernel similarity of drug. Here, we created matrix SD to save integrated drug similarity. The equation of integrated drug similarity is as follows: (7)${\displaystyle SD\left(d_i,d_j\right)=\left\{\begin{array}{c}\frac{DS1\left(d_i,d_j\right)+DS2\left(d_i,d_j\right)}{2}{d}_i\mathrm{and}{d}_j\text{have}\text{structural}\text{similarity}\mathrm{or}\text{side}\text{effect}\text{similarity}\hfill \\ GD\left(d_i,d_j\right)\text{otherwise}\hfill \end{array}.\right.}$

For integrated microbe similarity, we could obtain the integrated microbe similarity by integrating the microbe sequence similarity and Gaussian interaction profile kernel similarity of microbe. The integrated formula is as follows: (8)${\displaystyle SM\left(m_i,m_j\right)=\left\{\begin{array}{c}MS\left(m_i,m_j\right)m_i\mathrm{and}{m}_j\text{have}\text{sequence}\text{similarity}\hfill \\ GM\left(m_i,m_j\right)\text{otherwise}\hfill \end{array}.\right.}$

NIRBMMDA

In this article, we proposed a computational model of NIRBMMDA by employing multiple biological data including drug similarity, microbe similarity and known microbe-drug associations. In the model, we carried out NI and RBM to identify potential microbe-drug associations, respectively. Then, an ensemble learning was implemented to integrate the two models for gaining final score of potential microbe-drug associations. The whole flowchart of the NIRBMMDA is shown as Fig. 1. The details of the NIRBMMDA are shown as follows.

Neighborhood-based inference

The neighborhood-based method is a collaborative filtering algorithm and can recommend potential preference for a user based on preference of similar users (Su & Khoshgoftaar, 2009). In this article, we presented a based model of NI for predicting new microbe-drug associations. First, we constructed NI model based on integrated drug similarity. For a drug d_{i,i=1,2...,nd}, its neighbors can be obtained by filtering similarity scores based on threshold σ. The set of neighbors for drug d_i can be defined as {d_i|IDS_u,i > σ, u ≠ i}. Based on the set above, the potential association score score1_i,j between drug d_i and microbe m_j can be obtained by computing the sum associations between the microbe m_j and neighbors of drug d_i, which can be described as Eq. (9). (9)${\displaystyle score{1}_{i,j}=\frac{\sum _{i=1}^{nd}\sum _{j=1,u\ne i,ID{S}_{u,i}\ge \sigma }^{nm}{A}_{u,j}\times ID{S}_{u,i}}{\sum _{i=1}^{nd}\sum _{j=1,u\ne i,ID{S}_{u,i}\ge \sigma }^{nm}ID{S}_{u,i}}}$

where nd represents the number of drug, nm represents the number of microbe, A_u,j denotes the element of the u − th row and j − th column in A, and A^T represents the transpose of A. The IDS_u,j denotes integrated drug similarity between drug d_u and drug d_j.

Since a model of neighborhood-based inference was build based on threshold σ, we generate multiple thresholds {σ₁, σ₂, …, σ_ns} to build multiple basic models for reducing the bias of neighbor selection. The value of thresholds σ_i(i =1 ,2,…,ns) is between 0 and 0.5 with step size 0.05. After that, an upper bound parameter σ_upper is used for determining the multiple thresholds that are defined as σ_threshold = {σ_i|σ_i ≤ σ_upper, i = 1, 2, …, n}(ns = |σ_threshold|). In this way, ns thresholds {σ₁, σ₂, …, σ_ns} are used to build ns basic models.

Then, we integrated ns basic models to predict potential microbe-drug associations score by using average strategy, which can be described as follows: (10)${\displaystyle M\text{\_}IDS=\frac{\sum _{k=1}^{ns}score1\text{\_}k}{ns}}$

where ns denotes the number of basic models, score1_k represents predicted microbe-drug associations score based on the k − th threshold, and M_IDS denotes potential microbe-drug associations score based on drug similarity.

Moreover, we constructed a NI model based on integrated microbe similarity. The process of building NI model based on integrated microbe similarity is similar to the process of NI model based on integrated drug similarity. For microbe m_j, its neighbors can be filtrated through using integrated microbe similarity with threshold σ. The set of neighbors for microbe m_j is defined as {m_j|IMS_t,j > σ, t ≠ j}. Based on the set above, the potential association score score2_i,j between drug d_i and microbe m_j was obtained by calculating the sum of associations between the drug d_i and neighbors of microbe m_j as follows: (11)${\displaystyle score{2}_{i,j}=\frac{\sum _{i=1}^{nd}\sum _{j=1,t\ne j,IM{S}_{t,j}\ge \sigma }^{nm}{A}_{i,t}\times IM{S}_{t,j}}{\sum _{i=1}^{nd}\sum _{j=1,t\ne j,IM{S}_{t,j}\ge \sigma }^{nm}IM{S}_{t,j}}}$

where nm represents number of microbe, nd represents number of drug and A_i,t denotes the element of the i − th row and t − th column of A. The IMS_t,j denotes integrated microbe similarity between microbe m_t and microbe m_j.

We used multiple thresholds to build multiple basic models for reducing the bias caused by neighbor selection. Then, we integrated multiple basic models to predict potential microbe-drug associations score by using average strategy, which can be described as follows: (12)${\displaystyle M\text{\_}IMS=\frac{\sum _{k=1}^{ns}score2\text{\_}k}{ns}}$

where ns denotes number of basic models and score2_k represents predicted microbe-drug associations score based on the k − th threshold and M_IMS denotes potential microbe-drug associations score based on microbe similarity.

At last, we obtained final prediction score S1 based on integrated drug similarity and integrated microbe similarity, which can be described as follows: (13)${\displaystyle S1=\frac{M\text{\_}IDS+M\text{\_}IMS}{2}.}$

Restricted Boltzmann machine model

Restricted Boltzmann Machine (RBM), a stochastic neural network, can be used to learn potential probability distribution (Smolensky, 1986). Recently, RBM have been used in numerous fields including movie recommendation, image identification, speech recognition and association prediction in bioinformatics (Hinton, 2012; Wang & Zeng, 2013). In this article, we employed RBM to build a based model for predicting potential microbe-drug associations. As depicted in Fig. 2, RBM is a two-layer network including visible layer and hidden layer, where each layer includes many units. For a RBM, assume that there is a total of nm visible layer units and s hidden layer units. We used v = (v_i, v₂, …, v_nm) to denote set of visible layer units and employed h = (h₁, h₂, …, h_s) to denote set of hidden layer units. Because there is no intra-layer connection for visible layer units or hidden layer units of the RBM, the energy function between v and h can be defined as follows. (14)${\displaystyle E\left(\mathbf{v},\mathbf{h}\right)=-{\sum }_{i=1}^{nm}{b}_i{v}_i-{\sum }_{j=1}^s{c}_j{h}_j-{\sum }_{i=1}^{i=nm}{\sum }_{j=1}^s{w}_{ij}{v}_i{h}_j}$

where nm denotes number of visible layer units, s represents number of visible layer units, b_i is bias of i − th visible layer unit v_i, c_j is bias of j − th hidden layer unit h_j and w_ij represents weight between v_i and h_j.

Based on Eq. (14), we obtained marginal distribution over visible layer units by following equation. (15)${\displaystyle P\left(\mathbf{v}\right)=\sum _{\mathbf{h}}P\left(\mathbf{v},\mathbf{h}\right)=\frac{1}{Z}\sum _{\mathbf{h}}{e}^{-E\left(\mathbf{v},\mathbf{h}\right)}}$

where Z is called the partition function as follows. (16)${\displaystyle Z={\sum }_{i=1}^{nm}{\sum }_{j=1}^s{e}^{-E\left(v_i,h_j\right)}.}$

Because distributions of units in each layer of RBM are independent, the conditional probabilities of visible layer units and hidden layer units can be defined respectively as follows. (17)${\displaystyle P\left(v_i=1|\mathbf{h}\right)=\sigma \left({\sum }_{j=1}^s{w}_{ij}{h}_j+b_i\right)}$ (18)${\displaystyle P\left(h_j=1|\mathbf{v}\right)=\sigma \left({\sum }_{i=1}^{nm}{w}_{ij}{v}_i+c_j\right)}$

where σ(x) = 1/(1 + e^−x) is sigmoid function.

Given a dataset with nm microbes, nd drugs and known microbe-drug associations, a RBM with nm visible layer units and s hidden layer units is built to predict potential microbe-drug associations. For each drug, the observation d_k = {e_k,1, e_k,2, …, e_k,nm} with binary value denotes whether drug d_k is associated with nm microbes. For example, value of e_k,1 is 1 when drug d_k is associated with microbe m₁. Finally, nd drugs have nd observations. When predicting potential associated microbes for a drug, its observation is employed as input of RBM. After that, the prediction is conducted by following two equations. (19)${\displaystyle P_j=P\left(h_j=1|e_{k,1},e_{k,2},\dots ,e_{k,nm}\right)=\sigma \left(\sum _{i=1}^{nm}{w}_{ij}{e}_{k,i}+c_j\right),j=1,2,\dots ,s}$ (20)${\displaystyle e_{k,i}^{{}^{\prime }}=P\left(v_i|P_1,P_2,\dots ,P_s\right)=\sigma \left(\sum _{j=1}^s{w}_{ij}{P}_i+b_i\right),i=1,2,\dots ,nm}$

where σ(x) = 1/(1 + e^−x) is sigmoid function. The output of RBM is defined as score_{k, k = 1 ,2,…,nd} = {e_k,1′, e_k,2′, …, e_k,nm′} which denotes the predicted association score between drug d_k and nm microbes. Here, we defined S2 to save the predicted microbe-drug associations score.

Ensemble learning

Because generalization ability of individual predictor is poor, ensemble learning usually is used to integrate the several wake predictors to obtain a more stronger predictor (Zhou, 2009). Over the last decades, ensemble learning has been successfully employed to solve many problems including feature selection, computer-aided medical diagnosis and text categorization (Keyvanpour & Imani, 2013; Mohebian et al., 2017; Polikar, 2012). In this article, we used ensemble learning to integrate NI and RBM for inferring potential microbe-drug associations. To obtain common scale score ranged from 0 to 1, prediction scores of NI and RBM are normalized by following two functions (Polikar, 2006). (21)${\displaystyle S_{NI}=\frac{\left(tanh\left(0.1\times \left(\frac{S1-{\mu }_1}{{\sigma }_1}\right)\right)+1\right)}{2}}$ (22)${\displaystyle S_{RBM}=\frac{\left(tanh\left(0.1\times \left(\frac{S2-{\mu }_2}{{\sigma }_2}\right)\right)+1\right)}{2}}$

where μ₁ and σ₁ are mean and standard deviation of scores produced by the NI. Similarly, μ₂ and σ₂ are mean and standard deviation of scores produced by the RBM. Subsequently, the different weights were allocated for NI and RBM to derive better prediction performance. Here, we created matrix S to save the potential microbe-drug association score as follows. (23)${\displaystyle S=w_1{S}_{NI}+w_2{S}_{RBM}}$

where w₁ is weight for NI and w₂ is weight for RBM. The sum of w₁ and w₂ is 1.

Performance evaluation

We employed global LOOCV, local LOOCV and five-fold cross validation to evaluate the predicted performance of NIRBMMDA based on the three datasets of DrugVirus (Andersen et al., 2020; Long et al., 2020a), MDAD (Sun et al., 2018) and aBiofilm (Rajput et al., 2018), respectively. In LOOCV, each known microbe-drug association was selected in turn as test sample and remaining known microbe-drug associations were used as training samples. For global LOOCV, all unknown microbe-drug pairs were employed as candidate samples. Then, we used training samples to train model and used the trained model to predict score of test samples and candidate samples. We further ranked test sample with candidate samples based on predicted scores in global LOOCV. At last, we obtained the ranking of all test samples. In local LOOCV, score of test sample was ranked with scores of candidate samples which included the investigated drug of the test samples. At last, we also obtained the ranking of all test samples. In five-fold cross validation, the known microbe-drug associations were randomly divided into five subsets where each subset was regarded as test sample in turn and other four subsets were considered as training samples. All unknown microbe-drug pairs would be treated as candidate samples. Subsequently, we ranked score of each test sample with scores of candidate samples. Finally, we obtained the ranking of all test samples. Particularly, the five-fold cross validation was performed 100 times for avoiding bias caused by random sample divisions. If the ranking of test sample exceeded the given threshold, NIRBMMDA would be considered to achieve a correct prediction. Further, receiver operating characteristics (ROC) curve was plotted through true positive rate (TPR, sensitivity) against false positive rate (FPR, 1-specificity) at diverse thresholds. Sensitivity is the proportion of the test samples which rank over the pre-set threshold, while the specificity is the percentage of candidate samples whose ranking are lower than the appointed threshold. AUC could be used to evaluate the predictive performance of NIRBMMDA. The NIRBMMDA’s prediction is random when the value of AUC is 0.5. If the AUC’s value is 1, the predictive result of NIRBMMDA is perfect.

Because NIRBMMDA integrated two based model of NI and RBM, weights of NI and RBM would affect the performance of NIRBMMDA. We tested 11 group of wights of NI and RBM with a range from 0 to 1 (step size 0.1) for three datasets of DrugVirus, MDAD and aBiofilm respectively based on five-fold cross validation (see Table 2). For DrugVirus, the result showed that NIRBMMDA obtained the best performance of AUC and standard deviation with 0.8569 ±0.0027 when weight of NI is 0.6 and weight of RBM is 0.4. Based on the selected two weights, we compared the performance of NIRBMMDA with other four classical models of HGIMDA (Chen et al., 2016), IMCMDA (Chen et al., 2018a), KATZMDA (Chen et al., 2017) and MDGHIMDA (Chen et al., 2018b) according to five-fold cross validation. The evaluation result showed that our model is better than HGIMDA (0.6995 ±0.0024), IMCMDA (0.6776 ±0.0034), KATZMDA (0.8229 ±0.0022) and MDGHIMDA (0.8293 ±0.0033). Then, according to the two selected weights mentioned above, we compared NIRBMMDA with the four identical comparison models based on global LOOCV and local LOOCV, respectively. In the global LOOCV, NIRBMMDA obtained better performance with AUC of 0.8666 than HGIMDA (0.7048), IMCMDA (0.6901), KATZMDA (0.8305), MDGHIMDA (0.8518) (see Fig. 3). In the local LOOCV, the AUC of NIRBMMDA is 0.8512, which is better than HGIMDA (0.7537), IMCMDA (0.7425), KATZMDA (0.8216), MDGHIMDA (0.8509) (see Fig. 3).

For MDAD dataset, based on five-fold cross validation, NIRBMMDA obtained the best AUC and standard deviation of 0.9248 ±0.0014 when weight of NI is 0.8 and weight of RBM is 0.2 (see Table 2). As comparison algorithms, AUCs and the standard deviation of HGIMDA (0.8152 ±0.0012), IMCMDA (0.7849 ±0.0025), KATZMDA (0.9173 ±9.6340e−04) and MDGHIMDA (0.8153 ±0.0019) are less than the evaluation result of NIRBMMDA. Then, based on weights used in five-fold cross validation, we calculated AUCs of global LOOCV and local LOOCV for NIRBMMDA, HCIMDA, IMCMDA, KATZMDA and MDGHIMDA, respectively. As a result, in the global LOOCV, NIRBMMDA obtained the AUC with 0.9413, which is better than HCIMDA (0.8173), IMCMDA (0.7891), KATZMDA (0.9247) and MDGHIMDA (0.8446) (see Fig. 4). In the local LOOCV, the AUCs of HCIMDA (0.8301), IMCMDA (0.8035), KATZMDA (0.9119) and MDGHIMDA (0.8537) are less than NIRBMMDA (0.9204) (see Fig. 4).

For aBiofilm dataset, based on five-fold cross validation, the NIRBMMDA obtained the best AUC and the standard deviation with 0.9369 ±0.0020 when NI’s weight is 0.8 and RBM’s weight is 0.2 (see Table 2). As comparison algorithms, HGIMDA (0.8412 ±0.0014), IMCMDA (0.7509 ±0.0073), KATZMDA (0.9305 ±8.0311e−04), MDGHIMDA (0.8201 ±0.0022) are less than NIRBMMDA (0.9369 ±0.0020). Subsequently, based on the selected two weights mentioned above, we computed the AUCs of global LOOCV and local LOOCV for NIRBMMDA, HCIMDA, IMCMDA, KATZMDA and MDGHIMDA, respectively. In the global LOOCV, NIRBMMDA derived an AUC of 0.9557, which is higher than HCIMDA (0.8482), IMCMDA (0.7584), KATZMDA (0.9378) and MDGHIMDA (0.8491) (see Fig. 5). In the local LOOCV, NIRBMMDA obtained better AUC with 0.9414 than AUCs of other four classical models for HCIMDA, IMCMDA, KATZMDA and MDGHIMDA with 0.8837, 0.7718, 0.9302 and 0.8707 respectively (see Fig. 5).

In summary, NIRBMMDA obtained better prediction accuracy compared with four state-of-the-art models for datasets of DrugVirus, MDAD and aBiofilm based on LOOCV and five-fold cross validation. These results indicated that NIRBMMDA has an outstanding and stable performance in predicting potential microbe-drug associations.

Discussing parameters of model

For NIRBMMDA, there are some key parameters needed to be determined including threshold σ_upper used in NI. According to the previous study (Zhang et al., 2016), we tested 10 candidate values of σ_upper with a range from 0.05 to 0.5 (step 0.05) and calculated corresponding 10 AUPR scores based on five-fold cross validation on the training samples. After that, the σ_upper with the best AUPR score was selected to predict potential microbe-drug associations based on test sample.

Similarly, for another parameter s used in RBM, we tested 11 candidate values ranging from 20 to 120 with step size 10, computed corresponding 11 AUPR scores and employed the s with the best AUPR score to carry out prediction. Moreover, visible layer units bias b_i, hidden layer units bias c_j as well as weight w_ij between i − th visible layer unit v_i and j − th hidden layer unit h_j are also used in RBM. The contrastive divergence (CD) algorithm (Hinton, 2002) is employed to determined b_i, c_j and w_ij on the basis of training temples. The specific process of the CD algorithm is illustrated in Table 3.

Case studies

To further validate the prediction performance of NIRBMMDA, we implemented three type of case studies. In first type of case study, we predicted potential drugs for SARS-COV-2 through implementing NIRBMMDA based on the DrugVirus dataset. For the second and third type of case studies, we predicted potential microbes for ciprofloxacin and minocycline through implementing NIRBMMDA based on the MDAD dataset and the aBiofilm dataset, respectively. Our main research interest is in computational bioinformatics. Therefore, we usually confirmed the predicted results presented in case study by databases and published literatures. For some predicted association information that is not validated by any study, we hope the predicted associations can be further confirmed by biologist based on biological experiments in the future.

SARS-COV-2 is a kind of coronavirus with high transmission efficiency, which emerged at the end of 2019 and posed a huge threat to human health (Hu et al., 2021; Hui et al., 2020; Wu, Leung & Leung, 2020). SARS-CoV-2 can cause severe respiratory lesions and lung damages after entering cells (Zhu et al., 2020). Therefore, it is an urgent need to found effective drugs for SARS-CoV-2. Hu, Frieman & Wolfram (2020) found that chloroquine may have effect for treating for COVID-19 caused by SARS-CoV-2 based on study of nanomedicine. Moreover, Shannon et al. (2020) found that favipiravir can exerts an antiviral effect for SARS-CoV-2 by slowing RNA synthesis. Here, we used NIRBMMDA to predict potential drug for SARS-COV-2. Then, we ranked predicted drugs for SARS-COV-2 based on predicted score and further verified the top 20 potential drugs by finding literatures on PubMed. The result showed 13 of the first 20 predicted drugs for SARS-COV-2 were verified (see Table 4). For example, the association between SARS-COV-2 and idoxuridine was predicted and ranked fourth. Idoxuridine is a nucleoside analog and have been used as an antiviral drug for herpes (Almalki et al., 2021). Almalki et al. (2021) found that idoxuridine has significant antiviral activity for SARS-COV-2 through using molecular docking. Moreover, the association between SARS-COV-2 and glycyrrhizin was predicted and ranked sixth. Glycyrrhizin, also named glycyrrhizic acid, is a bioactive substance extracted from a medicinal herb of glycyrrhiza (He et al., 2019). Yu et al. (2021) found that glycyrrhizin was an efficient and nontoxic anti-SARS-COV-2 drug by using computer-aided drug design and biological verification.

Ciprofloxacin, second generation fluoroquinolone, shows outstanding antimicrobial activity with few side effects for treating bacterial infections over 30 years (Zhang et al., 2018). In this article, we employed NIRBMMDA to predict ciprofloxacin-related microbes. Then, we ranked the ciprofloxacin-related microbes according to predicted score and confirmed the top 20 potential associated microbes for ciprofloxacin by finding the literature on PubMed. The result showed that 17 out of the top 20 ciprofloxacin-related microbes were confirmed (see Table 5). For example, the top-ranked microbe for ciprofloxacin is Serratia marcescens. Serratia marcescens, a Gram-negative and non-sporulating bacillus, could cause lung infection, otitis and sepsis (Veraldi & Nazzaro, 2016). Veraldi & Nazzaro (2016) found ciprofloxacin can treat skin ulcers caused by Serratia marcescens through investigating three patients in hospital. Moreover, the association between Mycobacterium avium and ciprofloxacin was predicted and ranked third. Mycobacterium avium is an environmental microbe which exists in water, soil, bird and mammal hosts (Sangari, Parker & Bermudez, 1999). Klopman et al. (1993) found ciprofloxacin show activity against the Mycobacterium avium by using the microdilution method.

Minocycline, second generation tetracycline derivative, has good antibacterial activity (Jonas & Cunha, 1982; Nagarakanti & Bishburg, 2016). In addition, minocycline has been found to have non-antibiotic effects for inflammatory diseases based on open clinical trials (Garrido-Mesa, Zarzuelo & Gálvez, 2013). Particularly, minocycline has emerged effect in neuroprotection demonstrated by various studies in animal models (Garrido-Mesa, Zarzuelo & Gálvez, 2013; Romero-Miguel et al., 2021). Therefore, minocycline has been used for treating acne and could be a potential drug for neurodegenerative and inflammatory diseases such as dermatitis, Parkinson’s disease and Alzheimer’s disease (Garrido-Mesa, Zarzuelo & Gálvez, 2013; Romero-Miguel et al., 2021). In this case study, via the implementation of NIRBMMDA, we can predict potential microbes associated with drug of minocycline. Subsequently, we sorted predicted microbes for minocycline according to the predicted score and verified the top 20 potential microbes by finding the published literature. The result showed that 17 out of the top 20 microbes for minocycline were confirmed (see Table 6). Among the top 20 predicted microbes for minocycline, Pseudomonas aeruginosa was predicted with the first ranking. Pseudomonas aeruginosa, a common Gram-negative bacterium, can lead to severe infections for human (Chevalier et al., 2017). Chen et al. (2019) found that minocycline possessed antimicrobial activity for Pseudomonas aeruginosa in vitro experiment. Furthermore, the association between Streptococcus mutans and minocycline was predicted and ranked third. Streptococcus mutans possesses strong virulence factors including high acid production, ability to form compact biofilm and production of glucans (Abdel-Aziz, Emam & Raafat, 2020). Baker et al. (1983) found that minocycline can inhibit plaque formation caused by Streptococcus mutans in vitro pure cultures.