Neighborhood-based inference and restricted Boltzmann machine for small molecule-miRNA associations prediction

SM-miRNA associations

In this article, the known SM-miRNA associations information was obtained from SM2miR v1.0 (Liu et al., 2013), and we constructed two different SM-miRNA associations dataset of dataset 1 and dataset 2 (Jiang et al., 2009; Knox et al., 2011; Lu et al., 2008; Ruepp et al., 2010; Wang et al., 2009). Dataset 1 contains 664 known SM-miRNAs associations between 831 SMs and 541 miRNAs, in which 792 (out of 831, 95%) SMs do not have any known associated miRNAs according to SM2miR. By removing the information of SMs (miRNAs) that have any no known related miRNAs (SMs) in dataset 1, we constructed dataset 2 consists of 39 SMs and 286 miRNAs that are fully involved in the 664 known associations. Unsurprisingly, dataset 2 contains 664 known SMs-miRNAs associations between 39 SMs and 286 miRNAs. Moreover, an adjacency matrix A ∈ R^ns×nm was built to denote the associations between SMs and miRNAs, where ns represents number of SMs and nm indicates number of miRNAs. If the SM s(i) is associated with miRNA m(j), the entry A(i, j) is equal to 1, otherwise 0.

Integrated SM similarity

Referring to the earlier study (Lv et al., 2015), the integrated SM similarity was obtained by combining four types of SM similarity including SM chemical structure similarity $S_S^C$ (Hattori et al., 2003), disease phenotype-based similarity for SMs $S_S^D$ (Gottlieb et al., 2011), SM side effect similarity $S_S^S$ (Gottlieb et al., 2011) and gene functional consistency-based similarity $S_S^T$ (Lv et al., 2012). In this study, the SIMilar COMPound (SIMCOMP) approach was utilized to calculate the similarity between SM chemical structures based on SM chemical structure information (Hattori et al., 2003). Then, the Jaccard score was employed, utilizing information about diseases associated with each small molecule to assess the similarity of SMs based on disease phenotypes (Gottlieb et al., 2011). SM side effect similarity was gained with the method of Jaccard score (Gottlieb et al., 2011) based on SM-side effect association dataset. Gene functional consistency-based similarity was derived with Gene Set Functional Similarity (GSFS) (Lv et al., 2012) method on the dataset of SMs’ target genes. After that, we used the weighted average approach to calculate the four kind of SM similarity for obtaining integrated SM similarity, which can be described as follows. (1)${\displaystyle S_S=\frac{{\beta }_1{S}_S^D+{\beta }_2{S}_S^T+{\beta }_3{S}_S^C+{\beta }_4{S}_S^S}{\sum _{i=1}^4{\beta }_i}}$ where values of β_i(i = 1, 2, 3, 4) are set as 1 and S_S denotes integrated SM similarity.

Integrated miRNA similarity

Similarly, we obtained integrated miRNA similarity by merging disease phenotype-based similarity for miRNAs $S_M^D$ (Gottlieb et al., 2011) with gene functional consistency-based similarity for miRNAs $S_M^T$ (Lv et al., 2012). Specifically, we utilized the GSFS method to determine the functional consistency-based similarity of miRNA target gene sets (Lv et al., 2012). Furthermore, using the Jaccard score, we calculated the disease phenotype-based similarity for miRNAs based on the dataset of miRNA-related diseases (Gottlieb et al., 2011). We employed a weighted average approach to determine the integrated miRNA similarity SM, which is outlined below. (2)${\displaystyle SM=\frac{{\alpha }_1{S}_M^D+{\alpha }_2{S}_M^T}{\sum _{j=1}^2{\alpha }_j}}$ where values of α_j(j = 1, 2) are set as 1.

NIRBMSMMA

In this study, we presented a novel computational model, NIRBMSMMA, to identify potential SM-miRNA associations by leveraging known SM-miRNA associations, as well as integrated SM similarity and integrated miRNA similarity. First, we employed NI technique to anticipate potential associations between SMs and miRNAs by varying thresholds to compute the sum associations between the miRNAs (SMs) and neighbors of SM (miRNA). Second, a potential SM-miRNA association matrix was predicted using the contrastive divergence algorithm and sigmoid function of the restricted Boltzmann machine (RBM). Considering single method may be weak in generalization ability, ensemble learning was employed to integrate the predictions from NI and RBM to derive the final potential associations between SMs and miRNAs. The flowchart of NIRBMSMMA is shown as Fig. 1.

Neighborhood-based inference

Neighborhood-based method is a classic collaborative filtering algorithm, also referred to as memory-based algorithm, can recommend similar items for a user based on behaviors of similar users (Aggarwal, 2016). Here, we constructed a based model of NI to identify potential SM-miRNA associations. First, we built a NI model utilizing the integrated SM similarity. For a SM s_{i,i=1,2,…,ns}, by applying a threshold σ based on integrated SM similarity, the neighbors can be filtered, and subsequently, a defined set of neighbors can be established as {s_i|ISMS_d,i > σ, d ≠ i}. After that, the potential score of SM s_i associated with miRNA m_j can be gained by calculating the sum associations between the miRNA m_j and neighbors of SM s_i as follows. (3)${\displaystyle scor{e}_{i,j}^S=\frac{\sum _{i=1}^{ns}\sum _{j=1,d\ne i,ISM{S}_{d,i}\ge \sigma }^{nm}{A}_{d,j}\times ISM{S}_{d,i}}{\sum _{i=1}^{ns}\sum _{j=1,d\ne i,ISM{S}_{d,i}\ge TS}^{nm}ISM{S}_{d,i}}}$ where ns denotes the count of SM, nm denotes the count of miRNA, A_d,j denotes the entry between SM s_d and miRNA m_j in SM-miRNA associations matrix A. The ISMS_d,i denotes integrated SM similarity between SM s_d and SM s_i. The $scor{e}_{i,j}^S$ represents potential association score between SM s_i and miRNA m_j.

NI model was constructed by filtering neighbors with the threshold σ. To reduce bias of neighbor selection, we generated multiple thresholds {σ₁, σ₂, …, σ_n} in the range of 0.05 to 0.5 to construct multiple basic NI models. Particularly, an upper bound parameter σ_upper was employed to determine the multiple thresholds which were represented as σ_threshold = {σ_i|σ_i ≤ σ_upper, i = 1, 2, …, n(n = |σ_threshold|)}. Thus, n thresholds {σ₁, σ₂, …, σ_n} were employed to construct n basic NI models.

After that, the potential score for SM-miRNA associations was derived by integrating n basic NI models as follows.

Then, we integrated ns basic models to predict potential SMs-miRNA associations score by using average strategy, which can be described as follows: (4)${\displaystyle M\text{\_}ISMS=\frac{\sum _{k=1}^{n}scor{e}^S\text{\_}k}{n}}$ where n denotes the number of basic NI models, score^S_k denotes predicted SM-miRNA associations score based on the k − th threshold and M_ISMS is potential SM-miRNA associations score based on integrated SM similarity.

Second, we built a NI model based on integrated miRNA similarity by using the method above. For a miRNA m_{j,j=1,2,…,nm}, its neighbors can be filtrated by using a threshold σ based on integrated miRNA similarity and set of neighbors can be was defined as M_j|IMS_t,j > σ, t ≠ j. On the basis of the set of neighbors above, potential association score $scor{e}_{i,j}^M$ between SM s_i and miRNA m_j was obtained through computing the sum of associations between SM s_i and neighbors of miRNA m_j as Eq. (5). (5)${\displaystyle scor{e}_{i,j}^M=\frac{\sum _{i=1}^{ns}\sum _{j=1,t\ne i,IM{S}_{t,j}\ge \sigma }^{nm}{A}_{i,t}\times IM{S}_{t,j}}{\sum _{i=1}^{ns}\sum _{j=1,t\ne j,IM{S}_{t,j}\ge \sigma }^{nm}IM{S}_{t,j}}}$ where ns represents number of SM, nm represents number of miRNA, A_i,m denotes the entry between SM s_i and miRNA m_m in SM-miRNA association matrix A. The IMS_t,j represents integrated miRNA similarity between miRNA m_t and miRNA m_j.

To mitigate bias stemming from neighbor selection, we employed several distinct thresholds to construct multiple basic NI models. Next, we combined the multiple basic NI models to generate score for potential SM-miRNA associations, as outlined below. (6)${\displaystyle M\text{\_}IMS=\frac{\sum _{k=1}^{n}scor{e}^M\text{\_}k}{n}}$ where n denotes the number of basic NI models, score^M_k denotes predicted SM-miRNA associations score based on k − th threshold and M_IMS denotes potential SM-miRNA associations score based on integrated miRNA similarity.

Finally, the potential SM-miRNA association score S_NI was acquired by integrating NI models based on integrated SM similarity and NI models based on integrated miRNA similarity as follows. (7)${\displaystyle S_{NI}=\frac{M\text{\_}ISMS+M\text{\_}IMS}{2}}$

Restricted Boltzmann Machine (RBM)

RBM is probabilistic graphical model, also called stochastic neural network, can extract useful features from input data (Fischer & Igel, 2012; Zhang et al., 2018). In the previous decade, RBM has been extensively studied and employed in many fields including quantum physics (Melko et al., 2019), bioinformatics (Wang & Zeng, 2013) and text mining (Monti et al., 2016). In this article, we used RMB as a based model to predict potential SM-miRNA associations. As shown in Fig. 2, the RBM can be described as a two-layered neural network, consisting of visible layers and hidden layers, each of which is made up of many nodes (units). In an RBM, if we take into account the total number of nodes, there are nm in the visible layer and s nodes in the hidden layer. We employed v = (v_i, v₂, …, v_nm) to represent set of visible layer nodes and used h = (h₁, h₂, …, h_s) to represent set of hidden layer nodes. As there are no connections within the visible or hidden layer, the energy function between v and h can be expressed in the following manner. (8)${\displaystyle E\left(\mathbf{v},\mathbf{h}\right)=-\sum _{i=1}^{nm}{b}_i{v}_i-\sum _{j=1}^s{c}_j{h}_j-\sum _{i=1}^{i=nm}\sum _{j=1}^s{w}_{i,j}{v}_i{h}_j}$ where nm represents number of visible layer nodes, s denotes number of visible layer nodes, b_i denotes bias of i − th visible layer node v_i, c_j denotes bias of j − th hidden layer node h_j and w_ij denotes the weight between v_i and h_j.

Based on Eq. (8), the potential association probability of v and h can be defined as follows. (9)${\displaystyle P\left(\mathbf{v},\mathbf{h}\right)==\frac{1}{Z}{e}^{-E\left(\mathbf{v},\mathbf{h}\right)}}$ where Z = ∑_v,h − E(v, h) is called partition function.

Then, we can obtain marginal distribution of visible layer nodes by following equation. (10)${\displaystyle P\left(\mathbf{v}\right)=\sum _{\mathbf{h}}P\left(\mathbf{v},\mathbf{h}\right)=\frac{1}{Z}\sum _{\mathbf{h}}{e}^{-E\left(\mathbf{v},\mathbf{h}\right)}.}$

Since distributions of nodes of visible or hidden layer are independent, their conditional probabilities can be defined, respectively, as follows. (11)${\displaystyle P\left(v_i=1|\mathbf{h}\right)=\sigma \left(\sum _{j=1}^s{w}_{ij}{h}_j+b_i\right)}$ (12)${\displaystyle P\left(h_j=1|\mathbf{v}\right)=\sigma \left(\sum _{i=1}^{nm}{w}_{ij}{v}_i+c_j\right)}$ where σ(x) = 1/(1 + e^−x) is the logistic function.

Based on a dataset containing known SM-miRNA associations between nm miRNAs and ns SMs, an RBM is created for predicting potential SM-miRNA associations, with nm nodes in the visible layer and s nodes in the hidden layer. Each SM is associated with an observation sm_i = {A_i,1, A_i,2, …, A_i,nm}, which has a binary value indicating whether the SM s_i is associated with nmmiRNAs. Based on known SM-miRNA association matrix A ∈ R^ns×nm, we can obtain ns observations with nm dimension. The observation of an SM would be utilized as input for the RBM when predicting new miRNAs for it. Then, the prediction is implemented as two steps below. (13)${\displaystyle P_j=P\left(h_j=1|A_{i,1},A_{i,2},\dots ,A_{i,nm}\right)=\sigma \left(\sum _{t=1}^{nm}{w}_{ij}{A}_{i,t}+c_j\right),j=1,2,\dots ,s}$ (14)${\displaystyle S_{RBM}^i=P\left(v_i|P_1,P_2,\dots ,P_s\right)=\sigma \left(\sum _{j=1}^s{w}_{ij}{P}_i+b_i\right),i=1,2,\dots ,nm}$ where σ(x) = 1/(1 + e^−x) is the logistic function. $S_{RBM}^i$ is predicted score between SM s_i and nm miRNAs. Finally, we obtained the potential SM-miRNA association scores and defined S_RBM to save scores as follows: (15)${\displaystyle S_{RBM}=S_{RBM}^1,S_{RBM}^2,\dots ,S_{RBM}^{ns}}$

Ensemble learning

Because single predictor usually has poor generalization performances, ensemble learning is frequently used to combine multiple weak predictors and improve generalization performance. Dong et al. (2020). In the past decades, ensemble learning as one research hot spot have been studied and employed in many fields including data stream classification (Gomes et al., 2017), statistical mechanics (Krogh & Sollich, 1997), sentiment classification (Wang et al., 2014) and bioinformatics (Chen, Zhou & Zhao, 2018). Furthermore, NI and RBM were employed as individual base predictors for ensemble learning to increase the precision of predictions of potential SM-miRNA associations. To ensure the predicted score in range of 0 to 1, we normalized predicted scores of NI and RBM as follows: (16)${\displaystyle S1=\frac{\left(tanh\left(0.1\times \left(\frac{S_{NI}-{\mu }_1}{{\sigma }_1}\right)\right)+1\right)}{2}}$ (17)${\displaystyle S2=\frac{\left(tanh\left(0.1\times \left(\frac{S_{RBM}-{\mu }_2}{{\sigma }_2}\right)\right)+1\right)}{2}}$ where μ₁ and σ₁ refer to the mean and the standard deviation of the scores acquired by the NI. Likewise, μ₂ and σ₂ represent the mean and the standard deviation of the scores achieved by the RBM. After that, we used average strategy to integrate NI and RBM as follows: (18)${\displaystyle S=0.5\times S1+0.5\times S2}$ where S is final predicted SM-miRNA association score matrix.

Performance evaluation

We utilized global LOOCV, miRNA-fixed local LOOCV and SM-fixed local LOOCV as well as five-fold cross validation to evaluate the predictive ability of NIRBMSMMA by using two datasets based on SM2miR database (Liu et al., 2013). Dataset 1 contains 644 known SM-miRNA associations between 831 SMs and 541 miRNAs. After removing SMs (miRNAs) that have any no known related miRNAs (SMs) in dataset 1, we constructed dataset 2 with 644 known SM-miRNA associations between 39 SMs and 286 miRNAs.

In the LOOCV, each known SM-miRNA association was taken as a test sample in rotation, while the other 633 known SM-miRNA associations were employed as training samples. All unknown SM-miRNA pairs were considered as potential candidate samples. After that, we employed the training samples to train NIRBMSMMA and applied trained NIRBMSMMA to predict scores of test sample and candidate samples. In global LOOCV, we ranked score of test sample according to the scores of all candidate samples. In the SM-fixed LOOCV, based on predicted scores, we sorted test sample with candidate samples that involved fixed SMs. Similarly, in the miRNA-fixed LOOCV, we compared score of test sample with scores of candidate samples that contained fixed miRNAs. Next, the receiver operating characteristics (ROC) curves (AUCs) were plotted by using true positive rate (TPR, sensitivity) as abscissa axis and false positive rate (FPR, 1-specificity) as ordinate axis based on different thresholds. TPR denotes the percentage of candidate samples whose ranking over given threshold, while FPR is the portion of candidate samples with lower rankings than the given threshold. Then, AUCs were computed to compare the performance of NIRBMSMMA with four classic prediction models of TLHNSMMA (Qu et al., 2018), SMiR-NBI (Li et al., 2016), GISMMA (Guan et al., 2018) and SLHGISMMA (Yin et al., 2019). If the value of AUC is equals to 1, it means that NIRBMSMMA has a perfect prediction ability. If the value of AUC is 0.5, it means that prediction of NIRBMSMMA is random. As a result, in global LOOCV, the performance comparisons with some classical algorithms showed that NIRBMSMMA achieved an AUC of 0.9912 and is obviously better than TLHNSMMA (0.9859), SMiR-NBI (0.8843), GISMMA (0.9291), SLHGISMMA (0.9273) based on datasets 1, which was shown in Fig. 3. NIRBMSMMA achieved an AUC of 0.8645 and is obviously better than TLHNSMMA (0.8149), SMiR-NBI (0.7264), GISMMA (0.8203) and SLHGISMMA (0.7897) based on dataset 2 shown in Fig. 3. In the miRNA-fixed local LOOCV, NIRBMSMMA derived an outstanding performance with AUCs of 0.9875, which is better than TLHNSMMA (0.9845), SMiR-NBI (0.8837), GISMMA (0.9505) and SLHGISMMA (0.9553) based on datasets 1 shown in Fig. 4. NIRBMSMMA derived AUC of 0.8720, which is better than TLHNSMMA (0.8244), SMiR-NBI (0.7846), GISMMA (0.8203) and SLHGISMMA (0.8106) based on dataset 2 shown in Fig. 4. In the SM-fixed local LOOCV, NIRBMSMMA gained AUCs of 0.8394, which is better than TLHNSMMA (0.7645), SMiR-NBI (0.7497), GISMMA (0.7702) and SLHGISMMA (0.7702) based on dataset 1 shown in Fig. 5. In the SM-fixed local LOOCV, NIRBMSMMA gained AUCs of 0.7066, the AUCs of other comparison algorithms were TLHNSMMA (0.6057), SMiR-NBI (0.6100), GISMMA (0.6591) and SLHGISMMA (0.6565) based on dataset 2 shown in Fig. 5.

We further employed five-fold cross validation to evaluate the performance of the model. In the five-fold cross validation, 664 known SM-miRNA associations were randomly divided into five portions where four portions contain 133 known SM-miRNA associations, respectively, remaining portion contains 132 known SM-miRNA associations. Each portion was used as test sample in turn and remaining four portions were employed as training samples. Similarly, all unknown SM-miRNA pairs were considered as candidate samples. Additionally, we utilized the training samples to train NIRBMSMMA and applied trained NIRBMSMMA to predict scores of test sample and candidate samples. It’s worth noting that we repeated the five-fold cross-validation process 100 times to avoid bias from the sample division. Consequently, during the five-fold cross-validation process, the AUCs and standard deviations of NIRBMSMMA is 0.9898 ± 0.0009, which is better than TLHNSMMA (0.9851 ± 0.0012), SMiR-NBI (0.8554 ±  0.0063), GISMMA (0.9263 ± 0.0026), SLHGISMMA (0.9446 ±  0.0036) based on dataset 1. Simultaneously, the AUCs and standard deviations of NIRBMSMMA is 0.8547 ± 0.0046, which is better than TLHNSMMA (0.8168 ± 0.0022), SMiR-NBI (0.7104 ±  0.0087), GISMMA (0.8088 ± 0.0044), SLHGISMMA (0.7276 ±  0.0045) based on dataset 2.

Dataset 1 and dataset 2 contain the same information of 664 known SM-miRNA associations. The results showed that NIRBMSMMA have better performance in dataset 1 than dataset 2 based on cross validation. The reason is that datasets 1 and dataset 2 have the same positive samples (divided into training and testing samples), but compared to dataset 2, dataset 1 has more candidate samples. Candidate samples in dataset 1 contain more SMs/miRNAs that have any no known associations with miRNAs/SMs. Therefore, the predicted scores of candidate samples in dataset 1 are relatively low compared to the test samples, resulting in higher AUC values based on dataset 1 than dataset 2. Unsurprisingly, adding more known SM-miRNA associations to the dataset 1 is expected to improve the accuracy of NIRBMSMA.

In addition, we further collected 130 known SM-miRNA associations by searching the latest references. Subsequently, the collected dataset was integrated into dataset 1 to build a new dataset of dataset 3. Therefore, dataset 3 contains 794 known SM-miRNA associations between 831 SMs and 531 miRNAs. To evaluate whether the proposed model is applicable to the new dataset, we further implemented LOOCV and 5-fold cross validation based on the new built dataset 3 that was constructed according to dataset 1 and latest references. The results showed that in the global LOOCV, NIRBMSMMA obtained AUC of 0.9884, which is better than TLHNSMMA (0.9831), SMiR-NBI (0.8994), GISMMA (0.9228) and SLHGISMMA (0.9150) shown in Fig. 3. In the miRNA-fixed local LOOCV, NIRBMSMMA and TLHNSMMA derived AUC of 0.9802 0.9814 respectively, which is better than SMiR-NBI (0.8984), GISMMA (0.9338) and SLHGISMMA (0.8594) shown in Fig. 4. In the SM-fixed local LOOCV, NIRBMSMMA gained AUC of 0.8239, which is better than TLHNSMMA (0.7600), SMiR-NBI (0.7568), GISMMA (0.7700) and SLHGISMMA (0.7664) shown in Fig. 5. In the five-fold cross validation, 794 known SM-miRNA associations were randomly divided into five portions where four portions contain 159 known SM-miRNA associations respectively, remaining portion contains 158 known SM-miRNA associations. Similarity, during the five-fold cross-validation process, the AUCs and standard deviations of NIRBMSMMA is 0.9870 ± 0.0015, which is better than TLHNSMMA (0.9818 ± 0.0016), SMiR-NBI (0.8720 ± 0.0066), GISMMA (0.9185 ± 0034), SLHGISMMA (0.8013 ± 0.1151).

Discussing parameters of model

In the NIRBMMMA, there are some parameters, including threshold σ_upper used in NI, play a crucial role in predictive performance and needed to be determined. According to previous work (Zhang et al., 2016), we tested 10 different values of σ_upper ranging from 0.05 to 0.5 with step 0.05 and computed corresponding 10 AUPR values by employing five-fold cross validation on the training samples. Subsequently, the value of σ_upper that produced the highest AUPR value was selected and adopted as the optimal parameter to identify new SM-miRNA associations on the basis of test sample.

Likewise, for parameter of number of hidden layer node s used in RBM, we tested 11 candidate values in the range 20 to 120 with step 10, calculated corresponding 11 AUPR values and used S with the best AUPR value to implement prediction. Besides, visible layer node bias b_i, hidden layer node bias c_j as well as weight w_ij between i − th visible layer node v_i and j − th hidden layer node h_j also play important roles in RBM. We used contrastive divergence (CD) algorithm (Hinton, 2002) to determined b_i, c_j and w_ij based on training temples.

Case studies

To further evaluate prediction performance of NIRBMSMMA, we implemented two different types of case studies. For the first type of case studies, we used NIRBMSMMA to predict potential association scores for all unknown SM-miRNA pairs. Then, based on predicted scores, we sorted predicted SM-miRNA associations in descending order and confirmed the top 50 predicted SM-miRNA associations by searching the published literature on PubMed. As a result, five (10) out of the top 20 (50) potential SM-miRNA associations were confirmed (see Table 1). For example, the association between gemcitabine and mir-24-2 was predicted and ranked third. An experiment implemented by Pandita et al. (2015) indicated that mir-24-2 overexpression cells reduced 37%–50% activity when these cells were treated by using gemcitabine with lower than IC₅₀ dose. The association between mir-203a and 5-aza-2′-deoxycytidine ranked the fifth according to predicted scores. Liu et al. (2016) found that mir-203a expression increased when using 5-aza-2′-deoxycytidine to treat esophageal cancer cells. The association between 5-Fluorouracil (5-FU) and let-7c ranked the tenth according to predicted scores. Peng et al. (2015)found that overexpression of let-7c reduce Akt2 expression and under expression of Akt2 can enhance the sensitivity of renal cell to 5-Fluorouracil.

Also, to assess ability of NIRBMSMMA to predict potential miRNAs for new SM that without any known associated miRNA, based on dataset 1, we conducted the second type of case study on the two investigated SMs of 5-FU and Gemcitabine. First, for an investigated SM, we removed all known associated miRNAs for the investigated SM and used NIRBMSMMA to predict potential miRNAs for the investigated SM. After that, we ranked the top 50 predicted associations between investigated SM and miRNAs according to predicted scores.

5-FU is a kind of chemotherapeutical drug that can prevent cell proliferation by inhibiting enzyme thymidylate synthase and has been used to treat several cancers including head-and-neck (H&N), colorectal and breast (Goirand et al., 2018; Tozer et al., 2019; Wigmore et al., 2010). After implementing NIRBMSMMA, we acquired miRNAs that are potentially associated with 5-FU. Then, we sorted potential associations between 5-FU and miRNAs in descending order and confirmed the top 50 potential associations by using SM2miR database and searching the published literature on PubMed. Particularly, in Tables 2 and 3, we employed PubMed ID 26198104 to denote SM2miR database. As a result, 13 (32) out of the top 20 (50) potential miRNAs correlated with 5-FU were confirmed (see Table 2). For example, the mir-874 was predicted to be associated with 5-FU and the associations ranked the ninth. Han et al. (2016) found that mir-874 decrease the resistance of colorectal cancer cells to 5-FU and inhibit proliferation of colorectal cancer cells. The mir-455 was predicted to associate with 5-FU and the association ranked the tenth. An experiment conducted by Hummel et al. (2011) showed that the expression level of mir-455-3p was decreased after treatment of 5-FU by using TargetScan and bioinformatic analysis. The associations between mir-299 and 5-FU ranked the thirty-fourth according to predicted score. Chen, Lu & Hu (2019) found that mir-299 overexpression can increase sensibility of Hexokinase 1 (HK1) cells to 5-FU and further decrease invasion ability of HK1 cells.

Gemcitabine, a pyrimidine nucleoside analog anticancer drug, has potent activity for a wide spectrum of solid tumors (Miao, Chen & Luan, 2020; Mini et al., 2006; Rizzuto, Ghazaly & Peters, 2017). After implemented NIRBMSMMA, we gained potential Gemcitabine-miRNA associations score. After that, we sorted potential Gemcitabine-miRNA associations in descending order and confirmed the top 50 potential associations by searching the published literature on PubMed. Result showed that 11 (32) out of the top 20 (50) potential miRNAs associated with Gemcitabine were confirmed (see Table 3). For example, the mir-17 was predicted to be associated with gemcitabine and the association ranked the twelfth. Yan et al. (2012) found that mir-17-5p inhibitor can enhance sensitivity of pancreatic cancer cells to gemcitabine by upregulating Bim protein expression. Mir-27b was also predicted to associate with gemcitabine. A study implemented by Bera et al. (2014) demonstrated that the expression level of mir-27b is significantly decreased in gemcitabine resistant pancreatic ductal adenocarcinoma cells by testing with quantitative polymerase chain reaction (qPCR) analysis. Moreover, the mir-125a was predicted to be associated with gemcitabine and the association ranked twenty-third. Yao et al. (2016) found that mir-125a can enhance chemo-resistance of pancreatic cancer cells to gemcitabine via targeting A20 gene.

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.