Prediction of microbe-drug associations using a CNN-Bernoulli random forest model

Microbe-drug association

In this article, we introduced two datasets of known microbe-drug associations, separately for abiofilm (Rajput et al., 2018), Microbe-Drug Association Database (MDAD) (Sun et al., 2018). The abiofilm dataset records 2,884 known microbe-drug associations between 1,720 drugs and 140 microbes. Moreover, MDAD dataset collected 2,470 known microbe-drug associations between 1,373 drugs and 173 microbes. The statistics of the two datasets mentioned above are shown in Table 1. Furthermore, the adjacency matrix A ∈ R^nd×nm was used to represent the microbe-drug association information, where nd means the number of drugs and nm means the number of microbes. The value of entity A_ij is set to 1 when the drug d_i was associated with the microbe m_j. Otherwise, the value is set to 0. (1)${\displaystyle A_{ij}=\left\{\begin{array}{c}1,\text{if durg}{d}_i\text{is associated with microbe}{m}_j\hfill \\ 0,\text{otherwise}\hfill \end{array}\right.}$

Drug side effect similarity

The SIDER (Kuhn et al., 2010) database provides extensive information about drug side effects, and we downloaded the side effects of the drugs used in the this study from SIDER. Based on the view that the more side effects shared by two drugs, the more similar the two drugs. The set of side effects associated with drug d_i is defined as N(i). We defined the similarity between two drugs as 0 when there were no common side effects between the two drugs. The entity SS1(i, j) denoted the similarity of side effects between drug d_i and drug d_j. Jaccard (Gottlieb et al., 2011) was utilized to calculate drug side effect similarity Eq. (2) showed as following. (2)${\displaystyle SS1=\text{Jaccard score}=\frac{\left|N\left(i\right)\cap N\left(j\right)\right|}{\left|N\left(i\right)\cup N\left(j\right)\right|}.}$

Drug chemical structure similarity

SIMCOMP (http://www.genome.jp/tools/simcomp/) is a graph-based chemical structure comparison method that can be used to obtain the chemical structure similarity of different drugs. SIMCOMP is used to find maximal cliques in association graphs and furthermore to discover isomorphic maximal common subgraphs (Hattori et al., 2003). In the data pre-processing of this article, SIMCOMP2 quantifies the similarity between drugs by comparing their chemical structures. Specifically, we set the cutoff score for drug structure similarity at 0.5. The drug chemical structural similarity between drug d_i and drug d_j was represented by SS2.

Microbe sequence similarity

In this paper, we obtained two microbe sequence similarity based on two different known microbe-drug association datasets, respectively. For microbes in the MDAD dataset or the abiofilm dataset, microbe sequence similarity obtained by MAFFT and BioEdit calculations. MAFFT introduces the approximate distance calculation algorithm and the fast Fourier alignment algorithm to achieve multiple sequence matching. The sequence similarity matrix can be calculated by using the sequence identification matrix function in BioEdit (Tippmann, 2004). Based on the idea that the more sequences two microbes have in common, the more similar they are to each other. Hence, when two microbes have no sequence in common, the value of the microbe sequence similarity score between the two microbes is equal to 0. In this article, the microbe sequence similarity matrix was represented by the MV matrix, as in which MV(m_i, m_j) denotes the sequence similarity value between microbe m_i and microbe m_j.

Integrated similarity for microbe and drug

To fully leverage the raw information of drugs and microbes, we construct the integrated similarity for microbes using unprocessed microbe sequence similarity. For drugs, we aggregate the drug chemical structure similarity and drug side effect similarity and compute their average to obtain the integrated similarity for drugs. CNNBRF demonstrates outstanding performance by thoroughly exploiting the original data.

Moreover, to ensure consistency in data processing, we exclude the parts of models like NIRBMMDA, LAGCNMDA, and RFMDA that utilize Gaussian kernel similarity. We observe a significant decrease in performance when these parts are removed, while our model maintains its robustness. This resilience is one of the strengths of our model.

CNNBRFMDA

In this article, we propose a new computational model named CNNBRFMDA to predict potential microbe-drug associations. The flowchart of CNNBRFMDA is shown in Fig. 1. The proposed method consists of three main parts: Data preparation, CNN for latent feature extraction and BRF for predicting potential microbe-drug association.

Data preparation

Data preparation is the first step of the CNNBRFMDA algorithm, aimed at representing microbe-drug pairs as feature vectors. As previously mentioned, in our study, we constructed the adjacency matrix A of known microbe-drug association, the integrated drug similarity matrix SD(nd × nd) and the integrated microbe similarity matrix SM(nm × nm). We extracted nm and nd features separately for microbe and drugs from these matrices, respectively. By concatenating the feature vectors of the investigated microbes and drugs, we obtained nm + nd features for each microbe-drug pair. For any one of the two datasets, each drug is comprised of nd features (f₁, f₂, …, f_x…, f_nd), where the x − th feature f_x, represents the integrated similarity between the investigated drug and the x − th drug. Therefore, drug d_u can be represented by nd-dimensional vector FeV(d_u), which can be described as Eq. (3): (3)${\displaystyle FeV\left(d_u\right)=\left(f{\left(u\right)}_1,f{\left(u\right)}_2,\dots ,f{\left(u\right)}_x,\dots ,f{\left(u\right)}_{nd}\right)}$

where the x − th element f(u)_x of FeV(d_u) is the integrated similarity between drug d_x and drug d_u. Similarity, we denoted the nm-dimensional vector FeV(m_v) for microbe m_v as Eq. (4): (4)${\displaystyle FeV\left(m_v\right)=\left(f{\left(v\right)}_1,f{\left(v\right)}_2,\dots ,f{\left(v\right)}_y,\dots ,f{\left(v\right)}_{nm}\right)}$

where the y − th element f(v)_y of FeV(m_v) is the integrated similarity between microbe m_y and m_v.

Subsequently, by concatenating the feature vectors of FeV(m_v) and FeV(d_u), we obtained nm + nd features for each microbe-drug pair. As a result, there were nm × nd samples and the length of each sample was nm + nd, each sample corresponds to a microbe-drug pair. Next, we set the labels of the known microbe-drug association to 1, forming the positive sample set P. Simultaneously, negative samples were selected from the unlabeled microbe-drug pairs, and their labels were set as 0 to create a negative sample set N_CNN, ensuring a balanced 1:1 ratio of positive sample set and negative sample set. This method is feasible due to the low ratio of $\frac{\left(\text{the number of P}\right)}{nm\times nd-\left(\text{the number of P}\right)}$. After compiling P and N, the next step was combined them into a training set to construct low- dimension feature vectors.

Convolutional neural network for latent feature extraction

CNN is a deep learning algorithm (LeCun et al., 1989), the three basic layers of CNN are the convolutional layer, the pooling layer (the subsampling layer), and the fully connected layer. The convolutional layer obtains different features of the input data by learning filters, and then the pooling layer is used to reduce the feature dimension, enhance the translation invariance and robustness of the model, and finally classify in the fully connected layer. In the CNNBRFMDA model, we utilized CNN to extract complex features from microbe-drug feature vectors in a deep-learning framework. Specifically, we perform multiple convolution operations on input samples using different kernels to generate activation maps. The activation map Q_l at layer l can be described as: (5)${\displaystyle Q_l=\sigma \left(Q_{l-1}\odot W_l+b_l\right)}$

where σ(k) denotes the activation function, W_l the convolution kernel at layer l, b_l the offset vector, and ⊙ the convolution operation. Subsampling layers are used to compress the feature vectors of microbe-drug pairs and minimize overfitting. The formula for the subsampling layer Q_l can be expressed as: (6)${\displaystyle Q_l=Subsampling\left(Q_{l-1}\right).}$

In order to preserve the most salient features of each filtered region, we employ max-pooling in the subsampling layer. During the training process, CNN was optimized to minimize the network’s loss function. The training data was fed into CNN, enabling it to capture salient features from the input feature vectors of microbe-drug pairs. We conducted several experiments to identify the optimal hyperparameters of the CNN and obtain the best predictive model. For the convolutional operation, we incorporated 16 filters with a size of 1 × 16 in the convolutional layer, based on the configuration used in Deepthi, Jereesh & Liu (2021), which has been shown to perform effectively for similar tasks. In the subsampling layer, we established a filter size of 1 × 2. In our model, At the convolution and fully connected layers, we employed the rectified linear unit (ReLU (Nair & Hinton, 2010)) as the activation function. Additionally, the sigmoid function was used at the output layer. The model was implemented with the binary cross-entropy error function and optimized using Adam. Furthermore, to avoid overfitting, two dropout layer (Srivastava et al., 2014) are added between the convolutional and fully connected layers. Finally, we retrieved the weight data learned after multiple convolution and pooling operations. Ultimately, we updated the feature vectors of all microbe-drug pairs using the weight data learned from CNN. Specifically, we used the output data of the second-to-last layer (fully connected layer) as the low-dimension feature vectors for all microbe-drug pairs. Here, we reduced the dimensionality to $\frac{nd+nm}{2}$.

The model prevents overfitting through several strategies, including the use of dropout layers that randomly deactivate neurons to reduce model dependency on any specific set of features, as well as employing five-fold cross-validation. This cross-validation technique helps ensure that the model’s performance is robust across different subsets of the data, further mitigating the risk of overfitting.

It is worth mentioning that the model consists of only three layers with trainable parameters: one convolutional layer and two fully connected layers. This streamlined architecture significantly reduces computational complexity, decreases memory usage, and accelerates the training process, making it exceptionally efficient for deployment in environments with limited computational resources. Furthermore, the simplicity of having fewer trainable layers reduces the risk of overfitting and streamlines the model tuning and optimization process, ensuring robust performance across diverse datasets. Consequently, the model remains both practical and effective for applications that demand rapid processing with reliable accuracy.

Bernoulli random forest based prediction

As a learning algorithm, random forest (RF) is widely used in various prediction tasks and has achieved notable performance (Chen et al., 2018; Wang et al., 2019). This technique involves three key processes: random bootstrap sampling, random attribute bagging, and tree node splitting.

First, random bootstrap sampling is a method of sampling with replacement. For each tree, the random forest algorithm selects a subset of the training data, known as a “bootstrap sample”, which may contain duplicate data points. Second, random attribute bagging involves selecting a random subset of features for each node split rather than using all available features. Finally, tree node splitting is the core step in constructing decision trees. At each node, the algorithm selects the optimal feature and split point based on criteria such as Gini impurity or information gain (see Fig. S1).

However, RF cannot simultaneously account for both theoretical consistency and empirical reliability. The BRF, a variant of RF introduced in this article, demonstrates good experimental performance while maintaining theoretical consistency. Compared to RF, BRF achieves theoretical consistency while retaining strong performance. The key differences lie in their approaches to random bootstrap sampling, random attribute bagging, and tree node splitting

BRF divides the training data into a structural part and an estimation part. The structural part is used to construct the tree’s shape, while the estimation part is reserved exclusively for sample prediction and not used in tree construction. This contrasts with RF, where the entire bootstrap sample is used for both tree construction and prediction.

In the BRF, the feature selection for node splitting is not conducted by merely randomly selecting a subset of features. Instead, Bernoulli trials are employed to decide whether to select a single feature or a subset of features. This method introduces controlled randomness into the feature selection process, thereby enhancing model robustness. As illustrated in Fig. S2, event B1 chooses a value from 0 or 1 with probability p1, where B1 follows a Bernoulli distribution. Here, p1 corresponds to choosing one candidate feature, while 1−p1 corresponds to selecting $\sqrt{D}$ candidate features from the D features in training data. In contrast, traditional RF selects features purely at random without such control. After selecting features and samples, RF deterministically splits tree nodes using conventional decision tree algorithms. Conversely, BRF utilizes Bernoulli trials to determine whether to select the splitting point based on the impurity criterion or random sampling. This decision process is controlled by B2: Bernoulli trial controlled tree node splitting, where p2 represents random sampling, and 1−p2 represents splitting based on the impurity criterion. This additional layer of randomness adds complexity to the tree structure. These three differences collectively ensure that as data size grows indefinitely large, BRF converges to the optimal solution, thereby achieving consistency.

Performance evaluation

We conducted crossover experiments to assess the performance of CNNBRFMDA based on two representative datasets: MDAD and abiofilm. Each dataset was trained separately, with the abiofilm dataset containing 2,884 known associations between 1,720 drugs and 140 microbes, the MDAD dataset containing 2,470 known associations between 1,373 drugs and 173 microbes. To validate the predictive performance of CNNBRFMDA, 5-fold cross-validation was performed. All known microbe-drug associations were randomly divided into five equally sized subsets, with each subset being used as the test set in rotation. In comparison, the remaining four subsets were used as the training set. CNNBRFMDA was used to score all unlabeled microbe-drug pairs and test samples, and the scores of each test sample were compared with the scores of all unlabeled pairs to determine their ranking. Additionally, by setting the association values for the microbe-drug pairs to be predicted as 0, we simulated a scenario where the model predicts new, unseen microbes or drugs that are not present in the current datasets. This approach mirrors real-world conditions, where the model would encounter novel entities, providing a realistic estimate of its predictive capabilities in such situations.

Based on the CNN parameters derived from Deepthi, Jereesh & Liu (2021), we conducted a comparative analysis of CNNBRFMDA against five models: CNNRFMDA, BRFMDA (Wang et al., 2018), NIRBMMDA (Cheng et al., 2022), LAGCNMDA (Qu et al., 2024), RFMDA (Chen et al., 2018), employing a five-fold cross-validation.

For MDAD dataset, as shown in Fig. S3, CNNBRFMDA obtained an AUC of 0.9074, which was superior to CNNRFMDA (0.9043), BRFMDA (0.8873), NIRBMMDA (0.8589), LAGCNMDA (0.8241), RFMDA (0.7834).

For abiofilm dataset, as shown in Fig. S4, CNNBRFMDA obtained an AUC of 0.9221, which was superior to CNNRFMDA (0.9079), BRFMDA (0.8816), NIRBMMDA (0.8660), LAGCNMDA (0.8425), RFMDA (0.8313).

The corresponding Precision-Recall (PR) curves for the CNNBRF model are presented in Figs. S5 and S6, based on the MDAD and abiofilm datasets, respectively. We observed that CNNBRF significantly outperforms both the random forest model and the baseline random model in terms of precision and recall. This indicates that the CNNBRF model is more effective at identifying true microbe-drug associations when processing the MDAD and abiofilm datasets. From the graphs, it is evident that the CNNBRF model maintains a high level of precision across most levels of recall, demonstrating that the model can retain high result reliability without sacrificing significant recall. In contrast, the random forest model shows a more pronounced decline in precision at higher recall levels. Moreover, the overall precision performance of CNNBRF is stable and does not rely on optimization for any specific threshold, indicating that the model performs consistently under varying conditions.

Taking the MDAD dataset as an example, the CNN model initially reduces the dimensionality of the input data, resulting in 773-dimensional feature vectors. These vectors are then input into the BRF, where they are used to construct decision trees. To improve interpretability, we have focused on the features utilized for splitting nodes within these trees. We have generated correlation coefficient plots of these features, as shown in Fig. S7. This visualization demonstrates low inter-feature correlation, suggesting that the features used for splitting in the BRF possess good independence. This independence is key to understanding the model’s robust predictive performance.

To mitigate bias resulting from the random partitioning of known microbe-drug associations, the 5-fold cross-validation process was repeated 10 times. As a result, for the two datasets of MDAD, and abiofilm, CNNBRFMDA acquired the average AUC and standard deviation of 0.9017+/−0.0032 and 0.9146+/−0.0041, respectively, which were higher than those of five previous models. Additionally, we integrated logistic regression and k-nearest neighbors into our comparative analysis to further validate the performance of our model against these well-established techniques (see Tables S1, S2). It is worth mentioning that all prediction models underwent evaluation against CNNBRFMDA, utilizing the same dataset in 5-fold cross-validation.

To validate the efficacy of CNNBRF across different biological contexts, we employed the DrugVirus dataset (Andersen et al., 2020), which contains 933 associations between 175 drugs and 95 viruses. After conducting ten times five-fold cross-validation, we observed a performance metric of 0.7963 ± 0.0017. This result is lower compared to the performances achieved with the MDAD and abiofilm datasets. We believe this underperformance can largely be attributed to the smaller size of the DrugVirus dataset, which limits the model’s ability to generalize as effectively as it does with larger datasets. The results highlight the challenges associated with smaller datasets and emphasize the need for further refinement of CNNBRFMDA to enhance its robustness and accuracy. Improving model performance across various data scales and biological contexts will be a key focus of our future work.

To investigate the impact of different epochs on model performance, we conducted performance tests at epochs 10, 20, 50, and 100 on the MDAD and abiofilm datasets, as depicted in (see Tables S3, S4). We found that at epoch 20, the performance achieved by the training was optimally balanced with computational resource consumption.

Case studies

In our study, we conducted two case studies using the CNNBRFMDA model, building upon previous research to further validate its predictive capabilities. These case studies were informed by the methodologies and findings of earlier studies, as referenced in Cheng et al. (2022), Fan, Wang & Zhu (2023), Li et al. (2023a), Long et al. (2020a), Tian et al. (2023). The first case study focused on predicting potential microbe associations of the drug ciprofloxacin based on the abiofilm dataset using CNNBRFMDA. Ciprofloxacin is one of a new generation of fluorinated quinolones. The results of clinical trials with orally and intravenously administered ciprofloxacin have confirmed its potential for use in a wide range of infections (Campoli-Richards et al., 1988). We employed CNNBRFMDA to predict ciprofloxacin-related microbes, subsequently ranking the top 50 potential ciprofloxacin-related microbes. These predictions were then validated by reviewing the relevant literature on PubMed. The results showed that 41 out of the top 50 predicted ciprofloxacin-related microbes were confirmed through literature review (see Table S5). For instance, the top-ranked ciprofloxacin-related microbe is Escherichia coli, a Gram-negative bacterium in the family Enterobacteriaceae. It can cause various disease syndromes such as different diarrheal diseases, wound infections, meningitis, septicemia, and atherosclerosis (Olsvik et al., 1991). Jakobsen, Lundberg & Frimodt-Møller (2020) found that ciprofloxacin was highly effective in eliminating susceptible Escherichia coli in urine and kidney tissues. This specific effectiveness is crucial, given the increasing antibiotic resistance among E. coli strains, which complicates the management of such infections. The activity of ciprofloxacin against E. coli not only highlights the importance of this antibiotic in the arsenal against Gram-negative bacterial infections but also emphasizes the need for targeted antibiotic therapy. This approach minimizes the misuse of broad-spectrum antibiotics and helps preserve the efficacy of these drugs, which is essential for slowing the development of resistance.

Shigellosis, predicted to be related to ciprofloxacin and ranked 13th, is a major global cause of dysentery (Thompson, Duy & Baker, 2015). It often manifests with symptoms such as diarrhea, abdominal pain, fever, and vomiting, and in severe cases, it can lead to dehydration and other complications. Gharpure et al. (2022) investigated an outbreak of multidrug-resistant Shigella sonnei infections in a retirement community and found that the isolated strains were susceptible to ciprofloxacin. The effectiveness of ciprofloxacin against drug-resistant Shigella sonnei provides a reliable treatment pathway, particularly in outbreak scenarios where a quick and effective response is crucial. This underscores the importance of maintaining ciprofloxacin as a viable treatment option in the face of rising antibiotic resistance, ensuring that healthcare providers can effectively manage outbreaks and prevent widespread transmission and morbidity.

Additionally, Salmonella enterica was predicted to be related to ciprofloxacin and ranked 14th. Salmonella enterica, a Gram-negative rod-shaped bacterium, is commonly transmitted through contaminated water and food and can cause several different disease syndromes, with the most common being gastroenteritis, followed by bacteremia and typhoid fever (Han et al., 2024). Li et al. (2023b) discovered that ciprofloxacin can be used in the treatment of Salmonella enterica serovar Typhimurium infection. The model’s identification of ciprofloxacin as effective against Salmonella enterica serovar Typhimurium is particularly important given the bacterium’s role in significant public health issues like gastroenteritis and typhoid fever. Ciprofloxacin’s efficacy highlights its importance not only in treating common infections but also severe systemic infections caused by Salmonella, providing a reliable treatment option in the face of increasing antibiotic resistance. This underscores the value of targeted antibiotic therapy, allowing for more precise treatment strategies and helping to preserve the effectiveness of essential drugs against resistant strains.

In the second case study, we aimed to verify the predictive ability of CNNBRFMDA on microbe potentially associated with new drugs. To achieve this object, we removed all known associations of the drug moxifloxacin in the MDAD dataset and utilized the processed MDAD dataset to predict potential associated microbe of the new drug moxifloxacin using CNNBRFMDA. The results showed that 38 out of the top 50 predicted moxifloxacin-related microbes were confirmed through literature review (see Table S6). For example, with the highest correlation score to moxifloxacin, Haemophilus influenzae, a pathogenic bacterium causing respiratory and otolaryngological infections (Honda et al., 2020), has been experimentally confirmed by Honda et al. to be susceptible to moxifloxacin treatment. This finding is particularly important for guiding antibiotic stewardship and treatment strategies. By identifying Haemophilus influenzae as highly responsive to moxifloxacin, healthcare providers can more accurately prescribe treatments that are likely to be effective, reducing the need for broader spectrum antibiotics that may contribute to the development of further resistance. Additionally, this outcome highlights the potential of moxifloxacin as a valuable antibiotic in managing infections that are difficult to treat due to either the location or the nature of the bacterial resistance. Additionally, Candida albicans, which ranks second in correlation to moxifloxacin, is a commonly encountered fungal pathogen associated with diseases ranging from superficial mucosal complaints to life-threatening systemic disorders (Wang, 2015). Jadhav et al. (2017) have demonstrated the potential impact of moxifloxacin on the growth, morphogenesis, and abiofilm formation of Candida albicans, with minimal likelihood of Candida albicans developing resistance to moxifloxacin. This efficacy against Candida albicans is particularly significant given the pathogen’s association with a range of diseases from mild mucosal ailments to severe systemic infections that can be life-threatening. The ability of moxifloxacin to inhibit biofilm formation—a key factor in the virulence and drug resistance of Candida infections—further enhances its potential utility in clinical settings where fungal infections are prevalent or complicated by biofilms. The implications of these findings are broad, not only for the treatment of fungal infections but also for the potential development of moxifloxacin as a multi-use antimicrobial agent. This could lead to its strategic use in cases where fungal and bacterial co-infections are suspected, thus simplifying treatment regimes and potentially reducing the need for combination therapies that may have higher toxicities or drug interaction risks. Staphylococcus aureus, which ranks sixth in correlation to moxifloxacin, is a clinically important pathogen causing a wide range of human infections, from minor skin infections to severe tissue infection (Ahmad-Mansour et al., 2021). Ince, Zhang & Hooper (2003) have confirmed that moxifloxacin exhibits enhanced potency against Staphylococcus aureus, lower propensity to select for resistant mutants, and higher bactericidal activity against highly resistant strains. Moxifloxacin’s ability to effectively target Staphylococcus aureus with a lower propensity for developing resistance is particularly significant in the context of increasing antibiotic resistance worldwide. This characteristic makes moxifloxacin an important option in the antibiotic arsenal, especially in settings where drug resistance is prevalent, and traditional treatments fail. Its higher bactericidal activity is also critical for ensuring effective eradication of infections, minimizing the risk of recurrence and the spread of infection within healthcare settings. Moreover, the broad efficacy of moxifloxacin against Staphylococcus aureus enhances its utility not just for typical bacterial infections but also for more complicated clinical scenarios where the bacterial load and resistance factors complicate treatment strategies. This versatility can lead to better clinical outcomes and may reduce the duration and complexity of antibiotic therapy needed to manage infections effectively.