The dataset used in this study sourced from [44] , is provided in CSV format. It contains information about three types of entities: drugs, diseases, and proteins. In total, the dataset includes 1186 drugs from DrugBank, 451 diseases from OMIM, and 1467 proteins from UniProt. The dataset contains three major binary interaction matrices that capture the relationships among these entities:

1) Drug-Protein Interaction Matrix (DP): A binary matrix of size 1186 × 1467, where each entry indicates whether a given drug interacts with a specific protein (1 for interaction, 0 for no interaction).

2) Disease-Protein Interaction Matrix (XP): A binary matrix of size 451 × 1467, representing the associations between diseases and proteins, indicating which proteins are linked to each disease.

3) Drug-Disease Interaction Matrix (DX): A binary matrix of size 1186 × 451, showing known therapeutic relationships where a drug is used to treat a particular disease (1 indicates treatment, 0 indicates no known association).

For both ML and GNN models, the dataset has been split into 80% training and 20% testing sets. Output labels have been derived from the drug-disease association matrix $D_X$.

In this study, negative samples are drug-disease pairs with no known association. These are marked as 0 in the drug-disease matrix $D_X$. On the other hand, positive samples are marked as 1. To keep the dataset balanced, a fixed ratio $r$ has been used. This means $r$ times of positive samples have been chosen as negative samples randomly from the dataset. The ratio is defined in (1).

$\text{Number of negative samples}=r\times \text{Number of positive samples}$(1)

If $r<1$, the model will learn to focus more on positive associations. If $r>1$, it could become biased toward negative predictions due to the imbalance. In this work, a value of $r=1$ has been used.

For each drug-disease pair, a feature vector is constructed by combining the protein association profiles of both entities. The presence or absence of proteins associated with a drug-disease pair (i, j), where i is the drug index and j is the disease index, is encoded in the feature vector for each protein k as outlined in (2).

$p_{ijk}\mathrm{<mo>\; =\; </mo>}\{\begin{array}{l}\mathrm{0.0,}\text{if}\text{neither}\text{drug}i\text{nor}\text{disease}j\text{is}\text{associated}\text{with}\text{protein}k\\ \mathrm{0.3,}\text{if}\text{only}\text{drug}i\text{is}\text{associated}\text{with}\text{protein}k\\ \mathrm{0.6,}\text{if}\text{only}\text{disease}j\text{is}\text{associated}\text{with}\text{protein}k\\ \mathrm{1.0,}\text{if}\text{both}\text{drug}i\text{and}\text{disease}j\text{are}\text{associated}\text{with}\text{protein}k\end{array}$(2)

The label for each drug-disease pair is obtained from the Drug-disease Interaction matrix (DX) by converting the original two-dimensional structure into a one-dimensional array, where each element denotes whether a known association exists (1) or not (0). This flattened array serves as the output label for the ML classifier. A small-scale example involving 3 drugs and 2 diseases is shown in Figure 1 to illustrate this flattening process.

A graph has been constructed from the dataset to be used with a GNN classifier. In this graph, drugs and diseases have been represented as nodes. Each node has association with a binary feature vector that reflects its interactions with proteins. Let p be the total number of proteins in the dataset, where p = 1467. Accordingly, each drug and disease node has been associated with a binary vector of length 1467. These feature vectors have been derived from two binary association matrices defined in the dataset: the Drug-Protein (DP) matrix and the Disease-Protein (XP) matrix.

The graph has been modeled as a bipartite structure consisting of two distinct types of nodes: drugs and diseases. Edges between them have been defined using the Drug-Disease interaction matrix (DX). An entry $D{X}_{i\mathrm{,}j}=1$ means there is a known link between drug i and disease j, and results in an edge between the corresponding nodes. Entries where $D{X}_{i\mathrm{,}j}=0$ are treated as negative edges. The total count of positive edges is represented as $E^{+}$, corresponds to the number of ones in the DX matrix. In this dataset, the number of known drug-disease associations is $\left|E^{+}\right|=1827$.

All nodes, regardless of type, share the same feature space. Each node $v_i$ is associated with a binary feature vector $x_i$ of dimension $p$. The total number of nodes in the graph is $\left|V\right|=n_d+n_x$, where $n_d$ is the number of drugs and $n_x$ is the number of diseases. To distinguish between the two node types, drugs have been assigned node IDs from 0 to $n_d-1$, and diseases from $n_d$ to $n_d+n_x-1$. A simplified example of the graph construction process involving three drugs, two diseases, and six proteins has been presented in Figure 2 .

For using an ML classifier in drug repositioning, the task is set up as a binary classification problem. The goal is to predict if a specific drug-disease pair shows a valid therapeutic connection (label 1) or not (label 0). Each entity in the dataset has been represented by a feature vector of length 1467, corresponding to the total number of proteins. Due to this high dimensionality, the Random Forest classifier has been chosen for its ability to efficiently handle high-dimensional data in this study.

The implementation workflow has been illustrated in Figure 3 . It has started with the preparation of feature vectors for each drug-disease pair using information extracted from the dataset. These feature vectors have been used to train the Random Forest classifier, which has learned to distinguish between valid and invalid treatment associations. Later, the model has been evaluated using standard performance metrics. Afterwards, drug-disease pairs predicted as valid treatments but absent in the original dataset have been identified as potential candidates for drug repositioning.

The key hyperparameters of the Random Forest classifier used in our experiments are summarized in Table 1 .

In this study, drug repositioning has been treated as a link prediction problem. The goal is to predict if a connection (or edge) exists between a drug and a disease, indicating if the drug can potentially treat the disease. To solve this, a Graph Neural Network (GNN) model, GraphSAGE has been used. GraphSAGE is suitable for this task because, unlike some traditional GNNs that need the full graph during training, it learns from a node’s neighbors by sampling and combining their features. This makes it useful for large biological networks, where new drugs or diseases might appear later, and the model can still make predictions for them.

The implementation has been done in several steps. First, a graph has been constructed using the dataset, where drugs and diseases are nodes, and edges show known links between them. Then, GraphSAGE layers have been applied to generate embeddings for each node by collecting information from nearby nodes. Finally, a decoder has been used that takes the embeddings of a drug and a disease and predicts whether a link exists between them. The complete workflow has been shown in Figure 4 , and the core components involved in this process are explained in 2.6.1 and 2.6.2.

Table 2 shows the selected hyperparameters used for the GraphSAGE model, and later the core components involved in this process are explained in 2.6.1 and 2.6.2.

The encoder in GraphSAGE transforms the high-dimensional protein association features of each node into compact and meaningful vector representations called embeddings. These embeddings include information not only about the node itself but also about its nearby nodes. By capturing information about neighbors, the model learns to recognize drugs and diseases that are similar based on their protein associations. This helps the model recommend similar drugs for similar diseases during evaluation. Later, using this strategy, repurposed drugs can be identified. The encoding process follows the steps below:

1) The encoder takes two inputs: a feature matrix and an edge index. Each node representing a drug or a disease starts with a feature vector of 1467 values, based on protein interaction profiles. The edge index defines the connections between nodes in the graph.

2) After input, the encoder uses GraphSAGE layers to transform the original features into lower-dimensional embeddings. In each layer, node features are multiplied by trainable weight matrices. Then, each node updates its features by combining its own transformed values with information received from neighboring nodes. This process allows the model to understand the structure of the graph as well as from node-level data. As the data passes through the layers, the feature dimensions change from 1467 to 128 in the initial layer, and then to 32 at the next layer.

3) Afterwards, the training of the model uses the Binary Cross-Entropy (BCE) loss function. This loss helps guide the learning process by measuring how well the predicted outputs match the actual labels. Minimizing this loss leads to more accurate and meaningful node embeddings

Once the Encoder produces the node embeddings, the Decoder predicts whether an edge exists between a drug and a disease. The decoding process works as follows:

1) First, at first, the model is trained using the Encoder on both existing drug-disease links (positive edges) and absent links (negative edges) from the dataset. The purpose is to learn the model to predict whether a connection exists between a given drug and disease.

2) After the encoding, each drug and disease has its own embedding vector generated by the encoder. These vectors are low-dimensional representations that capture important features of each node. The decoder then calculates the similarity between a drug and a disease by taking the dot product of their embedding vectors. A higher similarity indicates a more significant link between the drug and the disease.

3) The dot product output undergoes a nonlinear activation function, the sigmoid function in this study to map the predicted value into a probabilistic range [0, 1]. This probability is then compared against a predefined threshold $\tau$. If the probability exceeds $\tau$, the model infers the presence of an edge, otherwise, it infers its absence as explained in (3) where H denotes the embedding of nodes.

$y\mathrm{<mo>\; =\; </mo>}\{\begin{array}{l}\mathrm{1,}\text{if}\sigma \left(H_{\text{drug}}.H_{\text{disease}}\right)>\tau \\ \mathrm{0,}\text{otherwise}\end{array}$(3)

TwinNetDR improves drug repositioning predictions by combining GraphSAGE and Random Forest outputs. Both models are trained on the same drug-disease pairs for consistent evaluation. During testing, each model independently predicts associations, and TwinNetDR merges these predictions for the final decision. The workflow is shown in Figure 5 . The following subsections explain how TwinNetDR handles rare repositioning cases and its decision-making strategy.

Identifying new purposes for existing drugs is not common because true repositioning cases are not frequent. When our system predicts a drug could treat a disease that it’s not currently known to help, this could mean one of two things. First, it might be a mistake—what we call a false positive. Second, it might actually be a valuable new discovery that hasn’t been documented yet. This uncertainty is exactly why we developed TwinNetDR. TwinNetDR works smarter by using two different methods together. It combines Random Forest, which analyzes the data in table format, with GraphSAGE, which examines how drugs and diseases connect in a network. This dual approach gives us much more reliable results. When one method might make an error, the other can often catch and correct it. More importantly, when both methods agree on a potential new use that isn’t in our original data, we can be more confident it’s worth further investigation rather than just being a computer error. This way, we reduce simple mistakes while still finding promising new treatment possibilities.

Since repositioning cases are rare, most false positives are likely errors. TwinNetDR prioritizes high-confidence predictions: it favors negative predictions (0) to minimize false positives. A drug-disease pair is flagged only if:

1) Both models agree: If GraphSAGE and Random Forest both predict an association ( $P_{\text{SAGE}}=1$ and $P_{\text{RF}}=1$), the association is confirmed.

2) Models disagree: If one predicts an association and the other does not, TwinNetDR first calculates the average confidence score as described in (4).

$P_{\text{avg}}=\frac{P_{\text{SAGE}}+P_{\text{RF}}}{2}$(4)

The decision-making process is explained in Algorithm 1, where $P_{\text{SAGE}}\left(X_i\right)$ and $P_{\text{RF}}\left(X_i\right)$ represent the sigmoid scores produced for the $i^{\text{th}}$ drug-disease pair $X_i$ by the GraphSAGE and Random Forest models, respectively, during evaluation.

TwinNet-DR performs even better. It achieves the highest AUC score of 97.53 percent, which is better than both GraphSAGE and Random Forest. Its accuracy is slightly lower than GraphSAGE, but it gives a much higher precision of 95.92 percent. In drug repositioning, high precision is very important because it means fewer false positives. If a model predicts fewer wrong drug-disease associations, its results are more trustworthy. So, TwinNet-DR provides more reliable suggestions for repurposed drugs with fewer errors. In addition, Figure 6 presents the combined ROC curves of all three models. The curve for TwinNet-DR lies closest to the top-left corner, clearly indicating superior overall performance compared to the others.

In this study, selecting an appropriate threshold is crucial. If the threshold is set too low, the model produces a large number of false positives. This leads to confusion between true repurposed candidates and incorrect ones, as many negative samples are mistakenly predicted as positive. Conversely, if the threshold is too high, the model becomes overly conservative, predicting most samples as negative. This drastically reduces the chance of identifying true repurposed drug-disease pairs, rendering the model ineffective. Even a midpoint threshold of 0.5 may not be ideal. Although it offers a balanced outcome, it lacks the necessary bias toward predicting negatives in this study’s context.

The model was tested using different decision thresholds. For each threshold, precision, recall, and F1-score were calculated. The precision-recall curve in Figure 7 shows how performance changed with varying thresholds. The best performance was observed at threshold 0.6, where the model achieved an F1-score of 95.26%, with precision 95.92% and recall 94.61%. This threshold also gave the highest accuracy of 90.20%. Based on these results, threshold 0.6 was selected as the final decision point.

Figure 8 further illustrates how the performance metrics vary with changes in the threshold. As the threshold increases, accuracy, recall, and F1-score show a rising trend up to a certain point. Notably, at a threshold of 0.6, most of these metrics reach their highest values, indicating that this point offers the best balance between identifying true positives and minimizing false positives for this study.

In summary, extremely low or high thresholds reduce model effectiveness. A threshold of 0.6 provides the best balance, providing the best accuracy while maintaining a good balance between precision and recall.

The highest ranked drug-disease association predicted by the proposed TwinNetDR is between Amoxapine, with DrugBank ID DB00543 and Drug ID 922 in the dataset, and Diabetes Mellitus, with OMIM ID 125853 and Disease ID 42 in the dataset. This association received a sigmoid score of 91.3% from TwinNet-DR. The same association scored 88.5% using the Random Forest model and 94.1% using the GraphSAGE model. This pair was not part of the known positive samples. This prediction is also supported by the study of Li et al. [43] , which reported Amoxapine as a candidate for diabetes treatment using a drug–target–disease graph neural network.