Survey Paper on Predicting Drug Combination Risk Levels

Drug combination therapy has gradually become a promising treatment strategy for complex or co-existing diseases. As drug-drug interactions (DDIs) may cause unexpected adverse drug reactions, DDI prediction is an important task in pharmacology and clinical applications. Recently, researchers have proposed several deep learning methods to predict DDIs. However, these methods mainly exploit the chemical or biological features of drugs, which is insufficient and limits the performances of DDI prediction. Here, we propose a new deep multimodal feature fusion framework for DDI prediction, DMFDDI, which fuses drug molecular graph, DDI network and the biochemical similarity features of drugs to predict DDIs. To fully extract drug molecular structure, we introduce an attention-gated graph neural network for capturing the global features of the molecular graph and the local features of each atom. A sparse graph convolution network is introduced to learn the topological structure information of the DDI network. In the multimodal feature fusion module, an attention mechanism is used to efficiently fuse different features. To validate the performance of DMFDDI, we compare it with 10 state-of-the-art methods. The comparison results demonstrate that DMFDDI achieves better performance in DDI prediction. Our method DMFDDI is implemented in Python using the Pytorch machine-learning library, and it is freely available at https://github.com/DHUDEBLab/DMFDDI.git.

As treatment plans involve more medications, anticipating and preventing drug-drug interactions (DDIs) becomes increasingly important. Such interactions can result in harmful side effects and may reduce therapy effectiveness. Currently, most computational approaches for DDI prediction rely heavily on complex feature engineering and require chemical information to be structured in specific formats to enable accurate detection of potential interactions. This study presents the first investigation of the application of Large Language Models (LLMs) for DDI prediction using drug characteristics expressed solely in free-text form. Specifically, we use SMILES notations, target organisms, and gene associations as inputs in purpose-designed prompts, allowing LLMs to learn the underlying relationships among these descriptors and accordingly predict possible DDIs. We evaluated the performance of 18 distinct LLMs under zero-shot, few-shot, and fine-tuning settings on the DrugBank dataset (version 5.1.12) to identify the most effective paradigm. We then assessed the generalizability of the fine-tuned models on 13 external DDI datasets against well-known machine learning baselines. The results demonstrated that, while zero-shot and few-shot paradigms showed only modest utility, fine-tuned models achieved superior sensitivity while maintaining competitive accuracy and F1-score compared to baselines. Notably, despite its small size, the Phi-3.5 2.7B model attained a sensitivity of 0.978 and an accuracy of 0.919. These findings suggest that computational efficiency and task-specific adaptation are more important than model size in order to capture the complex patterns inherent in drug interactions, and outline a more accessible paradigm for DDI prediction that can be integrated into clinical decision support systems.

Induction of cytochrome P450 (CYP) 3A4 potentially reduces the blood concentrations of substrate drugs to less than one-tenth, which results in ineffective pharmacotherapy. Although the prediction of drug-drug interactions (DDIs) that are mediated by induction of CYP3A4 has been performed mainly on the basis of in vitro information, such methods have met with limited success in terms of their accuracy and applicability. Therefore, a realistic method for the prediction of CYP3A4-mediated inductive DDIs is of major clinical importance. The objective of the present study was to construct a robust and accurate method for the prediction of CYP3A4-mediated inductive DDIs. Such a method was developed on the basis of the principle applied for prediction of inhibitory DDIs in a previous report. A unique character of this principle is that the extent of alterations in the area under the plasma concentration-time curve (AUC) is predicted on the basis of in vivo information from minimal clinical studies without using in vitro data. The analysis is based on 42 DDI studies in humans reported in 37 published articles over the period 1983-2007. Kinetic analysis revealed that the reduction in the AUC of a substrate of CYP3A4 produced by consecutive administration of an inducer of CYP3A4 could be approximated by the equation 1/(1 + CRCYP3A4 * ICCYP3A4), where CRCYP3A4 is the ratio of the apparent contribution of CYP3A4 to the oral clearance of a substrate and ICCYP3A4 is the apparent increase in clearance of a substrate produced by induction of CYP3A4. Using this equation, the ICCYP3A4 was calculated for seven inducers (bosentan, carbamazepine, efavirenz, phenytoin, pioglitazone, rifampicin [rifampin], and St John's wort [hypericum]) on the basis of the reduction in the AUC of a coadministered standard substrate of CYP3A4, such as simvastatin, in ten DDI studies. The CRCYP3A4 was calculated for 22 substrates on the basis of the previously reported method from inhibitory DDI studies using a potent CYP3A4 inhibitor such as itraconazole or ketoconazole. The proposed method enabled the prediction of AUC reduction by CYP3A4 induction with any combination of these substrates and inducers (total 154 matches). To assess the accuracy of the prediction, the AUC reductions in 32 studies were analysed. We found that the magnitude of the deviation between the mean values of the observed and predicted AUCs of all substrate drugs was <20% of the AUCs of the respective substrate drugs before administration of the inducers. In addition, rifampicin was found to be the most potent inducer among the compounds analysed in the present study, with an ICCYP3A4 value of 7.7, followed by phenytoin and carbamazepine, with values of 4.7 and 3.0, respectively. The ICCYP3A4 values of the other CYP3A4 inducers analysed in the present study were approximately 1 or less, which suggests that the AUCs of coadministered drugs may not be reduced to less than approximately half, even if the drug is metabolized solely by CYP3A4. By using the method reported in the present study, the susceptibilities of a substrate drug of CYP3A4 to inductive DDIs can be predicted quantitatively. It was indicated that coadministration of rifampicin, phenytoin and carbamazepine may reduce plasma AUCs to less than half for a broad range of CYP3A4 substrate drugs, with CRCYP3A4 values greater than 0.13, 0.21 and 0.33, respectively.

Uncovering novel drug-drug interactions (DDIs) plays a pivotal role in advancing drug development and improving clinical treatment. The outstanding effectiveness of graph neural networks (GNNs) has garnered significant interest in the field of DDI prediction. Consequently, there has been a notable surge in the development of network-based computational approaches for predicting DDIs. However, current approaches face limitations in capturing the spatial relationships between neighboring nodes and their higher-level features during the aggregation of neighbor representations. To address this issue, this study introduces a novel model, KGCNN, designed to comprehensively tackle DDI prediction tasks by considering spatial relationships between molecules within the biomedical knowledge graph (BKG). KGCNN is built upon a message-passing GNN framework, consisting of propagation and aggregation. In the context of the BKG, KGCNN governs the propagation of information based on semantic relationships, which determine the flow and exchange of information between different molecules. In contrast to traditional linear aggregators, KGCNN introduces a spatial-aware capsule aggregator, which effectively captures the spatial relationships among neighboring molecules and their higher-level features within the graph structure. The ultimate goal is to leverage these learned drug representations to predict potential DDIs. To evaluate the effectiveness of KGCNN, it undergoes testing on two datasets. Extensive experimental results demonstrate its superiority in DDI predictions and quantified performance.

Introduction Advanced artificial intelligence (AI) frameworks particularly, large language models (LLMs) have recently attracted attention for automating Drug–drug interactions (DDIs) extraction and prediction tasks. However, there is a scarcity of reviews on how LLMs can rapidly identify known and novel DDIs. Areas covered This review summarizes the state of LLM-based DDI extraction and prediction, based on a broad literature search from PubMed, Embase, Web of Science, Scopus, IEEE Xplore, the Cochrane Library, ACM Digital Library, Google Scholar, and Semantic Scholar published between January 2000 and February 2025. For DDI extraction from biomedical text and databases, we detail methods utilizing transformer-based models, such as domain-specific BioBERT and general GPT-based architectures. For DDI prediction, we discuss prediction frameworks including hybrid models (e.g. SmileGNN, DrugDAGT), conversational agents (e.g. ChatGPT), and prompt-based methods (e.g. DDIPrompt). Expert opinion LLMs offer potential for advancing pharmacovigilance and clinical decision support. However, realizing this and establishing clinical trust requires urgently addressing current limitations, particularly enhancing model explainability, improving reliability (mitigating hallucinations), and resolving data quality issues. Future research must prioritize rigorous clinical validation (prospective studies), developing robust explainable AI (XAI) techniques, refining data curation, and integrating multimodal patient data.

Multi-modal fusion model for Time-Varying medical Data: Addressing Long-Term dependencies and memory challenges in sequence fusion.

Graph neural networks (GNNs) is an information - processing system that uses message passing among graph nodes. In recent years, GNN variants including graph attention network (GAT), graph convolutional network (GCN), and graph recurrent network (GRN) have shown revolutionary performance in computer vision applications using deep learning and artificial intelligence. These neural network model extensions, collect information in the form of graphs. GNN may be divided into three groups based on the challenges it solves: link prediction, node classification, graph classification. Machines can differentiate and recognise objects in image and video using standard CNNs. Extensive amount of research work needs to be done before robots can have same visual intuition as humans. GNN architectures, on the other hand, may be used to solve various image categorization and video challenges. The number of GNN applications in computer vision not limited, continues to expand. Human-object interaction, actin understanding, image categorization from a few shots and many more. In this paper use of GNN in image and video understanding, design aspects, architecture, applications and implementation challenges towards computer vision is described. GNN is a strong tool for analysing graph data and is still a relatively active area that needs further researches attention to solve many computer vision applications.

Suicide is the fourth leading cause of death among people aged 15-29. More than 720, 000 people commit suicide every year. During the COVID-19 pandemic, we saw an increase in people seeking out mental health support on anonymous forums like Reddit. These anonymous forums allow people to express their suicidal ideation without judgment and give them a support structure that not everyone has. The aim of this project is to detect suicidal ideation using Reddit. In this work, we propose SIRGEL (Suicidal Ideation on Reddit using Graph Embeddings and LLMs), a dual-pipeline approach that combines large language model (LLM)- based annotation with both traditional text classification and graph neural network (GNN)-based modeling. The first step is to create a Reddit dataset for related subreddits. The second step is to annotate them as suicidal or not suicidal using LLMs. We also detail the creation of a new dataset made using Reddit posts and its benchmarking using baseline machine learning models. We model this problem using two approaches: traditional text classification and graph representation learning. Additionally, experiments with different types of Graph Neural Networks are performed with optimizations to improve and achieve efficient results

Rich textual and topological information of textual graphs need to be modeled in real-world applications such as webpages, e-commerce, and academic articles. Practitioners have been long following the path of adopting a shallow text encoder and a subsequent graph neural network (GNN) to solve this problem. In light of recent advancements in large language models (LLMs), it is apparent that integrating LLMs for enhanced textual encoding can substantially improve the performance of textual graphs. Nevertheless, the efficiency of these methods poses a significant challenge. In this paper, we propose ENGINE, a parameter- and memory-efficient fine-tuning method for textual graphs with an LLM encoder. The key insight is to combine the LLMs and GNNs through a tunable side structure, which significantly reduces the training complexity without impairing the joint model's capacity. Extensive experiments on textual graphs demonstrate our method's effectiveness by achieving the best model performance, meanwhile having the lowest training cost compared to previous methods. Moreover, we introduce two variants with caching and dynamic early exit to further enhance training and inference speed. Specifically, caching accelerates ENGINE's training by 12x, and dynamic early exit achieves up to 5x faster inference with a negligible performance drop (at maximum 1.17% relevant drop across 7 datasets). Our codes are available at: https://github.com/ZhuYun97/ENGINE.

Drug–target interaction (DTI) prediction constitutes an important challenge of drug discovery, which provides computational substitutes to expensive laboratory experiments. Current models tend to be hampered by generalizability, scalability, and interpretability. To overcome these issues, the introduce a new multimodal deep learning architecture that combines Graph Neural Networks (GNNs), Kolmogorov– Arnold Networks (KANs), and Large Language Models (LLMs). In the proposed method, drug molecular structures are graph-learned as graphs by GNNs, proteins are represented using LLM-based embeddings, and KAN layers regularize outputs for nonlinear learning and explainability. A multimodal attention fusion layer fuses structural and semantic features into predicting interaction probabilities. The model was tested on BindingDB, Davis, Enzyme, and GPCR datasets, outperforming DeepDTA, GraphDTA, MolTrans, and DrugKANs with improvements of up to 5% in AUC and 4% in F1-score. In addition, SHAP-based interpretability identifies most important molecular substructures and protein motifs associated with binding. These findings show the promise of integrating symbolic, structural, and semantic representations for reliable and interpretable drug discovery.

Various CYP time-dependent inhibition (TDI) assays have been widely implemented in drug discovery and development which has led to great success in positively identifying compounds with mechanism-base inhibition liability. However, drug-drug interaction (DDI) predictions by various in-silico models utilizing kinetic parameters obtained from TDI assays have met with significant challenges including questionable kinetic data, over-simplified in-vitro models and unreliable mathematic algorithms. Although significant efforts have been made to standardize the TDI assay and refine mathematical models, recent evaluation studies have revealed that the kinetic parameters of TDI, the most important in-vitro data required by all DDI prediction models, are significantly impacted by a variety of experimental variables including microsomal protein concentration, metabolic stability, CYP-specific probes, and post-incubation time. This review attempts to provide medicinal chemists a brief overview on the current status of TDI assays, determination of kinetic parameters and in silico DDI predictions with emphasis on the complexity of the TDI kinetics and limitations of current in-vitro models and DDI prediction methodologies.

This study introduces a novel approach to drug discovery in regenerative medicine through the utilization of a graph neural network (GNN). The research methodology integrates the development and training of the GNN with a subsequent evaluation of its performance metrics. The first phase involves the generation of synthetic data simulating a biological network, employing networkX and NumPy libraries to construct a random graph with Erdos-Renyi topology. The data, representing cellular responses to biomaterials, is then converted into PyTorch tensors for compatibility with the GNN architecture. The GNN model, characterized by two fully connected layers with ReLU and log-softmax activations, captures intricate relationships within the graph-structured data. The second phase employs a stochastic gradient descent algorithm, specifically the Adam optimizer, to train the GNN over 100 epochs using the cross-entropy loss for multi-class classification. The research methodology extends to the evaluation phase, producing three distinct output graphs for analysis: Visualization of the graph structure, a comparison between predicted and true labels, and a plot illustrating training loss over epochs. Performance metrics, including accuracy, precision, recall, and F1-score, are computed to assess the model’s predictive capabilities quantitatively. The study concludes with a discussion on the nuances revealed by each graph and their implications for refining GNN models in the context of drug discovery for regenerative medicine.

Unlocking the potential of advanced large language models in medication review and reconciliation: A proof-of-concept investigation

Large Language Models (LLMs) exhibit strong In-Context Learning (ICL) capabilities when prompts with demonstrations are used.However, fine-tuning still remains crucial to further enhance their adaptability.Prompt-based fine-tuning proves to be an effective fine-tuning method in low-data scenarios, but high demands on computing resources limit its practicality.We address this issue by introducing a prompt-based parameter-efficient finetuning (PEFT) approach.GNNAVI leverages insights into ICL's information flow dynamics, which indicates that label words act in prompts as anchors for information propagation.GN-NAVI employs a Graph Neural Network (GNN) layer to precisely guide the aggregation and distribution of information flow during the processing of prompts by hardwiring the desired information flow into the GNN.Our experiments on text classification tasks with GPT-2 and Llama2 show GNNAVI surpasses standard prompt-based fine-tuning methods in few-shot settings by updating just 0.2% to 0.5% of parameters.We compare GNNAVI with prevalent PEFT approaches, such as prefix tuning, LoRA and Adapter in terms of performance and efficiency.Our analysis reveals that GNNAVI enhances information flow and ensures a clear aggregation process. 1

Heterogeneous graphs are an essential structure that models real-world data through different types of nodes and relationships between them, including multimodality, which comprises different types of data such as text, image, and audio. Graph Neural Networks (GNNs) are a prominent graph representation learning method that takes advantage of the graph structure and its attributes that, when applied to the multimodal heterogeneous graph, learn a unique semantic space for the different modalities. Consequently, it allows multimodal fusion through simple operators such as sum, average, or multiplication, generating unified representations considering the supplementary and complementarity relationships between the modalities. In multimodal heterogeneous graphs, the labeling process tends to be even more costly due to the multiple modalities analyzed, in addition to the imbalance of classes inherent to some applications. In order to overcome these problems in applications that comprise a class of interest, One-Class Learning (OCL) is used. Given the lack of studies on multimodal early fusion in heterogeneous graphs for OCL tasks, we proposed a method based on unsupervised GNN for heterogeneous graphs and evaluated different early fusion operators. In this paper, we extend another work by evaluating the behavior of the main GNN convolutions in the method. We highlight that using operators such as average, addition, and subtraction were the best early fusion operators. In addition, GNN layers that do not use an attention mechanism performed better. In this way, we argue for heterogeneous graph neural networks in multimodal using early fusion simple operators instead of well-often-used concatenation and less complex convolutions.