Accurate prediction of drug-protein interactions (DPIs) is a pivotal step in drug discovery. Classical deep learning models like BridgeDPI [3] have shown strong performance by using Graph Neural Networks (GNNs) with hyper-nodes to effectively learn protein-protein and drug-drug associations (PPAs and DDAs). Siamese-based models such as BioAct-Het use a heterogeneous siamese neural network to model the complex relationship between drugs and bioactivity classes, bringing them into a unified latent space. The final bioactivity class label is decided based on the distance between the final embeddings obtained from the classical twin networks. In our work, we propose a parameter-efficient QuantumHybrid Siamese model where the protein and drug embeddings, generated by the classical encoders are processed on by the Quantum Siamese circuit, much like the classical Siamese model. Using a switch qubit, helps us implement the idea of weight sharing across the twin instances for the pair of inputs, but by using only a single instance of the VQC block. We focus on using our proposed circuit with the classical protein and drug encoders proposed in the BridgeDPI model. Our experiments include testing the model variations on the unseen drugprotein split of the BindingDB dataset as suggested by Gao et. al. from https://github.com/IBM/InterpretableDTIP [3]. Upon experimenting with different variations of the classical encoders and the quantum circuits, we achieved the best performing hybrid model (10 MB) with performance metrics comparable to the best performing classical model$(40 \text{MB})$along with$4 x$size reduction and 540 trainable quantum parameters. The code for this work can be found at https://github.com/SoardRaspi/Qiamese.
Accurate prediction of drug-protein interactions (DPIs) is a pivotal step in drug discovery. Classical deep learning models like BridgeDPI [3] have shown strong performance by using Graph Neural Networks (GNNs) with hyper-nodes to effectively learn protein-protein and drug-drug associations (PPAs and DDAs). Siamese-based models such as BioAct-Het use a heterogeneous siamese neural network to model the complex relationship between drugs and bioactivity classes, bringing them into a unified latent space. The final bioactivity class label is decided based on the distance between the final embeddings obtained from the classical twin networks. In our work, we propose a parameter-efficient QuantumHybrid Siamese model where the protein and drug embeddings, generated by the classical encoders are processed on by the Quantum Siamese circuit, much like the classical Siamese model. Using a switch qubit, helps us implement the idea of weight sharing across the twin instances for the pair of inputs, but by using only a single instance of the VQC block. We focus on using our proposed circuit with the classical protein and drug encoders proposed in the BridgeDPI model. Our experiments include testing the model variations on the unseen drugprotein split of the BindingDB dataset as suggested by Gao et. al. from https://github.com/IBM/InterpretableDTIP [3]. Upon experimenting with different variations of the classical encoders and the quantum circuits, we achieved the best performing hybrid model (10 MB) with performance metrics comparable to the best performing classical model(40MB)along with4xsize reduction and 540 trainable quantum parameters. The code for this work can be found at https://github.com/SoardRaspi/Qiamese.
Protein-ligand binding affinity is critical in drug discovery, but experimentally determining it is time-consuming and expensive. Artificial intelligence (AI) has been used to predict binding affinity, significantly accelerating this process. However, the high-performance requirements and vast datasets involved in affinity prediction demand increasingly large AI models, requiring substantial computational resources and training time. Quantum machine learning has emerged as a promising solution to these challenges. In particular, hybrid quantum-classical models can reduce the number of parameters while maintaining or improving performance compared to classical counterparts. Despite these advantages, challenges persist: why hybrid quantum models achieve these benefits, whether quantum neural networks (QNNs) can replace classical neural networks, and whether such models are feasible on noisy intermediate-scale quantum (NISQ) devices. This study addresses these challenges by proposing a hybrid quantum neural network (HQNN) that empirically demonstrates the capability to approximate non-linear functions in the latent feature space derived from classical embedding. The primary goal of this study is to achieve a parameter-efficient model in binding affinity prediction while ensuring feasibility on NISQ devices. Numerical results indicate that HQNN achieves comparable or superior performance and parameter efficiency compared to classical neural networks, underscoring its potential as a viable replacement. This study highlights the potential of hybrid QML in computational drug discovery, offering insights into its applicability and advantages in addressing the computational challenges of protein-ligand binding affinity prediction.
Finding drug-like molecules is a time- and money-consuming task that requires virtual screening techniques capable of accurately representing non-linear intermolecular relationships. Classical ML models usually consider molecular graphs or SMILES sequences, but cannot jointly exploit both structural and sequential information, let alone simultaneously. Of greater importance, numerous deep learning methods are hindered by poor generalisation, black-box predictions, and inadequate interpretability, reducing their feasibility for application in real-life drug discovery pipelines. In this paper, we present QDrugDiscoverAI, a hybrid quantum–classical deep learning framework for predicting molecular properties and optimising molecules to overcome these challenges. We present a framework for multimodal molecular representation learning that combines graph attention networks and transformer-based sequence encoders with variational quantum circuits to increase expressive capacity within a single architecture. It implements a dual-stream representation-learning mechanism with a multimodal feature fusion layer. It includes an explainability module that uses attention weights and SHAP values to identify substructures that contribute to model predictions and enhance interpretability. Moreover, QDrugDiscoverAI also includes a quantum–assisted reinforcement-learning module for a molecule-optimisation workflow, enabling more efficient exploration of the chemical design space while leveraging the controllability of quantum simulation. The experiments on three benchmark datasets, i.e., Lipophilicity, HIV, and BBBP, show improvements of 12.4%, 9.6%, and 11.1% in accuracy, AUC, and F1-score, respectively, over the matched classical and unimodal baselines. Quantitative evaluation of quantum components and fusion strategies is also conducted through ablation studies to demonstrate their roles further. In summary, the framework proposed serves as a systematic integration of interpretable hybrid learning, showing one way it could be organised to become increasingly efficient, and demonstrating the advantage of using quantum-simulated components to enhance classical deep learning models for exploration and predictive performance in drug discovery workflows. The implementation code and supporting scripts for the proposed QDrugDiscoverAI framework are publicly available at: https://github.com/Sripada-RamaSree/QuantumDrugAI
Drug repositioning is a less expensive and time-consuming method than the traditional method of drug discovery. It is a strategy for identifying new uses for approved or investigational drugs that are outside the scope of the original medical indication. A key strategy in repositioning approved or investigational drugs is determining the binding affinity of these drugs to target proteins. The large increase in available experimental data has helped deep learning methods to demonstrate superior performance compared to conventional prediction and other traditional computational methods in precise binding affinity prediction. However, these methods are complex and time-consuming, presenting a significant barrier to their development and practical application. In this context, quantum computing (QC) and quantum machine learning (QML) theoretically offer promising solutions to effectively address these challenges. In this work, we introduce a hybrid quantum–classical framework to predict binding affinity. Our approach involves, initially, the implementation of an efficient classical model using convolutional neural networks (CNNs) for feature extraction and three fully connected layers for prediction. Subsequently, retaining the classical module for feature extraction, we implement various quantum and classical modules for binding affinity prediction, which accept the concatenated features as input. Quantum predicted modules are implemented with Variational Quantum Regressions (VQRs), while classical predicted modules are implemented with various fully connected layers. Our findings clearly show that hybrid quantum–classical models accelerate the training process in terms of epochs and achieve faster stabilization. Also, these models demonstrate quantum superiority in terms of complexity, accuracy, and generalization, thereby indicating a promising direction for QML.
Drug discovery hinges on the accurate prediction of binding affinity between prospective drug molecules and target proteins that influence disease progression, which is financially and computationally demanding. Although classical and hybrid quantum machine learning models have been employed in previous studies to aid in binding affinity prediction, they encounter several issues related to convergence stability and prediction accuracy. In this regard, this paper introduces a novel hybrid quantum-classical deep learning model tailored for binding affinity prediction in drug discovery. Specifically, the proposed model synergistically integrates 3D and spatial graph convolutional neural networks within an optimized quantum circuit architecture. Simulation results demonstrate a 6% improvement in prediction accuracy relative to existing classical models, as well as a significantly more stable convergence performance compared to previous classical approaches. Moreover, to scalably deploy the proposed framework over today's noisy intermediate-scale quantum (NISQ) devices, a novel quantum error mitigation algorithm is proposed. This algorithm outperforms existing techniques and is capable of mitigating errors with gate noise probabilities,p≤0.05, while resulting in no additional overhead during the training and testing phases.
One of the main challenges in drug discovery is to find molecules that bind specifically and strongly to their target protein while having minimal binding to other proteins. By predicting binding affinity, it is possible to identify the most promising candidates from a large pool of potential compounds, reducing the number of compounds that need to be tested experimentally. Recently, deep learning methods have shown superior performance than traditional computational methods for making accurate predictions on large datasets. However, the complexity and time-consuming nature of these methods have limited their usage and development. Quantum machine learning is an emerging technology that has the potential to improve many classical machine learning algorithms. In this work we present a hybrid quantum-classical convolutional neural network, which is able to reduce by 20% the complexity of the classical network while maintaining optimal performance in the predictions. Additionally, it results in a significant time savings of up to 40% in the training process, which means a meaningful speed up of the drug discovery process.
A central question in hybrid quantum–classical learning is whether variational quantum circuits (VQCs) preserve task-relevant information more efficiently than classical nonlinearities of equal output dimension. We study this question for drug–target interaction prediction in Alzheimer’s disease drug-repurposing workflows using a Hybrid Quantum–Classical Graph Neural Network (HQGNN). The model compresses 512-dimensional GNN embeddings into a four-qubit VQC (128 × compression) before a classical prediction head. On the DAVIS kinase benchmark, HQGNN achieves concordance index (CI) =0.527 and root mean squared error (RMSE) =0.850 , outperforming a dimension-matched classical bottleneck baseline (CI =0.500 ; random ranking). Under the same compression, the quantum layer recovers 51% of the ranking signal lost by compression relative to the full classical model and yields a 5.4% relative CI gain over the classical bottleneck. This effect is obtained with 24 trainable quantum parameters (representational yield Δ CI/parameter ≈1.13×10−3 ). Under depolarizing noise, RMSE degradation remains below 3.3% at realistic noisy intermediate-scale quantum (NISQ) levels ( p≤0.01 ). Because the present benchmark is evaluated on a single transductive split, these results are interpreted as controlled proof-of-concept evidence of representational efficiency rather than definitive deployment-level generalization.