Quantum computing applications in drug discovery Authors: Jing Li, Leyi Wei, Henry H Y Tong, Quan Zou Published: 2026 ABSTRACT AbstractIn early drug discovery, virtual screening based on deep learning, virtual screening based on molecular docking, and molecular dynamics are three widely used computational strategies, but they always face a trade-off between throughput, search stability, and physical fidelity. This article discusses how quantum computing can be integrated into these processes under the constraints of Noisy Intermediate-Scale Quantum (NISQ). At present, the most realistic role of quantum computing is not the complete replacement of classical processes, but modular coprocessing for selected decision-sensitive subroutines. In the screening of deep learning, quantum modules are mainly inserted into selected components of the model. In predictive models, they are used to enhance representation learning or feature extraction. In generative models, they serve as priors or generators. In docking screening, quantum integration is suitable for specific substeps such as site recognition, pose search, and flexible docking. In molecular dynamics, representative examples include ground state ab initio molecular dynamics, annealer-based trajectory propagation, and excited state molecular dynamics, while most large-scale sampling is still done by classical methods. The actual problem in these scenarios is not whether the quantum module can be inserted, but whether it can provide repeatable gains related to decision-making under the constraints of actual running time and resources. Therefore, we emphasize strong classical baselines, reliable ranking and calibration, transparent resource reporting, and evaluation at downstream decision points as key criteria for assessing progress in the near term. Introduction Current status and challenges in drug discovery General context of drug discovery and development Modern drug research and discovery can be broadly divided into two major phases: discovery and development. The discovery phase encompasses target identification and validation, as well as lead identification and optimization, while the development phase involves systematic preclinical evaluation, process and formulation development, and subsequent validation through human trials. Despite these structured stages, the drug development pipeline typically spans 10–15 years and often exceeds USD 1–2 billion in cost, with high attrition rates resulting in only a small fraction of candidates ultimately reaching the market [1]. Among the multiple stages of drug discovery and development, early stage computational methods are particularly important for supporting candidate discovery and refinement. In this context, computational methods are widely used to improve efficiency and guide experimental testing [2–5] (Fig. 1a). Background of drug discovery and representative calculation methods in early screening. Role of computational methods in drug discovery With the rapid expansion of target structural data, screening datasets, and chemical space, computer-aided drug design (CADD) has become an indispensable component of early stage drug discovery. By shifting much of the trial and error process to in silico workflows, CADD reduces experimental cost while improving hit identification rates [6]. In recent years, machine learning and deep learning have further enhanced the rapid evaluation and prioritization of candidate molecules. Consequently, virtual screening has evolved from approaches based on rules and scoring functions toward direct prediction driven by data, enabling efficient prefiltering of extremely large compound libraries (Fig. 1b). In early stage drug discovery, three representative computational roles are frequently encountered: rapid screening, pose generation, and dynamical assessment. Deep learning–based virtual screening takes as input molecular fingerprints, simplified molecular input line entry system (SMILES) strings, molecular graphs, and 3D geometric features. These models learn mappings from molecular structure to activity, binding affinity, or phenotypic response, and then perform batch inference to rank large compound libraries. Its key advantages are scalability and speed, which make it well suited for first-pass filtering across vast chemical spaces. The second category is molecular docking-based virtual screening, which generates plausible ligand binding poses within a defined binding pocket and applies scoring functions to obtain coarse estimates of binding affinity. Molecular docking is widely used to generate interpretable binding hypotheses [7–11] (Fig. 1b). The third category is molecular dynamics, which simulates the time evolution of molecular complexes, typically in explicit or implicit solvent depending on the model and application, thereby enabling the assessment of structural stability and changes in key interactions. Molecular dynamics is often used to evaluate the physical plausibility of docking poses and to analyse induced conformational changes. In a specific system, molecular dynamics simulation (sometimes combined with free energy methods) can compare candidate compounds more accurately. However, this advantage usually depends on the specific situation and is not universally observed in all target or drug discovery environments (Fig. 1b). In some projects, virtual screening based on deep learning, virtual screening based on docking, and molecular dynamics can be used in stages or complement each other [12–15]. However, these methods are not intrinsically sequential and do not always need to be used together. Depending on the specific task, goal, and available resources, each method can be used as an independent computing entrance. In many drug discovery projects, the choice of the final compound is not simply determined by the calculation ranking. More commonly, candidate compounds with higher rankings will be further screened by drug chemists and project teams based on factors such as synthesizability, novelty, developability, accessibility of synthesis, and target specificity knowledge before entering experimental verification. In this context, the calculation score should be regarded as one of the reference factors for project level decision-making, rather than an independent final decision-making rule. From this perspective, the actual value of quantum computing is unlikely to replace human judgment, but plays a role by improving the reliability of ranking and decision-making confidence in screening under expert guidance. Accuracy and throughput trade-offs in computational drug discovery Although informatics and computational methods have significantly improved the efficiency of early screening, the overall speed of drug research and development has not improved accordingly. One of the fundamental reasons is that there is always a trade-off between accuracy and scale. Higher physical fidelity usually means higher computational costs, and higher flux often depends on stronger approximations, which may introduce additional uncertainty [16, 17]. This contradiction is particularly prominent in the three key calculation methods introduced in this article: virtual screening based on deep learning, virtual screening based on molecular docking, and a molecular dynamics simulation. Speed and reliability limits in virtual screening based on deep learning The deep learning model is good at high-throughput reasoning of massive chemical libraries, so it can quickly screen compounds at a lower cost. But its limitations are also obvious. This kind of model mainly captures the patterns in the training data. When applied to new chemical skeletons, different detection conditions, or unfamiliar target categories, it will be affected by distribution shift and reduce performance. In addition, labeling datasets usually contain experimental noise, batch effects, and category imbalance deviations, which reduce the comparability and interpretability of scores [18, 19]. Therefore, when screening large compound libraries, deep learning is more suitable as a tool for rough screening and sorting, rather than for final decision-making. Molecular docking, molecular dynamics simulation, or experimental verification are still necessary means for priority sorting [20–22] (Fig. 1b). Interpretability and stability limits in virtual screening based on molecular docking The main task of molecular docking is to search for possible ligand poses in the binding pocket within a high-dimensional conformational space and rank them using scoring functions. The main computing cost comes from numerous sampling and repeated scoring. Accuracy is still a key limiting factor because many scoring functions rely on simplified assumptions about solvents, receptor flexibility, induced fit, and entropy. Differences across software packages and parameter settings can further amplify the uncertainty in screening results [23]. Therefore, molecular docking is more suitable to be regarded as a tool to generate explainable conformational assumptions and preliminary sorting, rather than a stable quantitative prediction method of combined affinity. In practice, it is usually necessary to adopt a variety of scoring strategies and downstream verification workflows to reduce error propagation (Fig. 1b). Physical realism and computational cost in molecular dynamics Molecular dynamics simulates the time evolution of molecular complexes in explicit or implicit solvent environments. The resulting trajectory can be used to evaluate structural stability, conformational changes, and key interactions, providing stronger physical support for verifying molecular docking postures and exploring molecular mechanisms. However, many biologically relevant processes occur on microsecond to millisecond timescales. Because traditional molecular dynamics uses a very small time step, it takes a lot of simulation steps to reach such a timescale, and the calculation cost is very high. In addition, a rugged energy landscape will capture the trajectory, resulting in the instability of the results based on finite sampling [24–27]. Enhanced sampling methods can partially alleviate these problems, but they introduce additional parameters, require the selection of reaction coordinates and reweighting procedures, and can add uncertainty. In practice, molecular dynamics remains computationally demanding and typically requires expert setup and careful parameter selection [28, 29] (Fig. 1b). The basic principles and potential advantages of quantum computing Basic concepts of quantum computing Quantum computing is a computing paradigm that uses quantum states to represent and manipulate information. It encodes the information in the quantum bit, and the quantum bit can be in the superposition state of the classical state 0 and 1 before measurement. When multiple quantum bits are combined, the accessible state space will grow rapidly as the number of quantum bits increases. Quantum entanglement can further introduce strong non-classical associations, thus providing computational strategies that are completely different from classical methods for certain types of problems [30–33]. In practice, quantum gates and quantum circuits realize the controlled operation of quantum amplitude and phase, and transform the superposition state and entanglement state into executable algorithm programs [34]. It should be noted that quantum advantage is not universally applicable. Quantum advantage depends on whether the problem structure can be effectively exploited by a quantum algorithm. It also depends on practical factors such as oracle construction, data loading, and measurement cost. Therefore, any claimed acceleration effect must be evaluated through appropriate benchmarking and fair comparison [35, 36]. The core background terms used in this review are summarized in Table 1. Table 1Core background terms.Key termFull nameDefinitionNISQNoisy Intermediate-Scale QuantumThe current stage of quantum hardware, characterized by device noise, limited qubit counts, and constrained coherence timesHybrid QCHybrid quantum–classical computingA hybrid computing paradigm in which quantum processors handle a small number of critical computations, while classical computing supports scalable processing and robust validationQMQuantum mechanicsQuantum mechanical calculations provide high physical accuracy but are computationally expensiveMMMolecular mechanicsClassical molecular mechanics and force field models are computationally efficient, but they may be insufficiently accurate for processes such as bond breaking and formation or metal coordinationVQEVariational quantum eigensolverHigh accuracy evaluation and correction of electronic energies for small subsystemsQAOAQuantum approximate optimization algorithmApproximate solution of discrete combinatorial optimization subproblemsRescoringRanking rescoringRefinement and calibration of coarse rankings produced by classical methodsQM/MMQuantum mechanics/molecular mechanicsPartitioning of a system into quantum and classical regions to balance accuracy and computational tractability Core background terms. Theoretical advantages of quantum computing Molecular systems are dominated by quantum mechanics (QM) at the microscopic level. However, high accuracy electronic structure calculations become prohibitively expensive as molecular system size increases, requiring trade-offs between accuracy and computational cost [37, 38]. Quantum computing is therefore often considered to be a natural fit with molecular simulation, because it can directly use quantum bits to represent quantum states and simulate Hamiltonian dynamics through quantum circuits. From the perspective of the calculation framework, this method is closer to the inherent description of multi-body quantum systems, thus providing a theoretical way for calculating molecular energy and properties with higher accuracy. Importantly, this theoretical advantage refers to a potential capability of the computational paradigm rather than guaranteed practical performance. In ideal fault-tolerant quantum computing, quantum algorithms can change the computational complexity and expand the accuracy limits that some core subroutines can achieve [31, 39]. The scope of this review In recent years, virtual screening based on deep learning, virtual screening based on molecular docking, and molecular dynamics simulation have become the three most widely used calculation methods in early drug discovery. The deep learning model can quickly prioritize the massive combination library; molecular docking can generate candidate ligand target binding configurations and provide rough sorting; molecular dynamics simulation can evaluate the stability, conformation changes, and binding free energy of complexes. Despite these advances, these three methods still face fundamental bottlenecks in balancing scale, accuracy, and sampling efficiency. The deep learning model is limited by the distribution and generalization ability of training data, the molecular docking is limited by conformational search and simplified scoring functions, and the molecular dynamics is limited by the accuracy of the force field and the challenges of evaluating long-term scale and rare event sampling. Accordingly, this review focuses on three early stage computational workflows in computer-aided drug discovery: virtual screening using deep learning, virtual screening using molecular docking, and molecular dynamics. Under the Noisy Intermediate-Scale Quantum (NISQ) constraints, it examines quantum computing as a modular addition to classical workflows rather than as a replacement for classical pipelines. This review does not attempt to cover the full drug discovery pipeline or downstream wet lab and clinical validation. As illustrated in Fig. 2, quantum modules may be introduced at selected subroutines within otherwise classical computational procedures. These three workflows are often connected in practical discovery pipelines, but they are not treated here as a compulsory fixed sequence. Rather, each is discussed as an analytically separable workflow in which quantum modules may be introduced independently at selected subroutines when justified by the task and resource budget. The integration of plugin quantum computing in three typical computational strategies for early drug discovery. A recent systematic mapping review in Briefings in Bioinformatics by Nałęcz-Charkiewicz et al. [40] surveyed quantum computing methods and trends across a wide range of bioinformatics domains. By contrast, the present review focuses specifically on early stage drug discovery. It is organized around three operational computational workflows: virtual screening based on deep learning, virtual screening based on molecular docking, and molecular dynamics, with an emphasis on modular quantum integration within otherwise classical pipelines under realistic NISQ constraints. Within this scope, the review mainly centers on small molecule discovery for protein targets. This focus reflects the current concentration of representative quantum-assisted studies in early stage CADD and allows deep learning, molecular docking, and molecular dynamics to be discussed within a coherent methodological framework. RNA- and DNA-targeted ligand discovery is an important emerging direction, but it is not treated within the workflows discussed in this Review, because our scope focuses on protein-ligand drug discovery and the directly relevant quantum computing literature for RNA- or DNA-targeted ligand discovery remains comparatively limited. In addition to methodological analysis at the workflow level, Table 2 summarizes representative public collaborations between major technology or quantum-focused players and biopharma partners, illustrating the current industrial exploration of quantum approaches in drug discovery at the early stage. Table 2Representative public collaborations relevant to quantum computing in drug discovery.Region of primary quantum playerQuantum computing playerBiopharma partnerCollaboration focusUSAIBM QuantumModernaPublic collaboration exploring quantum computing applications in biomedicine and drug discovery within the IBM Quantum Network ecosystem (IBM Newsroom)ChinaTencent Quantum LabJOINCAREStrategic collaboration on quantum computing and AI for synthetic biology and drug discovery (JOINCARE)USASandboxAQAstraZenecaAdvancing molecular simulation for drug discovery via the AQBioSim initiative (SandboxAQ)SwitzerlandQC WareRocheJoint exploration of quantum neural networks for drug development tasks (Roche) Representative public collaborations relevant to quantum computing in drug discovery. Classical bottlenecks and quantum opportunities in drug discovery Against this background, this section first outlines the main classical workflows in drug discovery at the early stage and then highlights the bottlenecks that create realistic opportunities for modular quantum integration. Rather than replacing classical systems from end to end, a quantum processor is more plausibly deployed as a coprocessing module for a small number of steps that are especially sensitive, computationally expensive, or decisive. For the purposes of this review, we focus on three bottlenecks specific to these workflows in which modular quantum integration may be particularly relevant: ranking stability in virtual screening based on deep learning, combinatorial search in docking, and local refinement in molecular simulation. Although these areas are often connected in practical discovery workflows, they may also be considered independently, and quantum modules need not be introduced across the full pipeline. Deep learning–based virtual screening In drug discovery, virtual screening based on deep learning typically follows a classical workflow that is well established. First, labeled datasets (such as compound activity, affinity, or phenotypic response pairs) are curated from public databases or internal projects. Molecules are then encoded into representations suitable for modeling, including molecular fingerprints [41], SMILES strings [42], 2D molecular graphs [43], or 3D geometric features [44]. Subsequently, predictive models are trained to perform classification or regression of molecular properties or activities. After model tuning and calibration, the large candidate compound library is screened in batches, scoring is generated, and the compounds are sorted to determine the priority as soon as possible. Only a small part of the top-ranked candidate compounds can enter the stage with higher calculation and experimental requirements, such as molecular docking, free energy calculation, or experimental verification. This hierarchical screening strategy improves the hit rate and reduces the cost of iterative testing. In virtual screening based on deep learning, the bottleneck often lies in the reliability of ranking rather than the calculation throughput. Label noise, data deviation, distribution offset, and category imbalance will all lead to score instability. Quantum modules are more suitable to provide support at the model level (see Section ‘Classical–quantum hybrid strategies for deep learning models’ for details) [45, 46]. Virtual screening based on molecular docking The classic virtual screening based on molecular docking includes extensive conformational sampling, and then screening according to the score. Usually, protein and ligand structures are first prepared by adding hydrogen atoms, assigning charges and protonation states, and defining binding pockets or docking grids [47]. The translation, rotation, and torsional degrees of freedom of the ligand are treated as variables, allowing many candidate poses to be generated across conformational space. Then each candidate pose is scored to roughly estimate the degree of favorability of the combination. The score usually integrates shape complementarity, hydrophobic interaction, hydrogen bond, electrostatic action, and desolvation effect. Promising poses may subsequently be locally optimized or rescored. The dominant computational costs arise from two sources. First, extensive sampling is required in a high-dimensional conformational space [48]. Second, scoring functions must be evaluated repeatedly for each sampled pose. This burden is further amplified in multi-pocket settings, ensemble docking, or large-scale screening [49, 50]. The main difficulty is the combinatorial growth of conformational and pose space in molecular docking-based virtual screening. Docking can be framed as a problem of discrete decisions and combinatorial optimization. Classical docking relies on heuristic search and approximate scoring for efficiency, but it may explore the space insufficiently or get trapped in local optima. Quantum modules are more naturally introduced at selected docking substeps. They can turn a few key discrete decisions into optimization subproblems, produce more choices under a limited call budget, and then return the results to the classical docking procedure for subsequent scoring and assessment [51, 52]. Classical methods provide broad coverage, whereas quantum modules are better suited to supporting a limited number of key decisions (see Section ‘Hybrid classical–quantum docking’ for more details). Molecular dynamics In molecular dynamics [53, 54], the primary task is to simulate atomic motion under a potential energy surface or force field using very small time steps to generate trajectories, from which statistics, such as conformational ensembles, interaction stability, and conformational transitions, can be extracted. The key bottlenecks of classical molecular dynamics fall into two categories—(i) force field accuracy is limited: widely used force fields are approximate and may become unreliable when electronic structure effects are significant, such as polarization, charge or proton transfer, bond breaking and forming, or in systems containing metal centers; (ii) sampling is slow: many biologically relevant processes are rare events that occur only on long timescales, causing computational costs to increase rapidly. In systems involving strong polarization, metal centers, bond changes, or proton transfer, classical force fields can introduce errors into trajectories and free energy estimates. Full-system quantum chemistry is usually too expensive for routine use. A more practical strategy is therefore to use classical molecular dynamics for most of the trajectory, while applying higher accuracy calculations only to selected regions or critical conformations [55, 56]. A small number of high accuracy calls may improve the reliability of key local results. By contrast, directly addressing global sampling bottlenecks with quantum resources remains less mature and is currently harder to justify under NISQ constraints (see Section ‘Hybrid classical–quantum molecular dynamics’ for more details). Hybrid classical–quantum computing Background This section focuses on practical quantum computing under NISQ conditions [57–59], at which hardware is limited by noise, limited number of available quantum bits, and limited coherence time. As a result, reliably running quantum algorithms that require deep circuits and full error correction remains challenging. Under these constraints, a more feasible approach is a hybrid quantum–classical paradigm, in which a quantum processor functions as a specialized module within a classical workflow, handling a small number of critical subroutines. Classical computing continues to manage large-scale data processing, candidate generation, and sampling, as well as postprocessing and validation. At the algorithmic level, NISQ applications typically rely on variational quantum algorithms and hybrid quantum–classical frameworks. In deep learning–based applications, these frameworks are commonly implemented through parameterized quantum circuits, variational quantum classifiers, quantum kernels, and related hybrid learning architectures. Other representative examples include variational quantum eigensolver (VQE) [60] and quantum approximate optimization algorithm (QAOA) [61, 62]. VQE is mainly used for energy estimation and correction in small-scale quantum chemistry and electronic structure problems, while QAOA is mainly used to obtain approximate solutions to discrete combination optimization problems. Since the practical application value of these hybrid strategies depends not only on the algorithm design, but also on the baseline strength, calibration, and affected downstream decision-making points. Table 3 outlines a minimum evaluation template, which will be used to explain the representative research discussed in Sections ‘Classical–quantum hybrid strategies for deep learning models’, ‘Hybrid classical–quantum docking’, and ‘Hybrid classical–quantum molecular dynamics’. Table 3Minimal evaluation template for quantum-assisted computational strategies in drug discovery.WorkflowQuantum moduleBaseline or comparisonEvaluation outputResource reportingDeep learning–based VSFeature stage augmentation in predictive models; quantum priors or generators in generative modelsFully classical predictor or generator for the same task; matched classical control when feasiblePrediction: AUROC, AUPRC, F1, RMSE, MAE, Pearson, or Spearman. Generation: validity, uniqueness, diversity, novelty, QED, SA, Log P, docking score, or activityQubit number, circuit depth, shots, simulator or hardware, noise setting, seeds, runtimeDocking-based VSSite identification, pose search, and fragment-based flexible dockingClassical site prediction, pose search, graph search, clique search, or fragment placement baseline matched to the selected docking substepSite accuracy, pose RMSD, binding mode recovery, success probability, ground state overlap, clique recovery, docking configuration quality, or enrichmentGraph or QUBO size, poses or fragments tested, qubits or modes, circuit depth or annealing setting, samples, hardware or simulator, runtimeMolecular dynamicsGround state AIMD, annealer-based MD propagation, and excited state MDClassical MD, classical AIMD, QM/MM, or high accuracy reference calculation matched to the quantum-assisted simulation taskEnergy error, force error, reference agreement, trajectory stability, structural deviation, vibrational accuracy, excited state agreement, or uncertaintySubsystem size, atoms or orbitals, corrected frames or states, qubits, circuit depth, shots, measurement strategy, hardware or simulator, runtime Minimal evaluation template for quantum-assisted computational strategies in drug discovery. Classical–quantum hybrid strategies for deep learning models Under practical NISQ conditions, quantum computing is better positioned as an augmentation module within drug screening and molecular design workflows based on deep learning rather than as a replacement for the full classical deep learning model [63]. In hybrid classical–quantum strategies, classical models handle scalable representation learning and coarse ranking or generation, whereas quantum modules are applied only to selected submodules under feasible resource limits. With evaluation criteria at the workflow level outlined in Table 3, related representative study designs, quantum components, classical comparators, and operational roles are summarized in Table 4, while indicative resource feasibility and data provenance, endpoints, and realism caveats are summarized in Tables 6 and 7. Table 4Representative quantum-assisted drug discovery studies and their practical evaluation context.IDRefs.TaskQuantum componentClassical comparator(s)Stage of use and metricsScope notes1[64]Affinity predictionHybrid quantum module inserted after classical feature encoding within HQDeepDTAFDeepDTAF; DeepDTA; Pafnucy; TopologyNet; AEScore; KDEEPPrioritization before experimental validation; MAE, RMSE, Pearson, SD, and CIRetrospective benchmark level affinity regression study; no wet lab validation and no calibration analysis2[65]Affinity predictionHybrid quantum layers in GIN, GCN, and CNN modelsGIN, GCN, and CNN; also ML-DTI and YuelScreening support; Pearson, Spearman, and RMSE; downstream support evaluated with MM/GBSA, MM/PBSA, and MDAffinity prediction support setting with downstream MM/GBSA, MM/PBSA, and MD checks; still benchmark scale rather than a full prospective screening workflow3[66]Affinity predictionHybrid quantum fusion model combining 3D-CNN and SG-CNN with a QNNClassical fusion model; prior classical 3D-CNN and SG-CNN approachesScreening support; R2, MAE, MSE, Pearson, and SpearmanAffinity prediction benchmark based on structural information; useful for screening support, but still retrospective and not a complete workflow for screening large libraries4[67]Property predictionQNEM + QEEM within hybrid QEGNNClassical GCN, CCN, GIN, and GATPrioritization before experimental validation; accuracy and loss were reportedRetrospective molecular property benchmark using standardized or balanced subsets; results may not reflect screening imbalance or deployment conditions in real settings5[68]KRAS inhibitor generationQCBM quantum prior + LSTM generator + Chemistry42 reward/filterREINVENT; SMILES-/SELFIES-VAE; MoFlow; SMILES-/SELFIES-LSTM-HC; GB-GA; JANUS; vanilla LSTMTriage before synthesis; Tartarus success rate, docking score, and protein ligand interaction ranking; 1 M compounds per model; top 15 compounds synthesized; SPR and cell assaysHybrid workflow from molecular generation to experimental validation with synthesis and assay confirmation, but still biased toward known KRAS chemistry and docking-based proxies; no definitive quantum advantage established6[69]Molecule generationHybrid quantum generator and hybrid quantum cycle component in MolGANMolGAN; Cycle-MolGAN; prior hybrid quantum GAN modelsMolecule generation; QED, SA, Log P, validity, uniqueness, and diversityBenchmark scale small molecule generation on QM9/PC9 using proxy molecular metrics; validation remained mainly simulator-based, with no experimental follow-up in drug discovery7[70]Docking site identificationModified Grover search on a protein interaction space representation, with quantum search and site evaluation modulesNo direct comparison with standard classical tools for binding site prediction was reportedBinding site localization before docking identification demonstrated on a quantum simulator and a real quantum computerBinding site localization before docking rather than full prediction of docking poses; simplified encoding of interaction space with only selected interaction types considered8[51]Docking pose searchQAOA/DCQAOA for optimization of the maximum weight clique in the protein ligand interaction graphConventional QAOA; exact diagonalization; PLIP consistency checkPose selection; success probability and overlap with the ground stateProof of concept docking pose selection on a very small set of protein ligand complexes; no large library screening or re-ranking evaluation at the candidate level9[71]Docking pose samplingGaussian Boson Sampling (GBS) for maximum weight clique search in a binding interaction graphClassical random search; greedy shrinking; local searchConfiguration search; success rate for maximum weight clique recovery and correct binding mode recoveryProof of principle study for docking configuration search in a simplified interaction graph; strongly dependent on graph construction and not a complete docking benchmark10[72]Docking pose searchGBS on a binding interaction graph reduced to a maximum weighted clique problemClassical postprocessing/reconstruction steps are used, but no full classical docking baseline is replacedDocking configuration search; predicted docking configurations are selected by clique weight; benchmarked on a known ligand receptor interaction rather than standard metrics for large-scale screeningUses a simplified graph representation based on pharmacophores and numerical simulation; explicitly improves configuration sampling rather than the main difficulty of scoring11[73]Flexible dockingFragment-based QUBO docking with clash, bond, and single placement constraints; solved by SQBM+Classical fragment placement + downstream energy minimization; no direct flexible docking benchmark reportedDocking configuration search; predicted docking configurations are selected by clique weight; benchmarked on a known ligand receptor interaction rather than standard metrics for large-scale screeningProof of concept re-docking on aldose reductase; more realistic than rigid matching because internal degrees of freedom are represented through fragment consistency, but still based on simulation12[74]Ground state AIMDVQE electronic energies with numerical gradients via HFT + correlated samplingClassical AIMD formulation; exact/FCI references for validationGround state AIMD; proof of concept H2 trajectories on IBM devices; energies and gradients are centralProof of concept for small systems; larger systems were assessed mainly through classical reference wave functions13[75]Annealer assisted MD propagationQDE framework using a D Wave quantum annealer, followed by classical greedy postprocessingAnalytical reference solution; classical greedy postprocessingMD propagation; H2 vibrational trajectories in harmonic, anharmonic, and dissociative regimes; trajectory accuracy/convergence centralProof of principle for very small systems; focuses on trajectory propagation rather than electronic structure calculation14[76]Excited state MDRestricted VQE excited state MD with SSVQE/VQST state preparationExact adiabatic S1 simulations as referenceExcited state MD; S1 dynamics of H2 and CH2NH; agreement with exact adiabatic results except near conical intersectionProof of concept for small molecules; effectively diabatic near conical intersections; not yet practical for large-scale MD Representative quantum-assisted drug discovery studies and their practical evaluation context. Integration at the model level uses quantum modules within otherwise classical learning models to enhance selected computational stages (Fig. 2a). In the drug discovery literature reviewed here, such integration has mainly appeared in two broad forms, while related representative study designs, quantum components, classical comparators, and operational roles across the reviewed studies are summarized in Table 4. The first comprises predictive models. Within this group, the main pattern is to introduce quantum modules at the representation or feature stage. For example, Jeong et al. [64] introduced a hybrid quantum neural network after classical feature encoding, while related hybrid affinity models [65, 66] similarly introduced small quantum submodules after classical encoders or feature fusion stages for binding affinity prediction. This predictive form also includes quantum embedded graph models for molecular property prediction, as illustrated by quantum embedded graph neural network (QEGNN) [67], where a quantum module is coupled to graph-based molecular representations. The second broad form comprises generative molecular design models, in which quantum components act as priors or generators, including generation of Kirsten rat sarcoma viral oncogene homolog (KRAS) inhibitor generation using a quantum circuit Born machine (QCBM) prior [68] and hybrid quantum cycle molecular generative adversarial network (HQ-Cycle-MolGAN) for small molecule generation [69]. Although task categories and model architectures vary, these studies show similar integration patterns: quantum components are inserted into specific submodules of classical deep learning models instead of replacing the entire predictor or generator. However, the current evidence is still uneven between different tasks and assessment settings. Most studies are still based on simulation, limited hardware verification, or benchmark test level evaluation. Only a few studies progress to prospective screening. The study of KRAS inhibitors is a significant exception, because it has developed from the computational generation stage to the stage of compound synthesis and experimental verification. Hybrid classical–quantum docking Under the constraints of NISQ, the more realistic quantum participation docking method is not to replace the entire docking software stack, but to use quantum computing as a modular plugin to handle selected sensitive and controllable docking steps. Classical computing continues to be responsible for pose generation, geometric constraint processing, and large-scale enumeration and rough scoring. Quantum modules only intervene in specific stages, such as position recognition, pose search, or fragment-based flexible docking. Their role is to support selected docking-related subproblems under a limited quantum call budget, rather than to replace the full docking workflow [56, 63]. The notes on the relevant representative research design, resource feasibility, and reality are summarized in Tables 5–7. Table 5Indicative resource feasibility for representative quantum use cases.IDRefs.Use caseQuantum scaleCircuit depthShotsError mitigationSettingPractical caveat1[64]Affinity prediction7–9 qubits; 9 qubits in the main configuration20 layers; reported cost ~260 versus ~10 480 in the compared settingNot reportedNo explicit QEM; future QEM was suggestedClassical simulation only (PennyLane/PyTorch); no real hardwareAmong the tested settings, NN angle appears the most plausible under NISQ conditions, although performance degrades under noise2[65]Affinity prediction6 qubits in graph layers; 10 qubits after drug–protein fusionPQC with m entanglement layers followed by one single qubit layer; exact m deferred to SINot reportedNot reportedClassically simulated PQC-based hybrid deep learningNo real hardware execution; noise, error mitigation, and hardware feasibility not analysed3[66]Fusion affinity prediction4 qubits (amplitude) or 8 qubits (hybrid angle)PQC Circuit 1, L = 10Not reportedDREM was effective mainly for P ≤ .05Classical simulation only; no execution on actual quantum hardware reportedSimulation only setting; DREM appears effective mainly under low noise4[67]Property prediction4–9 qubits; validated on 72 qubit Wukong hardwareShallow HE-VQC; about 4 CZ layersNot reportedNot reportedSimulator + Wukong hardware validationValidation on real hardware was reported, but only on small and standardized datasets5[68]KRAS generation16 qubits4-layer QCBMNot reportedMentioned, but not specifiedIBM Guadalupe + classical baselinesEarly experimental validation, but small synthesis scale, micromolar activity, and no definitive quantum advantage6[69]Molecule generation8 qubits; 2 ancilla3 variational layers +8 encoding layers2 × 105 shotsNo explicit QEM; only comparison between noisy and ideal simulatorsClassical training followed by noisy and ideal IBM Brisbane simulator runs; no execution on real quantum hardwareValidation was mostly based on simulation; the reported high entropy state does not by itself establish a stable practical advantage7[70]Docking site identification2 qubits per interaction site; total size depends on the encoding of ligand and protein interaction spaceModified Grover search combined with a site evaluation circuit; depth not reportedNot reportedNot reportedQiskit simulator + real quantum computer validationUpstream site identification only; simplified interaction space encoding using selected interaction types8[51]Docking pose selection6–12 qubitsQAOA/DCQAOA with p layers; p = 8 worked best in the DCQAOA case using 8 qubits5000 samples; 500 trialsNo explicit QEM was used; the analysis only considered depolarizing noiseMindQuantum simulation onlyUseful for small docking subproblems, but challenges in scaling and local optima remain9[71]Docking graph search24 optical modesPhotonic GBS; qubit gate depth is not applicable105 and 104 samplesNo explicit QEM; robustness to photon loss was discussedPhotonic simulation using a simplified coarse-grained graphSupports docking configuration search rather than full scoring function evaluation10[72]Docking pose searchPhotonic GBS; optical modes scale with the size of the binding interaction graph, with 24 interaction pairs in the benchmark caseGate circuit depth is not applicable; Gaussian optical networkNot reportedNo explicit QEM; robustness to photon loss discussedNumerical photonic simulation; no execution on real hardware in this studyImproves sampling of docking configurations in a simplified pharmacophore graph; does not address the main difficulty of scoring function evaluation11[73]Flexible dockingQUBO size depends on the fragment placement problem; solved by SQBM+ simulated quantum annealerBased on annealing; no gate circuit depth reportedNot reportedNot reportedProof of concept re-docking using SQBM+ simulation followed by classical energy minimizationMore realistic than rigid matching because internal degrees of freedom are represented through fragment consistency, but validation remains based on simulation12[74]Ground state AIMDH2 mapped from 4 qubits to 2, and further to 1 qubitUCCSD circuit; H2 implementations using 1 and 2 qubits295 × 103 on IBM Q VigoMeasurement error mitigation (Qiskit Ignis)IBM Q Vigo hardware + state vector simulationSmall H2 proof of concept; gradients are numerical rather than fully quantum-analytic; larger systems assessed mainly via classical validation13[75]MD propagation based on an annealerD-Wave 2000Q; 2048 physical qubits, up to 64 fully connected logical qubitsQUBO formulation with an annealing setting; gate circuit depth is not applicableNot reportedNo explicit QEM was used; the best accuracy was obtained after classical greedy postprocessingReal D-Wave 2000Q annealerH2 vibrational dynamics only; proof-of-principle trajectory propagation, not electronic structure MD14[76]Excited state MD4 qubits for H2; 6 qubits for CH2NHHE ansatz; SSVQE depth > 4 for H2 and > 12 for CH2NH; shallower VQST states also usedNot reportedNo explicit QEM; measurement partitioning on commuting observablesSampling simulations and runs on the IBM Kawasaki Falcon deviceProof of concept for small molecules; effectively diabatic near conical intersections Indicative resource feasibility for representative quantum use cases. Table 6Data provenance, supervision targets, and bias/realism caveats in representative quantum-assisted drug discovery studies.IDRefs.TaskData source/benchmark/scale/splitSupervision target/endpointPotential bias/realism caveat1[64]Affinity predictionPDBbind v2016 core benchmark; protein/pocket structures in PDB format and ligands represented by SDF/SMILES; core 2016 used for evaluation; fixed token lengths 1000/63/150Supervised regression: −log(Ki/Kd/IC50)Retrospective benchmark only; no wet lab validation; NISQ feasibility was assessed mainly through simulation, without demonstrated deployment on a real QPU2[65]Affinity predictionDavis, Metz, KIBA, and PDBbind benchmarks with totals of 31 824/35 259/160 278/13 196; reported train and test splits with five-fold cross validation; plus a liver disease screening set of 1500 pairs for application useSupervised regression for protein ligand binding affinity; labels derived from Kd, Ki, and KIBA; evaluated by Pearson, Spearman, and RMSERetrospective benchmark study with quantum layers evaluated only by simulation; no validation on real devices or in wet lab experiments; downstream screening was checked mainly by MM/GBSA, MM/PBSA, and MD3[66]Affinity predictionPDBbind 2020; refined set 5316 for training/validation and core set 285 for testing; 25% of the refined set used for validationSupervised regression: binding affinity, evaluated by R2, MAE, Pearson, Spearman, and MSEValidation based on simulation rather than real devices; DREM mitigation was shown mainly at P ≤ .054[67]Property predictionClinTox, HIV, and BACE; original sizes 1491, >40 k, and 1522; standardized to 500 each; 80:20 train/test splitSupervised classification: molecular property, activity, or toxicity labelDownsampling and label balancing reduce real world realism; still retrospective, although also tested on Wukong hardware5[68]KRAS generation~650 known KRAS inhibitors; 850 k STONED-SELFIES analogs; top 250 k from 100 M Enamine REAL; ~1.1 M total; 1 M generated compounds screened; 15 compounds synthesized; no explicit splitGoal-directed molecular generation using Chemistry42 and local filters, followed by docking and protein ligand interaction ranking; endpoints included SPR, MaMTH DS IC50/activity, and cell viabilityBiased toward known KRAS chemistry and docking proxies; only 15 compounds were tested; still early stage hits; no definitive quantum advantage established6[69]Molecule generationQM9 (~134 k) and PC9 (~99 k); single or combined dataset training; small validation sets (100/250); no standard split emphasisGoal-directed molecular generation using Chemistry42 and local filters, followed by docking and protein ligand interaction ranking; endpoints included SPR, MaMTH DS IC50/activity, and cell viabilitySmall validation sets; optimization based on proxy metrics; reported high entropy state; hardware assessment relied mainly on noisy simulation7[70]Docking site identificationRepresentation of protein ligand interaction space with validation on both simulator and real quantum computer; no standard benchmark split reportedNo supervised target; endpoint: identification of candidate docking sitesSimplified lattice/interaction space encoding; only selected interaction types were included, and the study addresses upstream site localization rather than full docking pose prediction8[51]Docking pose selectionThree case studies: 8SKH, 3HAC, and 5F4L; reduced to 6-, 8-, and 12-qubit BIG/QUBO instances; no train/test splitNo supervised target; endpoint was the optimal docking pose or the maximum weight cliqueRelies on substantial pharmacophore pruning and heuristics; one case used proximity to the actual binding pose for point selection; scaling remains limited by local optima and parameter growth9[71]Docking searchCase study of a known TACE aryl sulfonamide interaction; coarse-grained graph with 4 ligand nodes and 6 receptor nodes, giving 24 interaction pairs; no train and test split reportedNo supervised target; endpoint was the maximum weight clique or docking configurationDoes not address the limitations of scoring function evaluation; relies on a simplified coarse-grained representation; some benchmarking was performed after fixing the clique size10[72]Docking pose searchKnown interaction between TACE and aryl sulfonamide; coarse-grained pharmacophore graph with 4 ligand nodes and 6 receptor nodes, giving 24 interaction pairs; no train and test split reportedNo supervised target; endpoint: maximum weighted clique/docking configuration searchUses a simplified pharmacophore representation and numerical simulation; improves configuration sampling rather than addressing the main difficulty of scoring function evaluation; some benchmarking was performed after fixing the clique size11[73]Flexible dockingAldose reductase, ALDR, PDB ID: 2HV5; proof of concept re-docking using 343 subregions and 3005 candidate fragment placements; no train and test split reportedNo supervised target; endpoint: reconstructed binding pose/re-docking RMSDSingle target proof of concept using SQBM+ simulated quantum annealer; realism is limited by fragment decomposition and a manually defined docking region, although internal degrees of freedom are better represented than in rigid formulations12[74]Ground state AIMDProof-of-concept AIMD study for H2 on IBM quantum devices; larger molecules were assessed using classical FCI wave functions; no train and test split reportedNo supervised target; endpoint: AIMD trajectory propagation from VQE energies and numerical gradientsProof of concept for very small systems; real hardware MD was shown only for H2, while relevance to larger systems was assessed mainly through classical validation13[75]MD propagation based on an annealerH2 vibrational dynamics on D-Wave 2000Q; nearly harmonic, highly anharmonic, and dissociative regimes; no train/test splitNo supervised target; endpoint: trajectory propagation/convergence to analytical reference solutionProof of principle study using a diatomic molecule; focuses on trajectory propagation rather than electronic structure MD, with results improved by hybrid classical postprocessing14[76]Excited state MDExamples of S1 excited state MD for H2 and CH2NH small molecules; no train and test split reportedNo supervised target; endpoint: excited state MD trajectories compared with exact adiabatic S1 simulationsSmall molecule proof of concept; agreement degrades near conical intersection, where the method becomes effectively diabatic rather than fully adiabatic Data provenance, supervision targets, and bias/realism caveats in representative quantum-assisted drug discovery studies. Table 7Abbreviations used in this manuscript.AbbreviationFull nameADMETAbsorption, distribution, metabolism, excretion, and toxicityBACEβ-site amyloid precursor protein cleaving enzymeCADDComputer-aided drug designCASCIComplete active space configuration interactionCNNConvolutional neural networkDFTDensity functional theoryDTIDrug–target interactionGATGraph attention networkGCNGraph convolutional networkGINGraph isomorphism networkHFHartree–FockHIVHuman immunodeficiency virusHQNNHybrid quantum neural networkIBMInternational Business MachinesIC50Half maximal inhibitory concentrationKdDissociation constantKiInhibition constantLogPOctanol/water partition coefficientLSTMLong short-term memoryMAEMean absolute errorMDMolecular dynamicsMMMolecular mechanicsMM/GBSAMolecular mechanics/generalized Born surface areaMM/PBSAMolecular mechanics/Poisson–Boltzmann surface areaMSEMean squared errorNISQNoisy Intermediate-Scale QuantumPDBProtein Data BankPDBbindProtein Data Bank binding database/benchmark setPQCParameterized quantum circuitQAOAQuantum approximate optimization algorithmQCBMQuantum circuit born machineQEDQuantitative estimate of drug-likenessQMQuantum mechanicsQM/MMQuantum mechanics/molecular mechanicsQNNQuantum neural networkQPUQuantum processing unitQSVMQuantum support vector machineRPearson correlation coefficientR2Coefficient of determinationRMSERoot mean squared errorSASynthetic accessibilitySELFIESSELF-referencing embedded stringsSMILESSimplified molecular input line entry systemSPRSurface plasmon resonanceVAEVariational autoencoderVQEVariational quantum eigensolverVSVirtual screeningAIMD Ab initio molecular dynamicsGBSGaussian Boson samplingQUBOQuadratic unconstrained binary optimizationQEMQuantum error mitigationUCCSDUnitary coupled cluster with singles and doublesSSVQESubspace search variational quantum eigensolverVQSTVariational quantum state transcriptionHE ansatzHardware efficient ansatzHE-VQCHardware efficient variational quantum circuitPLIPProtein ligand interaction profilerSDStandard deviationCIConfidence interval Abbreviations used in this manuscript. In the molecular docking literature reviewed in this article, quantum integration mainly appears in three forms: site recognition, pose search, and flexible docking (Fig. 2b). In site identification, quantum search is used to locate candidate ligand binding regions before full pose generation. For example, in the research of Liliopoulos et al. [70], they applied the improved Grover search to the spatial representation of protein interaction. In posture search, molecular docking is re-expressed as graph search, group search, or matching problem to identify reasonable docking configuration. Representative examples include Banchi et al. [71] using Gaussian boson sampling for maximum weighted group search; Ding et al. [51] using quantum approximate optimization algorithm (QAOA) and digitized counterdiabatic QAOA (DCQAOA) in the combined interaction diagram; and Zha et al. [72] using grid point matching and characteristic atom matching to collect posture. The sample is coded as a quadratic unconstrained binary optimization (QUBO) instance. In flexible docking, fragment-based models begin to consider the internal degree of freedom of the ligand. For example, Yanagisawa et al. [73] formulated protein ligand flexible docking as a QUBO problem by incorporating protein fragment interactions, fragment clashes, covalent connectivity, and single placement constraints. Although there are differences in the definition of tasks, the reviewed studies follow a common practical logic: quantum modules are applied to specific subproblems related to docking, not the complete docking workflow. Current presentations are still limited in terms of scale and authenticity, and usually rely on simplified representations, proof-of-concept settings, or validation based on simulation. Hybrid classical–quantum molecular dynamics Under the constraints of NISQ, the real path of quantum computing in molecular dynamics does not replace the entire molecular dynamics process, but provides electronic structure level correction for specific areas where the classical force field is not reliable. Classical engines remain responsible for trajectory propagation, thermostat and barostat control, and long timescale sampling. Quantum modules are instead most relevant to local regions involving bond formation or breaking, metal coordination, proton transfer, or strong polarization, where fixed topology or fixed charge force fields may fail. In such settings, a quantum processor could be used to estimate local electronic energies, forces, or interaction components, which are then passed back to the classical molecular dynamics or QM/molecular mechanics framework. Current demonstrations remain proof of concept and are still limited to small systems, but they clarify the intended division of labor: classical computing provides scale and sampling, whereas quantum computing is used to improve the accuracy of selected electronically complex regions [77]. In the molecular dynamics literature reviewed here, representative quantum participation has mainly appeared in three forms: ground state ab initio molecular dynamics, molecular dynamics propagation based on annealers, and excited state molecular dynamics (Fig. 2c). The first form is ground state ab initio molecular dynamics. Fedorov et al. [74] used VQE to compute electronic energies on a quantum computer and propagated nuclei classically on the resulting potential energy surface. This enabled proof-of-concept ab initio molecular dynamics simulations of H2 on IBM quantum devices, with gradients evaluated numerically. The second form is based on the molecular dynamics propagation of the annealer. As Gaidai et al. [75] said, they developed a quantum differential equation framework and used the D-Wave 2000Q quantum annealer to propagate the vibration trajectory of H2. The third form is excited state molecular dynamics. Hirai [76] proposed an excited state molecular dynamics method based on variational quantum algorithms and demonstrated S1 state dynamics for H2 and CH2NH. In these studies, quantum components enter the molecular dynamics cycle at the level of energy evaluation, calculations related to forces, or trajectory updates, and the broader simulation workflow remains hybrid rather than completely quantum. Accordingly, under current NISQ constraints, the most practical role of quantum computing in molecular dynamics is selective local correction or targeted trajectory assistance within an otherwise classical sampling framework. Rather than being used for full-system sampling, it provides high accuracy electronic structure and energy information for a localized key region or supports the propagation of simplified trajectories or excited state dynamics in small systems, which is then fed back into the classical molecular dynamics loop. Given current limitations in qubit count, noise, and measurement overhead, practical implementations require strict control over subsystem size and the frequency of quantum calls. Consequently, a realistic engineering strategy is to call the quantum module only at a limited number of critical conformational states or time points, or to restrict it to small model systems and proof-of-concept dynamical settings, while leaving most trajectory propagation to classical force fields [78, 79]. Challenges and outlook Current major challenges Although the preceding sections describe practical strategies for embedding quantum modules into deep learning–based virtual screening, molecular docking-based virtual screening, and molecular dynamics, several key hurdles remain before these approaches can advance beyond the proof-of-concept stage and become reproducible, comparable, and deployable hybrid workflows. These challenges can be broadly grouped into four categories: hardware and noise constraints, interface alignment, benchmarking and reproducibility, and relevance to decision points, as described below: (1) Hardware and noise constraints Gate errors, decoherence, and measurement noise on NISQ devices require quantum modules to remain shallow, measurement-light, and sparingly invoked; otherwise, output variance and operational costs grow rapidly, undermining reproducibility, and obscuring marginal gains [80, 81]. (2) Interface alignment The mismatch between molecular inputs, quantum representations, and classical outputs remains a central issue. High-dimensional molecules and protein representations usually have to be compressed into low-dimensional forms that can be processed by quantum devices. In turn, quantum outputs (such as learned features or energy correction) must be aligned with classical fractions in scale and statistical interpretation to achieve interpretable fusion, calibration, and reliable reordering [82]. (3) Benchmark test and repeatability The reliability of hybrid methods depends on whether they are superior to the strong classical baseline under strict experimental conditions and whether they can withstand the effects of random seed and sampling noise (see Table 3). In addition to the overall indicators, the evaluation should also emphasize early enrichment, sorting stability, and calibration, while clearly reporting the circuit depth, number of samples, and optimizer settings [83]. (4) Relevance with decision-making points Even if the hybrid method shows measurable improvements in benchmark indicators, it does not automatically prove that it is reasonable to use quantum integration in practice. Its value ultimately depends on whether the gain is large enough to influence downstream decision-making points under the available runtime and quantum call budget, not just improve intermediate indicators [59, 63]. In order to meet these practical challenges, Table 3 provides a minimum evaluation template for the drug discovery workflow based on quantum computing, which clarifies the expected baseline, calibration indicators, and virtual screening based on deep learning, virtual screening based on molecular docking, and virtual screening of molecular dynamics under the constraints of running time. Table 4 summarizes a practical evaluation and reporting framework suitable for the typical study of quantum-assisted drug discovery, including the workflow stage, quantum action, classical baseline, decision-making impact, and benchmarking background. In addition, Table 5 summarizes the indicative dimensions of resource feasibility in typical use cases, including quantum bit size, circuit depth, number of experiments, error mitigation hypothesis, and hardware or analog settings. Table 6 further summarizes the data provenance, supervision targets, and caveats related to bias and realism in representative studies discussed in this review. Box 1 summarizes these practical requirements as a minimal benchmarking checklist for studies of quantum-assisted drug discovery. Together, these materials highlight that the key barrier is not whether a quantum module can be inserted into a workflow in principle, but whether it can deliver reproducible gains relevant to decisions under realistic NISQ constraints. Box 1. Minimal benchmarking checklist for quantum-assisted drug discovery studies Strong classical baselines. Compare against relevant classical baselines at the same workflow stage rather than against weak or mismatched references only. Strong classical baselines. Compare against relevant classical baselines at the same workflow stage rather than against weak or mismatched references only. Multiple random seeds. Report variability across repeated runs; for stochastic quantum components, separate seed effects from shot noise effects when possible. Multiple random seeds. Report variability across repeated runs; for stochastic quantum components, separate seed effects from shot noise effects when possible. Sensitivity analysis for shot noise. Evaluate whether conclusions remain stable under different shot budgets or realistic sampling noise. Sensitivity analysis for shot noise. Evaluate whether conclusions remain stable under different shot budgets or realistic sampling noise. Ablation studies. Distinguish gains from the quantum module itself from gains due to feature fusion, re-ranking strategy, workflow redesign, or other auxiliary components. Ablation studies. Distinguish gains from the quantum module itself from gains due to feature fusion, re-ranking strategy, workflow redesign, or other auxiliary components. Resource reporting. Explicitly report qubit count, circuit depth, shot budget, optimizer, and assumptions for error mitigation. Resource reporting. Explicitly report qubit count, circuit depth, shot budget, optimizer, and assumptions for error mitigation. Decision relevance. State the exact decision point affected by the quantum module and the corresponding runtime budget or intervention scope. Decision relevance. State the exact decision point affected by the quantum module and the corresponding runtime budget or intervention scope. Conclusions and future outlook In the foreseeable NISQ era, quantum computing in drug discovery is most appropriately framed as a coprocessor plugin, and its value lies in delivering verifiable marginal gains for a small number of critical subroutines, rather than replacing classical pipelines end-to-end. In many discovery workflows in the real world, computational ranking still needs to support difficult prioritization decisions at a late stage before experimental validation, reflecting the fact that current methods do not always provide sufficient confidence for final selection. From this perspective, a meaningful practical test for quantum computing is not merely whether it improves benchmark metrics, but whether it can improve ranking reliability and decision confidence at these important decision points. Across deep learning–based virtual screening, molecular docking-based virtual screening, and molecular dynamics, practical hybrid strategies are beginning to emerge. Classical computing continues to handle scalable representation learning, pose generation, coarse scoring, trajectory propagation, sampling, and robust validation, while quantum modules intervene either within predictive or generative deep learning models, or at selected substeps in docking and molecular dynamics. In the representative studies reviewed here, these roles include augmentation at the representation or feature stage and quantum priors or generators in deep learning; site identification, pose search, and flexible docking based on fragments in docking; and ground state ab initio molecular dynamics, molecular dynamics propagation based on annealers, and excited state molecular dynamics in molecular dynamics. Looking ahead, a realistic path for quantum computing in drug discovery is to remain within hybrid workflows and focus on a small set of decision-sensitive and computationally intensive steps. The goal is not to replace the full pipeline with quantum computing, but to achieve targeted improvements within hybrid workflows. Progress should therefore be judged by whether quantum modules deliver reproducible gains at specific downstream decision points under realistic runtime and resource constraints. Key Points In the Noisy Intermediate-Scale Quantum (NISQ) era, quantum computing in drug discovery is best developed as a modular coprocessor for selected subroutines that are sensitive for decisions, rather than as a replacement for full classical workflows. In the Noisy Intermediate-Scale Quantum (NISQ) era, quantum computing in drug discovery is best developed as a modular coprocessor for selected subroutines that are sensitive for decisions, rather than as a replacement for full classical workflows. In screening and design based on deep learning, the most plausible roles of quantum modules in the near term are augmentation at the representation or feature stage in predictive models and the use of quantum priors or generators in molecular design models. In screening and design based on deep learning, the most plausible roles of quantum modules in the near term are augmentation at the representation or feature stage in predictive models and the use of quantum priors or generators in molecular design models. In virtual screening based on docking, realistic quantum roles are concentrated at selected substeps related to docking, especially site identification, pose search, and flexible docking based on fragments, rather than full docking replacement. In virtual screening based on docking, realistic quantum roles are concentrated at selected substeps related to docking, especially site identification, pose search, and flexible docking based on fragments, rather than full docking replacement. In molecular dynamics, current representative quantum participation spans ground state ab initio molecular dynamics, trajectory propagation based on annealers, and excited state molecular dynamics, while most sampling at a large-scale remains classical. In molecular dynamics, current representative quantum participation spans ground state ab initio molecular dynamics, trajectory propagation based on annealers, and excited state molecular dynamics, while most sampling at a large-scale remains classical. Quantum computing should be evaluated not only by benchmark accuracy, but also by decision relevance, ranking and calibration reliability, and comparison with strong classical baselines under explicit runtime and resource constraints. Quantum computing should be evaluated not only by benchmark accuracy, but also by decision relevance, ranking and calibration reliability, and comparison with strong classical baselines under explicit runtime and resource constraints. In order to ensure the repeatability and comparability of the research results, the report should clearly explain the quantum bit scale, circuit depth, experimental budget, error mitigation hypothesis, optimizer settings, and the specific position and frequency of quantum calls in the workflow. In order to ensure the repeatability and comparability of the research results, the report should clearly explain the quantum bit scale, circuit depth, experimental budget, error mitigation hypothesis, optimizer settings, and the specific position and frequency of quantum calls in the workflow.