Beyond the hype: an empirical assessment of quantum computing in bioinformatics

1 Introduction

Quantum computing has emerged as a promising paradigm with the potential to address some of the most computationally intensive problems in bioinformatics, ranging from molecular simulation (Berryman et al., 2022; Brokowski et al., 2022) and drug discovery (Aramyan et al., 2022; Zhang Y. et al., 2026) to genomics (Boev et al., 2021; Nałęcz-Charkiewicz and Nowak, 2022) and machine learning (Li et al., 2021; Saggi et al., 2026). A rapidly growing body of review literature has explored this topic, providing valuable overviews of quantum algorithms, theoretical advantages, and prospective applications across biological domains (Baiardi et al., 2023; Chakraborty et al., 2023; Cordier et al., 2022; Ghamsari, 2025; Kumar et al., 2024; Maheshwari et al., 2022; Marchetti et al., 2022; Maurizio and Mazzola, 2025; Mokhtari et al., 2024; Nałęcz-Charkiewicz et al., 2024; Outeiral et al., 2021b; Pal et al., 2024a,b; Pallavi and Prasanna Kumar, 2025; Sung et al., 2026). These works have contributed to shaping the conceptual landscape of quantum bioinformatics, clarifying its scope, terminology, and potential impact.

However, a careful examination of the existing literature reveals a fundamental limitation: the majority of the reviews available are mostly theoretical and prospective. As widely acknowledged, current quantum hardware remains in the noisy intermediate-scale quantum (NISQ) computer era (Preskill, 2018), limiting the feasibility of large-scale, real-world applications, thus supporting the speculative nature of many claims.

Several key aspects that are central to assessing the practical relevance of quantum computing in bioinformatics remain largely unexplored in existing literature reviews. First, there is a lack of systematic analysis of empirical performance, with very limited attention paid to real experimental results obtained on quantum machines. Second, comparative evaluations across different hardware platforms are almost entirely absent, preventing a clear understanding of the technological landscape. Third, the literature lacks a rigorous discussion of benchmarks and classical baselines, making it difficult to assess whether the reported quantum approaches provide meaningful advantages over state-of-the-art classical methods. While there have been recent efforts toward standardized benchmarking of quantum applications (Koch et al., 2026) and quantum computers (Lall et al., 2025), we are not aware of any work specific to bioinformatics quantum benchmarking.

In addition, important aspects related to scientific reproducibility are often overlooked. The availability of source code, datasets, and executable workflows is rarely addressed, and there is no consolidated effort to evaluate the reproducibility of published quantum bioinformatics studies. This represents a critical gap, particularly in a field where results are often sensitive to implementation details, hardware constraints, and algorithmic choices.

The goal of this work is to provide an empirically grounded assessment of quantum computing applications in bioinformatics, similarly to Khalaj et al. (2026)'s recent work. Differently, our review places specific emphasis on the analysis of experimental studies and real-device implementations, the comparison of different quantum hardware technologies, the presence and quality of benchmarks and classical baselines, and the evaluation of reproducibility through the availability of code and data. By critically evaluating hardware-scaling bottlenecks, classical benchmarking, dataset-size constraints, and reproducibility, our work addresses the empirical and infrastructural limitations that currently constrain quantum computing applications in bioinformatics.

This paper begins by defining four quantum computing paradigms and evaluating their applications across four core bioinformatics domains. It then compares the execution platforms used, analyzes empirical performance against classical baselines, and assesses the reproducibility of the literature. Finally, it synthesizes these findings to discuss challenges and outline recommendations for future research.

2 Quantum methods and scope

2.1 Quantum computing paradigms

Applications of quantum computing in bioinformatics do not rely on a single computational model. There are multiple paradigms with different mathematical formulations, hardware requirements, and levels of experimental maturity (Nałęcz-Charkiewicz et al., 2024). For the purposes of this review, we distinguish between four main categories:

• Quantum annealing;
• Gate-based quantum computing;
• Quantum-inspired methods;
• Classical simulations.

This distinction is essential for interpreting the available literature, since studies that are described as “quantum” may differ substantially in whether they are executed on real quantum processing units, simulated on classical hardware, or implemented entirely as classical algorithms inspired by quantum principles (McGeoch, 2014; Nielsen and Chuang, 2011) (Figure 1).

2.1.1 Quantum annealing

Quantum annealing is currently one of the most common paradigms in quantum bioinformatics, particularly because many biological problems can be formulated as combinatorial optimization tasks (Nałęcz-Charkiewicz et al., 2024). In this setting, the problem is typically encoded as either a Quadratic Unconstrained Binary Optimization (QUBO) problem or an Ising model (Heng et al., 2022). In a QUBO formulation, the objective is to find the binary vector that minimizes a quadratic cost function, usually written as: where x is a binary vector and Q is a symmetric matrix encoding both individual variable contributions and pairwise interactions. Constraints are not represented separately, but are added to the objective function through penalty terms (Rieffel et al., 2015). This makes QUBO formulations particularly suitable for problems such as genome assembly, codon optimization, RNA folding, feature selection, classification, and molecular docking, where biological choices can be represented as binary decisions.

minx∈{0,1}nxTQx

QUBO formulations are computationally equivalent to the Ising model (Sato et al., 2019). The corresponding Hamiltonian function can be defined as: where σ_*i*, σ_*j*∈{−1, +1} are spin variables, and h_*j* (local fields or biases) and J_*ij* (coupling strengths) are real-valued parameters. Here, H(σ) is the energy of the system (Li et al., 2018), which physically represents the cost function of the optimization problem mapped onto an energy landscape. The goal is to identify the spin configuration that minimizes this energy (the ground state of the Hamiltonian), which corresponds precisely to the optimal solution of the encoded biological problem.

H(σ)=∑i,jJijσiσj+∑jhjσj

The QUBO and Ising formulations are mathematically equivalent and are related via a simple linear change of variables from binary to spin variables. This transformation is defined by the equations: By substituting this relationship into the QUBO objective function, the mathematical terms can be expanded and rearranged to convert the QUBO matrix Q into the Ising parameters h and J. Researchers can formulate a bioinformatics problem in a binary-optimization language and map it to a physical representation compatible with quantum annealers (e.g., D-Wave's) or Ising-type solvers (McGeoch, 2014).

σi=2xi-1 or equivalently xi=σi+12.

However, the practical use of quantum annealing requires more than simply writing a QUBO or Ising function. The biological problem must first be translated into binary variables, objective terms, and penalties that correctly encode the desired solution. In addition, the resulting mathematical problem must be embedded within the hardware's physical topology (Pelofske, 2025). This step can be highly complex because current annealers are limited by qubit number, connectivity, noise, embedding overhead, and sensitivity to hyperparameters such as penalty weights and annealing schedules. For this reason, reported performance depends not only on the biological formulation but also on implementation details that affect reproducibility and scalability.

2.1.2 Gate-based quantum computing

Gate-based quantum computing is a more general-purpose and theoretically universal model of quantum computation (Pyrkov et al., 2024). In this paradigm, information is encoded into qubits and manipulated through sequences of quantum gates arranged in circuits. For example, Figure 2 shows a circuit for Grover's celebrated quantum search algorithm (Grover, 1997). In bioinformatics, gate-based approaches have been proposed or explored for sequence search, approximate pattern matching, molecular similarity, network alignment, protein folding, graph kernels, and biomedical classification.

Compared with quantum annealing, gate-based quantum computing allows a broader range of algorithmic formulations. A problem does not necessarily need to be reduced to a QUBO or Ising model; instead, it can be expressed through data encoding procedures, unitary transformations, parameterized circuits, measurements, and hybrid quantum-classical optimization loops. This flexibility makes gate-based approaches attractive for domains such as quantum machine learning, molecular simulation, graph analysis, variational quantum eigensolvers, and quantum approximate optimization algorithms. Some of these methods are described in more detail in Section 3.

Still, real-device implementations remain constrained by the limitations of NISQ hardware (Preskill, 2018), including limited qubit counts, circuit depth restrictions, gate errors, decoherence, noise, and the absence of fully scalable quantum error correction. Limited qubit connectivity is an additional constraint for gate-based processors that lack native all-to-all interactions. As a result, many gate-based bioinformatics studies are still performed on simulators or are limited to very small problem instances.

2.1.3 Quantum-inspired algorithms

A third category comprises quantum-inspired methods. These approaches borrow concepts from quantum computing but are executed entirely on classical hardware. They can be useful and, in some cases, competitive with conventional classical algorithms. However, they do not provide direct evidence of quantum advantage because they do not exploit physical quantum resources, such as superposition, entanglement, or tunneling, on a quantum processing unit.

2.1.4 Classical simulation

Quantum emulators occupy an intermediate position in the current research landscape. They are classical software environments that reproduce the behavior of quantum circuits or quantum algorithms without using a physical quantum processor. They are useful for testing the logical correctness of an algorithm, estimating resource requirements, comparing QUBO or circuit designs, and validating small-scale examples before execution on quantum hardware. Some simulators, such as Qiskit (Javadi-Abhari et al., 2024), also support the inclusion of noise models, which approximate certain hardware imperfections. Regardless, simulation results should not be interpreted as equivalent to real quantum hardware execution, as classical simulation may not fully capture the complexity of device-specific constraints.

This distinction among paradigms has direct consequences for the empirical assessment of quantum computing in bioinformatics. Studies based on real quantum annealers provide a different type of evidence from studies based on gate-based quantum processors, classical simulators, or quantum-inspired algorithms. Similarly, results obtained on toy-scale instances cannot be directly generalized to biologically realistic datasets without careful consideration of scalability, benchmarking, and reproducibility. Therefore, in this review, the computational paradigm, problem formulation, and execution platform are treated as central criteria for evaluating each study.

2.2 Scope and methodology

This review focuses on experimental applications of quantum computing in bioinformatics, distinguishing between theoretical proposals, simulated validations, and implementations on real quantum hardware or quantum-inspired platforms. Unlike many previous reviews, this work adopts a more restrictive selection criterion, prioritizing studies that report experimental results, computational benchmarks, comparisons with classical methods, or sufficient details about the execution platform used.

The literature search included studies that applied quantum computing, quantum annealing, or quantum-inspired methods to bioinformatics and biomedical problems. The included topics cover sequence analysis, genome assembly, RNA folding, protein structure prediction, molecular simulation, drug discovery, phylogenetics, network analysis, and biomedical classification. Articles were evaluated not only by the biological domain addressed, but also by the computational formulation proposed and the type of evidence reported. Specifically, in this review, we did not consider classical HPC approaches to bioinformatics problems. The interested reader can find more information in, for example, Liu et al. (2016); Nobile et al. (2017); Zou et al. (2021); Jiang et al. (2021) and Liu et al. (2024).

To collect the studies for this systematic review, a comprehensive literature search was conducted using the Scopus search engine. The search strategy employed a combination of domain-specific keywords (e.g., “bioinformatics,” “biology,” and “medicine”) and technology-focused terms (e.g., “quantum,” “annealing,” “IBM,” and “D-Wave”) to identify articles on quantum computing applications in bioinformatics and closely related domains. The initial search results were filtered to exclude preprints, papers published before 2014, and non-peer-reviewed journals or conferences. Publications with zero recorded citations on Scopus were also excluded, except for those issued in more recent times (i.e., after January 2025). We increased this dataset by manually screening the reference lists of existing literature (Cordier et al., 2022; Mokhtari et al., 2024; Ghamsari, 2025; Sung et al., 2026; Baiardi et al., 2023; Nałęcz-Charkiewicz et al., 2024; Pal et al., 2024b) to identify additional studies. Finally, to align with the empirical scope of this review, we excluded purely theoretical proposals; we retained only publications reporting practical experiments conducted on classical simulators, quantum-inspired platforms, or real quantum hardware.

Tables 1–4 summarize the final dataset investigated for this review. The same dataset is also available in Supplementary Table 1. Figure 3 illustrates the search strategy based on the PRISMA 2020 model (Page et al., 2021a,b).

Table 1. Summary of reviewed literature (Group 1: Genome assembly, sequence alignment, and pattern matching).

Each row represents an evaluated study detailing the authors, topic, quantum paradigm (QA, Quantum annealing; GB, Gate-based; QI, Quantum-inspired; Sim, Simulation), execution platform, whether it was compared to a classical baseline, dataset size (RL, Similar to real-life cases; Red, Toy or reduced dataset), and whether the source code and test data are publicly available.

To assess the completeness of the Scopus search and reduce potential database selection bias, additional queries were run in PubMed, Web of Science, IEEE Xplore, and the ACM Digital Library using equivalent combinations of search parameters. The results showed substantial overlap with the records already retrieved through Scopus, and no additional relevant studies were identified. A detailed comparison of database coverage is provided in Supplementary Table 2.

For each study, we analyzed three main aspects, summarized in Figure 4. The first is the presence and quality of comparisons with classical methods. In particular, we assessed whether relevant classical baselines were used, whether the evaluation metrics were clearly defined, and whether the reported results supported a genuine improvement, comparable performance, or only a proof-of-concept demonstration. This aspect is particularly important because, in the absence of robust classical baselines, the performance of quantum or quantum-inspired methods remains difficult to interpret.

The second aspect is the execution modality. Studies were classified into implementations on real quantum processing units, quantum annealers, emulators, hybrid quantum-classical workflows, and quantum-inspired methods. When available, the specific platform was also recorded. This classification helps avoid ambiguity between studies actually executed on quantum hardware, simulated studies, and classical algorithms inspired by quantum computing.

The third aspect is reproducibility. For each study, we evaluated the availability of source code, datasets, execution parameters, hardware or simulator descriptions, and sufficient documentation to reproduce the reported results. This criterion is central because the performance of quantum and quantum-inspired algorithms can depend strongly on implementation details, such as the choice of penalty terms in QUBO formulations, embedding strategies, circuit depth, noise models, and solver settings. The lack of code and data limits independent verification and makes it more difficult to assess the field's actual progress.

Overall, our methodology allows studies to be classified not only by their bioinformatics application domain but also by their degree of empirical maturity. The aim is not to determine whether quantum computing has already achieved a generalizable advantage in bioinformatics, but rather to critically assess which results have actually been demonstrated on these platforms and at what level of reproducibility. Furthermore, we emphasize that our analysis exclusively examined peer-reviewed articles and did not consider preprints. Our aim was not to peer-review the examined works again to assess the quality of the studies. This has already been done by reviewers before publication. Rather, we defined a set of objective criteria (reproducibility, platform, dataset availability, and dataset size) to formally analyze the identified articles.

2.2.1 Methodological limitations

We acknowledge the absence of a formal risk-of-bias or standardized study-quality assessment tool as a methodological limitation of this review. Conventional risk-of-bias frameworks (such as Whiting et al., 2016 or Page et al., 2023) cannot be applied directly to computational benchmarking and evaluation of emerging quantum hardware. While our selection criteria relied on the peer-review process as a baseline quality filter and assessed each study across strict, objective technical criteria, the lack of a formalized quality-scoring tool remains a limitation.

3 Quantum computing applications across biological domains

This literature review shows that quantum computing and quantum-inspired paradigms are being examined across a wide variety of applications. To evaluate these advancements, we categorized the selected studies into four broad thematic domains based on their underlying biological objectives. These domains enclose sequence-level genomics, three-dimensional structural biology, evolutionary systems, and high-dimensional biomedical data classification. A visual summary of these four areas is presented in Figure 5.

3.1 Genome assembly, sequence alignment, and pattern matching

Sequence analysis is among the earliest and most extensively explored domains in bioinformatics (Castrignanò et al., 2006b; Picardi et al., 2011; Di Martino et al., 2026; Chiara et al., 2018). This is because of its many core tasks: exact pattern matching, approximate string search, sequence alignment, motif discovery, and sequence similarity scoring can all be reformulated as search or combinatorial optimization problems. This natural algorithmic affinity has made the field particularly attractive for quantum approaches. Genome and transcriptome assembly (Paris et al., 2025; Chiocchio et al., 2022; Libro et al., 2023), sequence alignment, and pattern matching all share a common computational profile: they involve large-scale combinatorial search over discrete biological sequences, with exact solutions that are either NP-hard or computationally prohibitive at the scales demanded by modern genomics. Quantum computing offers two principal mechanisms for addressing these bottlenecks: Grover-style amplitude amplification, which provides quadratic speedups for unstructured search problems, and quantum annealing or variational methods, which reformulate combinatorial optimization tasks as energy minimization over Ising or quadratic unconstrained binary optimization (QUBO) Hamiltonians. Despite this strong formal compatibility, the literature remains highly fragmented across different levels of technological maturity, ranging from purely theoretical proposals to simulation-based proofs of concept and only a limited number of experimentally grounded implementations.

3.1.1 Genome assembly

De novo genome assembly naturally maps to NP-hard graph traversal problems such as finding Hamiltonian paths through overlap graphs or resolving ambiguities in de Bruijn graphs caused by repeats, uneven coverage, and sequencing errors. For large eukaryotic genomes, classical assembly heuristics remain computationally demanding and cannot guarantee optimal solutions. Quantum annealing provides a direct way to solve these problems by encoding the assembly as a QUBO. Boev et al. (2021) demonstrated this by formulating genome assembly as a QUBO problem and solving it on D-Wave quantum annealers. Nałęcz-Charkiewicz and Nowak (2022) extended this line of research using a Genomic Signal Processing approach, detecting overlaps between reads via Pearson correlation coefficients and formulating the assembly as a traveling salesman problem-like (TSP-like) optimization task solved through a hybrid classical-quantum workflow on D-Wave hardware. Both studies highlight that current quantum hardware can only handle toy-scale instances. Still, they provide the clearest empirical evidence that the assembly-to-QUBO mapping is, in principle, feasible.

3.1.2 Sequence alignment and pattern matching

Multiple sequence alignment (MSA) is NP-hard in its exact formulation. Practical tools such as Clustal Omega (Sievers et al., 2011), MUSCLE (Edgar, 2022), and MAFFT (Katoh et al., 2009) rely on progressive heuristics that sacrifice optimality for tractability. Pattern matching at the genomic scale, whether exact or approximate, faces analogous scaling challenges when millions of short queries must be searched against databases containing billions of nucleotides (Chakraborty et al., 2023; Aramyan et al., 2022; Castrignanò et al., 2004, 2006a). Given a set of n elements, Grover's algorithm (Grover, 1997), Figure 2 offers a quadratic speedup for unstructured search, reducing the number of queries from O(n) to O(n), which has motivated several theoretical proposals. Early work was predominantly theoretical, focusing on reformulating these classical problems in terms of quantum algorithms. Prousalis and Konofaos (2018) recast sequence alignment as a quantum multi-pattern recognition problem combining a dot-matrix strategy with Grover-like search, presenting an algorithmic proof of concept without hardware implementation or experimental validation. Soni and Rasool (2021) proposed two quantum exact pattern-matching algorithms based on Grover-style search with both quantum-accessible classical memory and quantum random access memory-based (QRAM) access models, but they were not executed on real quantum hardware. Their subsequent extension, Soni and Rasool (2022), generalized this to multiple-pattern matching on a multicore quantum-processing abstraction, with improved methodological transparency, including publicly accessible simulation code. A similar theoretical orientation characterizes work on sequence-based similarity scoring: Chagneau et al. (2024) proposed quantum routines for computing protein similarity scores through quantum sequence generation and comparison, framing the task as a sequence-based similarity search for protein prioritization. Still, the study remained primarily methodological with simulated evaluation. Yang et al. (2018) proposed a quantum algorithm for molecular similarity based on quantum phase estimation using labeled molecular graphs, remaining entirely theoretical. Cao et al. (2015)) formulated DNA motif discovery as an adiabatic optimization problem.

3.1.3 Quantum-inspired heuristics

A distinct branch of the literature is represented by quantum-inspired methods, in which most practically benchmarked approaches remain classical in execution despite their quantum-inspired design principles. This distinction is important: quantum-inspired algorithms borrow representational or operational concepts from quantum mechanics, such as superposition-like probability amplitudes, quantum gate-inspired operators, or Ising-model formulations, but execute entirely on classical hardware, and their relationship to genuine quantum advantage is indirect. Kartous et al. (2014) introduced a quantum-inspired cuckoo search algorithm for MSA, combining a randomized progressive alignment strategy with quantum-inspired evolutionary operators, including qubit-based representation, measurement, interference, and mutation. Their method was empirically benchmarked against established classical MSA tools and showed competitive alignment quality, but was implemented entirely on classical hardware. Similarly, Giannakis et al. (2019) proposed a quantum-inspired general variable neighborhood search heuristic for MSA, executed classically and showing promising behavior on larger sequence sets. Together, these studies illustrate a recurring terminological ambiguity in the literature: many so-called quantum methods in bioinformatics are actually quantum-inspired heuristics on conventional hardware. For the double digest problem, Suo et al. (2020) proposed a quantum-inspired genetic algorithm, with simulation results showing modest acceleration over classical genetic algorithms. Guo et al. (2017) addressed DNA code design using the Bloch quantum chaos algorithm, another quantum-inspired optimization approach.

3.1.4 Implementations on quantum hardware

More practically grounded work on real quantum hardware emerged only recently. Fox et al. (2021) reformulated mRNA codon optimization as a QUBO problem and implemented it on the D-Wave quantum annealer, directly comparing it against a classical genetic algorithm under the same objective function. This study is one of the clearest experimentally grounded examples in sequence-related quantum bioinformatics, combining real hardware execution, explicit classical baselines, a biologically realistic optimization task, and detailed methodological disclosure. A similarly important step toward implementation-aware sequence analysis was taken by Sarkar et al. (2021a) as they introduced QiBAM, a gate-based quantum algorithm for approximate substring search applied to DNA read alignment. Unlike purely asymptotic Grover-based proposals, QiBAM addressed the biologically motivated problem of approximate read alignment in the presence of sequencing errors, extending Grover's search algorithm. The method was validated using the OpenQL compiler and the QX Simulator framework, moving beyond abstract oracle formulations toward compiler-executable, simulator-tested circuit-level design. However, it remains limited to simulation and small test instances.

Taken together, these studies show that quantum approaches to sequence analysis are conceptually mature but experimentally immature. Most published work remains dominated by theoretical speedup claims, idealized oracle models, or quantum-inspired heuristics executed classically. Only a small subset of studies move toward implementation-aware, comparatively reproducible evaluation. Sequence analysis represents one of the clearest examples of the current state of quantum bioinformatics: a domain with strong formal compatibility with quantum computing, but where practical quantum advantage remains largely prospective rather than experimentally demonstrated.

3.2 Protein Structure, molecular simulations, and drug discovery

The computational characterization of protein and RNA structure and function covers several key areas. These include folding, binding-site identification, molecular docking, conformational dynamics, and molecular simulation. This work stands at the intersection of a mature classical biophysical tradition and an emerging quantum computational paradigm, promising to resolve the fundamental limitations imposed by the exponential complexity of many-body quantum-mechanical systems (Cordier et al., 2022; Outeiral et al., 2021b).

3.2.1 RNA structure and codon optimization

RNA secondary structure prediction is NP-complete when pseudoknots are included (Fox et al., 2022). Pseudoknots are biologically important motifs appearing in ribosomal RNA, telomerase RNA, and viral frameshifting elements (Staple and Butcher, 2005). Yet, their prediction remains computationally intractable for classical methods at biologically relevant scales. A closely related combinatorial challenge is mRNA codon optimization (Fox et al., 2021): because the genetic code is degenerate, reverse-translating a desired protein sequence into an mRNA construct that maximizes expression requires selecting among up to six synonymous codons per amino acid position while simultaneously optimizing for codon usage bias, secondary structure stability, GC-content, and avoidance of regulatory motifs. This is an NP-hard problem with direct practical relevance for mRNA vaccine design, gene therapies, and recombinant protein production.

Fox et al. (2022) formulated RNA secondary structure prediction as a binary quadratic model designed to maximize consecutive base pairs and average stem length, solving it on a D-Wave quantum annealer. The authors achieved highly competitive performance compared to replica-exchange Monte Carlo and three other literature algorithms on known structures with pseudoknots. The same group (Fox et al., 2021) applied quantum annealing to mRNA codon optimization on D-Wave, directly comparing against a classical genetic algorithm under the same objective function.

3.2.2 Protein structure prediction and design

Protein structure prediction and design present even tougher combinatorial challenges. Despite the high accuracy of deep learning methods such as AlphaFold (Abramson et al., 2024) and RoseTTAFold (Baek et al., 2021) for well-ordered globular domains, significant limitations remain for intrinsically disordered regions, flexible loops, protein-protein interfaces (Castrignanò et al., 2007), and the prediction of binding-site geometries with the sub-angstrom accuracy required for drug design. Quantum algorithms offer an alternative route through variational quantum eigensolvers (VQE) (Cerezo et al., 2021) and quantum approximate optimization algorithms (QAOA) (Pellow-Jarman et al., 2024), which can encode the folding free-energy landscape as a parameterized quantum circuit optimized through hybrid quantum-classical loops. Zhang Y. et al. (2026) presented a VQE-based framework that represents amino acid connectivity on a tetrahedral lattice with steric, geometric, and chirality constraints encoded into a problem-specific Hamiltonian expressed as sparse Pauli operators. Evaluated on real protein fragments, the quantum framework outperforms standard models, demonstrating the practical utility on real gate-based hardware. A similar work by Uttarkar and Niranjan (2026) showed that error-mitigated VQE can resolve the folding free-energy landscape of disordered segments of ubiquitin C on both tensor-network simulators and trapped-ion processors. Quantum genetic algorithms and Hamiltonian-embedding lattice encoding have established proof-of-principle speedups on simplified hydrophobic-polar protein folding models (Atari and Majd, 2022; Outeiral et al., 2021a; Wong and Chang, 2021).

Protein design requires exploring combinatorial spaces of rotameric states that grow exponentially with protein size. Maguire et al. (2021) addressed the qubit bottleneck by developing XENet, a graph-convolutional architecture explicitly co-designed for quantum hardware that generates molecular graph representations, decreasing the size of the solution space by a factor of 165, and illustrating how classical machine learning can serve as a preprocessing stage to make quantum protein design tractable on near-term devices.

3.2.3 Molecular dynamics and simulations

Scaling quantum simulation to biologically realistic protein sizes remains a major challenge (Castrignanò et al., 2000; Madeddu et al., 2022; Botta et al., 2026). Classical molecular dynamics (MD) approximates quantum effects using empirical force fields that model atoms as classical particles interacting through simplified potential energy terms, but these parameterizations inadequately capture electronic polarization, charge transfer, and bond-breaking processes. Quantum-chemical methods can provide the necessary accuracy but scale polynomially to exponentially with system size on classical computers, limiting calculations to fragments containing at most a few hundred atoms. Berryman et al. (2022) demonstrated a hybrid approach in which classical forcefields are supplemented with a kernel-based machine learning method trained on quantum-mechanical fragment energies to correct the potential energy surface for DNA base pair stretching in explicit solvent, providing a candidate explanation for the long-standing discrepancy between single-molecule stretching experiments and classical simulations. Their correction method was integrated into the AMBER22 (Case et al., 2025) simulation software, suggesting a generalizable strategy for embedding quantum accuracy into classical MD workflows at a tractable computational cost. At the single-molecule level, Taniguchi et al. (2023) demonstrated nucleotide identification using a quantum gate circuit designed to explain the single-molecule conductance of adenosine electronically bonded between electrodes.

3.2.4 Molecular docking and drug discovery

In molecular docking and structure-based drug discovery, hybrid quantum-classical workflows are gaining traction. The combinatorial search space of molecular docking can be cast as a QUBO problem tractable on near-term quantum hardware. Aramyan et al. (2022) deployed QAOA-based MaxCut formulations combined with zero-differential overlap molecular pruning to characterize SP-A interactions with the SARS-CoV-2 spike protein on a NISQ device, and compared against state-of-the-art classical MaxCut algorithms. Chagneau et al. (2024) proposed quantum algorithms for computing protein similarity in drug design applications via quantum sequence generation and comparison, while Bhuvaneswari et al. (2023) explored Grover's algorithm for protein-protein interaction prediction.

Taken together, these studies demonstrate that quantum approaches to protein structure, molecular simulation, and drug discovery span the full range from theoretical proposals to hardware demonstrations on real quantum processors. The most advanced empirical results establish that quantum methods can already produce biologically meaningful results for carefully scoped problems. However, all current demonstrations remain constrained to small molecular systems or fragments.

3.3 Phylogenetics and genomic information theory

Phylogenetics and genomic information theory share a deep computational structure: both require exhaustive exploration of combinatorial spaces that grow super-exponentially with input size, and both mostly resist the heuristic shortcuts that have proven effective in other areas of bioinformatics. Phylogenetic tree inference aims to reconstruct the evolutionary relationships among biological entities from molecular sequence data. At the same time, algorithmic information theory provides a framework for quantifying the intrinsic complexity and structure of biological sequences without reference to an assumed generating distribution. Despite their distinct biological motivations, these problems converge on a common computational bottleneck: the need to evaluate an incredibly large space of discrete combinatorial objects, for which no efficient classical shortcut is known to guarantee optimal solutions.

3.3.1 Phylogenetic tree inference

Phylogenetic tree construction is central to comparative genomics, molecular evolution, epidemiology, and functional annotation. Distance-based methods such as UPGMA (Sokal and Sneath, 1963), Neighbor-Joining (Saitou and Nei, 1987), and Fitch-Margoliash (Fitch and Margoliash, 1967) compute pairwise evolutionary distances from sequence alignments and apply clustering algorithms to infer tree topology and branch lengths. Character-based and statistical methods, like maximum parsimony, maximum likelihood, and Bayesian inference, evaluate the probability or parsimony score of observed sequence data under explicit models of sequence evolution and search for the tree that optimizes the chosen criterion. The fundamental computational difficulty is that the number of possible unrooted binary tree topologies grows super-exponentially with the number of taxa. Even heuristic tree search strategies such as nearest-neighbor interchange, subtree pruning and regrafting, and tree bisection and reconnection can require days to weeks of computation for datasets with thousands of taxa, which are now routine in metagenomics and viral genomics (Santos et al., 2022; Onodera et al., 2023). Maximum likelihood and Bayesian methods increase this difficulty by requiring repeated evaluation of the likelihood function at each step of the tree search. For large-scale phylogenomic analyses involving hundreds of gene families across dozens of genomes, the aggregate computational cost becomes a significant bottleneck even on high-performance computing clusters.

Beyond computational complexity, phylogenetic analysis increasingly faces challenges related to data privacy and security. Genomic sequences used for phylogenetic reconstruction may contain sensitive personal information in human genomics, proprietary data in agricultural biotechnology and pharmaceutical research, or biosecurity-relevant information in pathogen surveillance. Collaborative phylogenetic analyses across institutions or national borders may require that individual sequences remain confidential while still contributing to a shared tree, creating a need for secure multiparty computation protocols that preserve privacy without sacrificing analytical accuracy (Santos et al., 2022). Classical secure computation frameworks exist, but add substantial computational overhead, and their security guarantees may be undermined by the advent of large-scale quantum computers capable of breaking current public-key cryptographic schemes.

Santos et al. (2022) addressed both the phylogenetic inference and privacy challenges by proposing a secure multiparty computation system assisted with quantum cryptographic protocols. Their system adapts three distance-based phylogenetic methods to a private setting where each party's sequences are not disclosed to other participants. The security infrastructure relies on Quantum Key Distribution (Scarani et al., 2009) and Quantum Oblivious Key Distribution (Panduranga Rao and Jakobi, 2013), which provide security guarantees robust against quantum-computational attacks. Their results demonstrate that quantum-enhanced cryptographic security can be integrated into phylogenetic workflows without significant computational penalty, an important consideration as genomic data sharing becomes increasingly regulated and as quantum threats to classical cryptography grow more concrete.

Onodera et al. (2023) took a complementary approach by applying Fujitsu's quantum-inspired digital annealer to phylogenetic tree construction via normalized minimum cut. The digital annealer, while not a true quantum device, implements a fully connected Ising model solver inspired by quantum annealing principles and can explore combinatorial landscapes more efficiently than conventional simulated annealing for certain problem classes. In this formulation, the phylogenetic problem is cast as a graph partitioning task: sequences are represented as nodes in a similarity graph, and the tree is constructed by recursively bisecting the graph to minimize a normalized cut criterion. Onodera et al. provided detailed data on both solution quality and processing time across datasets of varying sizes and divergences, establishing empirical baselines for the feasibility of quantum-inspired hardware for phylogenetic inference. While the method does not yet compete with highly optimized classical phylogenetic software for standard use cases, it demonstrates that the graph-partitioning formulation of phylogenetics can be translated to Ising-model solvers and could benefit from future improvements in quantum annealing hardware.

3.3.2 Algorithmic information theory

Algorithmic information theory, rooted in the concept of Kolmogorov complexity (Li and Vitányi, 2008), provides a fundamentally different lens for analyzing biological sequences. Rather than comparing sequences through alignment-based similarity measures, algorithmic complexity quantifies the information content of a sequence by the length of the shortest program that generates it on a universal Turing machine. This perspective connects to deep questions about the nature of biological information: highly repetitive genomic regions have low algorithmic complexity, while apparently random sequences have high complexity, reflecting the role of mutational processes, selection, and functional constraint. Practical applications include alignment-free phylogenetic distance estimation, detection of horizontal gene transfer, characterization of genomic complexity across species, and design principles for synthetic genomes (Sarkar et al., 2021b). The fundamental computational obstacle is that Kolmogorov complexity is uncomputable in the general case, and practical approximations require exhaustive enumeration of all programs up to a given length on a reference universal Turing machine.

Sarkar et al. (2021b) developed a quantum computing framework based on the circuit model that directly addresses this computational bottleneck. Their approach obtains the universal prior distribution over the space of automata as a quantum superposition, which is then conditioned to estimate algorithmic information metrics, including algorithmic probability and coding-theorem-based complexity. The authors identified specific cases in which the quantum implementation offers a polynomial advantage over exhaustive classical enumeration. Algorithmic information estimation is one of the more promising candidates for demonstrating genuine quantum advantage via Grover-style search in a setting where classical shortcuts are provably unavailable. While the current implementation remains limited by the available quantum hardware to small automata sizes, the theoretical analysis establishes a clear path to quantum advantage as device capabilities scale.

3.4 Quantum machine learning and classification for biomedical data

Modern biomedical research generates high-dimensional data from diverse experimental and clinical sources (Bolis et al., 2017), such as whole-genome sequencing, transcriptomics, proteomics, metabolomics, medical imaging, and electronic health records. Extracting biologically or clinically meaningful patterns from these data has become a central computational challenge in contemporary biology and medicine. Classification tasks can be applied to virtually every biomedical domain: identifying tumors into molecular subtypes from multi-omics profiles to guide treatment selection, predicting transcription factor binding specificity from DNA sequence features to understand gene regulation, diagnosing diseases from clinical measurements, and annotating proteins into functional families from sequence or structural descriptors (Li et al., 2018, 2021). Beyond sample-level classification, many biomedical problems involve structured data naturally represented as graphs, like protein-protein interaction networks, gene regulatory networks, metabolic networks, and brain connectivity maps. Measuring similarity between graphs or aligning nodes across networks is essential for comparative analysis and functional annotation (Kishi et al., 2022; Bai et al., 2015; Daskin et al., 2014).

3.4.1 Classification tasks

The computational bottlenecks in biomedical classification are varied and persistent. The burden of dimensionality is heavy: in genomics, datasets routinely contain expression measurements for tens of thousands of genes, but only tens to hundreds of patient samples, creating a regime where classical machine learning models are highly prone to overfitting and where reliable generalization requires careful regularization, feature selection, or dimensionality reduction (Li et al., 2018, 2021). Training complex models on such high-dimensional, small-sample data demands substantial computational resources, and the combinatorial explosion of possible feature interactions makes exhaustive model search infeasible. For graph-structured biological data, the challenges are even more severe. Computing graph kernels that capture meaningful topological features requires enumerating subgraphs or performing random walks, both of which scale poorly with graph size; kernels based on all-subgraph enumeration are exponentially expensive on classical hardware (Kishi et al., 2022; Bai et al., 2015). Network alignment across multiple biological networks involves computing the principal eigenvector of the Kronecker product of input adjacency matrices. This space quickly grows, exceeding classical memory and computation limits (Daskin et al., 2014).

Quantum computing offers several distinct mechanisms for addressing these bottlenecks. Quantum annealing can train machine learning models by formulating parameter optimization as a QUBO problem, potentially navigating complex energy landscapes more effectively than classical optimizers. Gate-based quantum circuits can encode and process exponentially large feature spaces through quantum superposition, enabling kernel evaluations and feature mappings that would be intractable classically. Quantum phase estimation (Yang et al., 2018) can extract spectral properties of large matrices exponentially faster than classical eigensolvers, directly benefiting network analysis tasks. Several empirical studies suggest that quantum and quantum-inspired approaches may exhibit particular advantages in small samples. The high-dimensional regime that characterizes much of biomedical data is precisely the setting where classical methods struggle most with overfitting (Li et al., 2021, 2018; Willsch et al., 2020).

3.4.2 Molecular and multi-omics classification

For molecular-level classification, Li et al. (2018) trained a D-Wave quantum annealer to classify and rank transcription factor binding specificity using simplified datasets derived from actual binding affinity experiments. Comparing against five classical methods, they observed a modest advantage in classification performance and comparable ranking performance for the quantum annealer on small training sets, suggesting that quantum annealing may be particularly suited to machine learning problems where training data are limited. Li et al. (2021) developed Ising-type algorithms inspired by quantum processor architectures for the classification of integrated, genome-wide multi-omics human cancer data from The Cancer Genome Atlas, representing the first application of such approaches to this data type. Their analysis revealed that Ising-type algorithms performed better with small datasets and exhibited less overfitting than classical alternatives. Willsch et al. (2020) trained support vector machines on a D-Wave quantum annealer using both synthetic and biological data, discovering that the quantum annealer produces an ensemble of diverse near-optimal solutions that often generalize better to unseen data than the single global minimum found by a conventionally trained support vector machine (SVM). This ensemble property may represent a genuinely novel mechanism for improving generalization in small-sample settings, distinct from classical ensemble methods such as bagging or boosting.

3.4.3 Clinical disease prediction and clustering

For clinical classification, Ullah et al. (2022) proposed a fully connected quantum convolutional neural network for the diagnosis of ischemic heart disease. Their architecture exploits the logarithmic depth scaling of quantum circuits to reduce the number of trainable parameters while maintaining expressive power, and was evaluated after classical preprocessing steps, including feature selection via maximum relevance minimum redundancy filtering and dimensionality reduction using SVM with recursive feature elimination. The model achieved notable improvements in test accuracy over previously published quantum and classical baselines, demonstrating that quantum circuit architectures can be competitive for structured clinical classification tasks even on current NISQ devices. Bharill et al. (2019) proposed an enhanced quantum-inspired evolutionary fuzzy c-means algorithm for unsupervised learning. By combining quantum computing concepts with the fuzzy c-means framework, their method simultaneously optimizes the fuzzifier parameter, the number of clusters, and the initial cluster centers. This approach demonstrated improved clustering accuracy and convergence behavior compared to classical fuzzy clustering and standard evolutionary alternatives.

3.4.4 Graph-structured data and networks

For graph-structured biological data, quantum approaches offer perhaps their most distinctive theoretical advantages. Kishi et al. (2022) proposed quantum graph kernels that exploit quantum superposition to encode features extracted from all subgraphs of an input graph into a single quantum state, enabling the computation of graph similarity based on the complete subgraph spectrum. This task is expensive on classical hardware because the number of subgraphs grows exponentially with graph size. Through detailed numerical simulation on bioinformatics classification problems, they demonstrated that the proposed quantum kernel frequently achieves better classification accuracy than established classical graph kernels, and proved that the quantum computer requires fewer queries to construct the feature vector than a classical sampler approximating the same vector. Bai et al. (2015) developed a complementary approach based on continuous-time quantum walks, using the quantum Jensen–Shannon divergence to measure information-theoretic dissimilarity between graphs represented as density matrices derived from the quantum walk dynamics. Their evaluation on standard graph datasets from bioinformatics and computer vision demonstrated competitive classification performance, establishing that quantum walk-based graph descriptors capture structural information complementary to classical graph invariants.

4 Execution platforms and technologies

The reviewed literature can be categorized into the four distinct quantum paradigms described in Section 2: quantum annealing, gate-based quantum computing, quantum-inspired algorithms, and simulation-based approaches (i.e., classical emulation of quantum applications). Each paradigm offers distinct computational trade-offs, and their adoption across the field of quantum bioinformatics is uneven.

Our findings are based on a curated collection of 53 experimental studies from the literature. A summary of these publications and their key elements is presented in Tables 1–4.

Table 2. Summary of reviewed literature (Group 2: Protein structure, molecular simulations, and drug discovery).

Each row represents an evaluated study detailing the authors, topic, and quantum paradigm (QA, Quantum annealing; GB, Gate-based; QI, Quantum-inspired; Sim, Simulation), execution platform, whether it was compared to a classical baseline, dataset size (RL, Similar to real-life cases; Red, Toy or reduced dataset), and whether the source code and test data are publicly available.

Table 3. Summary of reviewed literature (Group 3: Phylogenetics and genomic information theory).

Each row represents an evaluated study detailing the authors, topic, and quantum paradigm (QA, Quantum annealing; GB, Gate-based; QI, Quantum-inspired; Sim, Simulation), execution platform, whether it was compared to a classical baseline, dataset size (RL, Similar to real-life cases; Red, Toy or reduced dataset), and whether the source code and test data are publicly available.

Table 4. Summary of reviewed literature (Group 4: Quantum machine learning and classification for biomedical data).

Each row represents an evaluated study detailing the authors, topic, and quantum paradigm (QA, Quantum annealing; GB, Gate-based; QI, Quantum-inspired; Sim, Simulation), execution platform, whether it was compared to a classical baseline, dataset size (RL, Similar to real-life cases; Red, Toy or reduced dataset), and whether the source code and test data are publicly available.

Quantum-inspired solutions and classical simulations are both executed entirely on classical hardware, distinguishing them from the other two paradigms that rely on physical quantum processors. The distinctions between these two approaches can sometimes be unclear. Some techniques labeled “quantum-inspired” may borrow principles from quantum computing while overlapping in methodology and motivation with true circuit-simulation studies. This ambiguity means that, in certain contexts, it may be practical to consider these two approaches together, as they both rely on classical computational techniques to solve problems that might otherwise benefit from quantum mechanics.

Based on our analysis, classical simulations of quantum circuits represent the dominant paradigm (25 studies, 47%), followed by quantum annealing (11 studies, 21%), quantum-inspired algorithms (11 studies, 21%), and gate-based quantum computing (6 studies, 11%). This distribution is illustrated in Figure 6.

4.1 Quantum annealing vs. gate-based quantum computing

Among studies utilizing real quantum hardware, quantum annealers, and D-Wave systems in particular, dominate the landscape. The eleven quantum annealing studies employed D-Wave processors spanning three hardware generations: the D-Wave 2X (Calude et al., 2020; Li et al., 2018), the D-Wave 2000Q (Boev et al., 2021; Nałęcz-Charkiewicz and Nowak, 2022; Li et al., 2021; Willsch et al., 2020), and the D-Wave Advantage (1st gen.) (Fox et al., 2021, 2022; Maguire et al., 2021; Zaborniak et al., 2022; Fernández-Otero et al., 2026). By contrast, only six studies employed gate-based quantum processors, distributed across IBM hardware (Taniguchi et al., 2023; Zhang Y. et al., 2026), IonQ trapped-ion systems (Uttarkar and Niranjan, 2026; Uttarkar et al., 2026), Rigetti superconducting processors (Aramyan et al., 2022), and QuantumCTek Xiaohong (You et al., 2025).

Several factors may explain D-Wave's predominance in quantum bioinformatics, as illustrated in Section 2.1.1. First, many core bioinformatics problems (genome assembly, RNA folding, codon optimization, protein design, phylogenetic tree construction, and machine learning model training) can be naturally reformulated as combinatorial optimization problems that can be written as QUBO or Ising Hamiltonians (Nałęcz-Charkiewicz et al., 2024; Cordier et al., 2022). Quantum annealers are purpose-built for this exact class of problems, requiring no gate decomposition, circuit compilation, or error correction overhead.

Second, D-Wave systems offer significantly higher qubit counts than current gate-based processors. The D-Wave Advantage2^TM provides over 4,400 qubits, while the largest gate-based devices in the reviewed studies offer only hundreds. This difference in raw qubit availability allows quantum annealers to encode larger problem instances, even though the effective problem size remains constrained by connectivity and embedding overhead.

Third, the D-Wave programming model is comparatively accessible: users formulate their problem as a QUBO matrix and submit it to the cloud-accessible processor, without needing expertise in quantum circuit design, gate synthesis, or error mitigation strategies (Baiardi et al., 2023).

Gate-based quantum computers, by contrast, offer universal programmability and can in principle implement any quantum algorithm, including those with no natural QUBO formulation (Cordier et al., 2022). The studies employing gate-based hardware addressed problems that exploit this universality: VQE-based protein structure prediction (Zhang Y. et al., 2026), quantum circuit-based nucleotide identification (Taniguchi et al., 2023), QAOA-based molecular pruning for protein docking (Aramyan et al., 2022), and error-mitigated VQE for disordered protein folding (Uttarkar and Niranjan, 2026). However, gate-based processors currently suffer from severe limitations in qubit count, gate fidelity, coherence times, and connectivity, which restrict practical problem sizes to toy-scale instances and require sophisticated error mitigation techniques (Pal et al., 2024b). Consequently, gate-based approaches in bioinformatics remain largely confined to proof-of-concept demonstrations on small molecular fragments or simplified model systems.

4.2 Simulators vs. real quantum processors

A substantial majority of the reviewed studies (36 out of 53, 68%) did not execute on quantum hardware at all, instead relying on quantum circuit simulators, quantum-inspired classical algorithms, or hybrid classical workflows. This reflects the current reality that most bioinformatics problems of practical interest exceed the capacity of available quantum devices. Simulation remains the only viable route for algorithm development and validation at large scales (Nałęcz-Charkiewicz et al., 2024; Maurizio and Mazzola, 2025).

The distinction between quantum emulators and quantum-inspired methods is important. Quantum circuit simulators faithfully reproduce the behavior of a gate-based quantum computer by simulating the evolution of quantum states on classical hardware. These tools are invaluable for verifying circuit correctness, debugging quantum algorithms, and benchmarking performance before deployment on real devices. However, they scale exponentially with qubit count on classical hardware, limiting simulations to roughly 20–40 qubits on conventional clusters (Google Quantum AI, 2026; NVIDIA, 2022; Baruffa et al., 2022), which severely constrains the problem sizes that can be tested. Quantum-inspired algorithms, on the other hand, borrow operational concepts from quantum mechanics, such as qubit-based probability representations, quantum rotation gates, and superposition-like encodings, but execute entirely as classical heuristics with no claim to quantum speedup. Their relationship to genuine quantum advantage is indirect at best, and the terminological overlap with true quantum methods can create confusion about the nature of the computational contribution (Mokhtari et al., 2024).

The simulation and quantum-inspired landscape is notably fragmented across software platforms. Among the 36 studies in this category, we identified at least three distinct platforms: Qiskit (22%), MATLAB (17%), and PennyLane (14%). Custom or unspecified implementations represent 39% of studies.

This fragmentation reflects the absence of a dominant software ecosystem for quantum bioinformatics. The prevalence of unspecified or ad hoc implementations raises concerns about reproducibility and comparability across studies. Only 29% of these studies provide publicly available source code, further complicating the challenge of reproducibility.

5 Literature meta-analysis

A systematic examination of the 53 experimental studies included in this review reveals some notable patterns regarding performance evaluation and dataset scale across the quantum bioinformatics literature. These patterns show some positive signs, but they also reveal important limitations. Caution is required to interpret the findings presented in these articles. Tables 1–4 lists the key findings.

5.1 Performance relative to classical baselines

Of the 53 studies, 40 (75%) include an explicit comparison between their quantum or quantum-inspired implementation and at least one classical baseline method. Strikingly, every one of these 40 studies reports performance that is either comparable to or better than the corresponding classical approach. Not a single study in our corpus reports that a quantum implementation performed significantly worse than its classical counterpart on the tested problem instances. This pattern holds across all four thematic groups and all four quantum paradigms. However, we point out that these 40 papers focus on the quality of the results produced by their chosen methods. None of the considered articles explicitly demonstrates a quantum advantage over classical approaches.

Among quantum annealing studies, Fox et al. (2022) report that the D-Wave annealer is highly competitive with replica exchange Monte Carlo for RNA folding; Li et al. (2018) and Willsch et al. (2020) find that D-Wave-trained classifiers match or slightly exceed classical machine learning baselines for transcription factor binding prediction. Nałęcz-Charkiewicz and Nowak (2022) demonstrate that hybrid classical-quantum genome assembly produces results comparable to those of classical solvers. Among gate-based studies, Zhang Y. et al. (2026) reports that VQE-based protein structure prediction on the IBM Eagle processor outperforms AlphaFold3 in root mean square deviation and docking efficacy. Uttarkar and Niranjan (2026) show that error-mitigated VQE on the IonQ Aria resolves disordered protein-folding landscapes with similar results to classical methods. Among quantum-inspired and simulation studies, the pattern is equally consistent: Suo et al. (2020, 2023) report faster convergence for quantum-inspired genetic algorithms on the double digest problem, Dabba et al. (2021) find improved performance, and Saggi et al. (2026) demonstrate competitive multi-omics classification accuracy.

While these results are encouraging, the uniformly positive outcome across all 40 studies warrants cautious interpretation. Publication bias is the well-documented tendency for studies reporting positive or neutral results to be published preferentially over those reporting negative findings (Ioannidis et al., 2014; Song et al., 2013; Thornton, 2000; Marks-Anglin and Chen, 2020; Begg and Berlin, 1988). It is implausible that no research group has ever observed a quantum implementation performing worse than a classical baseline on a bioinformatics task; rather, such negative results are unlikely to survive the publication process. This systematic absence of negative results makes it difficult to assess the true distribution of quantum vs. classical performance across the problem landscape. Future work would benefit from registered reports or dedicated platforms for negative results to provide a more balanced empirical picture. In conclusion, while publication bias is a known phenomenon in research publishing, it remains to be seen whether the problem applies to quantum bioinformatics, too.

5.2 Dataset scale and the hardware-size bottleneck

A second notable pattern concerns the relationship between quantum paradigm and dataset scale. Of the 53 examined studies, 41 (77%) were evaluated exclusively on toy or substantially reduced datasets, while only 12 (23%) employed datasets approaching real-life scale.

5.2.1 Dataset size definition

In this review, the distinction between “Toy or reduced” and “Similar to real-life cases” datasets is determined by whether the problem size reflects true biological complexity or is artificially constrained due to current hardware limitations. Since bottlenecks manifest differently across various bioinformatics problems, we provide multiple definitions of “toy-scale” for each domain:

• Sequence analysis: Datasets restricted to very short fragments rather than to biologically realistic megabase-to-gigabase genomes or full-length mRNA transcripts.
• Structural biology: Problems simplified to coarse-grained lattice models or heavily pruned structural sub-pockets, as opposed to all-atom, full-length protein structures containing hundreds or thousands of residues.
• Phylogenetics and graph networks: Inputs limited to a handful of sequences for multiple sequence alignment, dozens of taxa for tree inference, or minimal graphs with tens of nodes. In contrast, real-world biological networks and phylogenies may span hundreds to thousands of interacting entities.
• Biomedical machine learning: Quantum machine learning often handles “real-life” datasets by applying severe classical dimensionality reduction to massive clinical cohorts before quantum processing. Therefore, in this specific domain, “toy-scale” denotes the use of synthetic, illustrative, or arbitrarily down-sampled data rather than validated, full-scale clinical cohorts.

5.2.2 Dataset scale distribution

The distribution of dataset sizes across paradigms is highly asymmetric. Among the 11 quantum annealing studies, 10 used toy or reduced datasets, with only Li et al. (2021), who applied D-Wave 2000Q to TCGA multi-omics cancer classification, testing on data of a realistic scale. Among the six gate-based studies, only Aramyan et al. (2022), who used the Rigetti Aspen-8 for SP-A/SARS-CoV-2 docking, employed a dataset comparable to real-life cases.

By contrast, among the 36 quantum-inspired or simulation-based studies, 10 (28%) were evaluated on datasets of realistic scale, including applications in cancer genomics (Akpinar and Oduncuoglu, 2025), graph classification (Bai et al., 2015; Kishi et al., 2022; Zhul et al., 2018), multi-omics classification (Saggi et al., 2026), disease prediction (Ullah et al., 2022), medical imaging (Youssry et al., 2016), gene regulation (Dabba et al., 2021), protein-protein interaction (Yang et al., 2019), and transcriptomics (Abrar-Ul-Haq and Bonny, 2020).

This disparity directly reflects the hardware constraints of current quantum devices. Even the largest available quantum annealers and gate-based processors impose severe limits on the problem sizes that can be natively encoded, due to qubit count, connectivity topology, and noise (Nałęcz-Charkiewicz et al., 2024). Bioinformatics problems of practical interest, such as genome assembly for organisms with megabase-scale genomes, protein folding for chains of hundreds of residues, or classification over thousands of patient samples with tens of thousands of features, far exceed these capacities. Quantum-inspired algorithms, executing on classical hardware without qubit limitations, face no such constraint and can therefore be tested on realistically sized inputs. However, this also means that the most practically relevant evaluations in the current literature come from methods that do not actually use quantum hardware, creating a paradox in which the approaches most likely to demonstrate genuine quantum advantage are precisely those least able to be tested at a meaningful scale.

6 Reproducibility assessment

Reproducibility is a foundational principle of the scientific method. The ability of independent researchers to verify, replicate, and build upon published results is essential for establishing the credibility of any computational method and for allowing incremental scientific progress (Munafò et al., 2017). In the context of quantum bioinformatics, a field still in its early stages, reproducibility takes particular importance: claims of quantum advantage or competitive performance must withstand independent scrutiny, and the research community must be able to access, modify, and extend proposed algorithms to advance the field collectively. When source code and experimental datasets are not made publicly available, the scientific community is forced to rely solely on the authors' reported results, with no mechanism for independent verification or error detection. Moreover, the absence of shared implementations significantly hinders future development, as researchers must re-implement algorithms from textual descriptions that may omit critical implementation details, hardware-specific parameter choices, or preprocessing steps.

6.1 Criteria for code and data availability assessment

To ensure a rigorous and consistent evaluation, we adopted strict criteria to determine whether a study's source code and experimental data are available for reproducibility. Source code is considered available only if it is deposited in a freely and publicly accessible repository that can be accessed by any researcher without restriction or solicitation. Pseudocode, partial code snippets embedded in the manuscript, or proprietary implementations do not qualify. Similarly, statements indicating that code is “available upon request” are not considered valid, as such mechanisms introduce gatekeeping that undermines the principle of unsolicited, immediate access.

For experimental data, we consider datasets available if the authors either directly provide the raw data in a public repository or supplementary material, or if they specify the exact database, accession numbers, or repository identifiers from which their test cases were obtained, provided that such repositories are freely and publicly accessible. Vague references to data sources without precise identifiers, or data stored behind institutional access restrictions, do not meet our threshold. These criteria reflect established best practices for computational reproducibility in scientific research, which emphasize that all materials necessary to verify and extend published results must be immediately and unconditionally accessible to the community.

6.2 Code and data availability

Our analysis of 53 experimental studies reveals a concerning level of reproducibility (Figure 7). Only 18 studies (34%) provide publicly available source code, while 35 (66%) do not share their implementations. Data availability is somewhat better: 37 studies (70%) provide access to their experimental datasets, either by directly publishing the data or by pointing to the exact public repository from which test cases were obtained, while 16 studies (30%) do not make their data accessible. Critically, only 16 studies (30%) make both code and data simultaneously available, which is the minimum requirement for full computational reproducibility. At the other extreme, 14 studies (26%) provide neither code nor data, rendering their results entirely unverifiable by independent researchers.

6.3 Challenges specific to quantum reproducibility

Beyond the general challenges shared with all computational disciplines, quantum computing introduces several unique obstacles that make reproducibility particularly difficult, even when authors act in good faith.

First, access to quantum processing units (QPUs) is currently restricted and expensive. Most quantum hardware is available only through cloud-based services with limited or no free tiers, institutional access agreements, or pay-per-use pricing models (Kordzanganeh et al., 2023). Researchers trying to reproduce a published result may not have access to the same device or face prohibitive costs to run the required number of experiments. This contrasts sharply with classical computing, where most researchers with a standard workstation can typically reproduce a simulation-based study.

Second, quantum hardware is not static: QPU calibration parameters, including qubit coherence times, gate fidelities, readout error rates, and crosstalk characteristics, change continuously, often over hours to days (Dasgupta and Humble, 2024; Proctor et al., 2020; Senapati et al., 2023). A circuit that produces high-quality results on a given device today may yield substantially different outcomes tomorrow, even with identical compilation and execution parameters. This temporal variability means that exact numerical reproduction of quantum hardware results is generally not achievable, and researchers must instead aim for statistical consistency within noise margins. An additional problem is that older hardware generations are routinely decommissioned. For example, the D-Wave 2X and 2000Q systems used in several of the reviewed studies are no longer commercially available, making literal reproduction of those experiments impossible.

Third, the probabilistic nature of quantum computation introduces inherent stochasticity. Quantum measurements are probabilistic, and finite shot counts introduce sampling noise into all results (Mari et al., 2021; Barron et al., 2024). While this can be mitigated by increasing the number of shots, i.e., experiment runs, the precise statistical fluctuations in any given run are irreproducible by definition. Authors rarely report the full distribution of outcomes across multiple independent runs, instead presenting only summary statistics that may obscure run-to-run variability.

Fourth, quantum hardware platforms often operate as black boxes with limited transparency regarding their internal operations (Roy, 2025; Roy et al., 2025). Users typically have little control over, and limited information about, the physical qubit-to-logical qubit mapping, the specific error-mitigation strategies applied by the backend, the compilation and transpilation passes applied to submitted circuits, or the scheduling and queuing policies that may affect execution conditions. These hidden variables can significantly influence results but are rarely documented in sufficient detail to reproduce the results.

Fifth, quantum algorithms are highly sensitive to hyperparameters and configuration choices (Lubinski et al., 2024; Pellow-Jarman et al., 2024; Bonet-Monroig et al., 2023). For quantum annealing, these include the annealing schedule, chain strength for minor embedding, number of reads, and post-processing strategies. For gate-based variational algorithms, critical choices include the ansatz architecture, parameter initialization, selection of a classical optimizer, learning rate, number of variational layers, and the error mitigation technique. Small changes in any of these parameters can dramatically alter solution quality. Yet, many publications do not report these details, making faithful reproduction difficult even when the source code is available.

6.4 Implications and recommendations

The combination of low code-sharing rates and the inherent challenges of quantum reproducibility creates a situation in which the vast majority of results in the quantum bioinformatics literature cannot be independently verified. This is particularly problematic given that, as noted in our benchmarking analysis, all comparative studies report performance equal to or better than classical baselines. This uniformly positive outcome would benefit from independent confirmation. The field would be substantially strengthened by adopting community standards for reporting quantum experiments, including mandatory disclosure of all hardware parameters, compilation settings, and hyperparameter choices; publishing source code in persistent, public repositories; and using standardized benchmark datasets that enable fair cross-study comparisons.

7 Discussion

The studies reviewed in this manuscript highlight a field that is simultaneously promising and immature. Quantum bioinformatics has evolved from theoretical proposals to a diverse collection of experimental studies covering four major areas: sequence analysis and genomics, protein structure and molecular simulation, phylogenetics and genomic information theory, and quantum machine learning for biomedical classification. These studies utilize all three main quantum computing paradigms: quantum annealing, gate-based quantum computing, and quantum-inspired or simulation-based approaches. This broad range of activities, achieved over roughly two decades of effort, demonstrates that the intersection of quantum computing and life sciences is not just theoretical; it has yielded concrete algorithmic contributions to a variety of biologically relevant problems.

7.1 Encouraging findings

Several aspects of the current landscape are genuinely encouraging. First, the positive performance reported across all comparative studies indicates that quantum and quantum-inspired methods are at least as competitive as classical baselines on the tested problem instances. In some cases, quantum approaches have demonstrated structural advantages that are difficult to replicate classically. These findings suggest that quantum computing may not only allow faster execution of existing algorithms but also introduce qualitatively different computational strategies that complement classical approaches in ways not yet fully understood.

7.2 Persistent challenges

Despite these encouraging signs, the reviewed literature exposes several fundamental challenges that must be addressed before quantum bioinformatics can transition from proof-of-concept demonstrations to practical impact.

The most pressing limitation is the severe constraint on problem scale imposed by current quantum hardware. As our analysis demonstrated, 88% of quantum annealing and gate-based studies were restricted to toy or substantially reduced datasets. Even the largest available quantum processors cannot natively encode bioinformatics problems of practical interest yet: genome assembly for organisms with megabase-scale genomes, protein folding for chains of hundreds of residues, or classification over thousands of patient samples with tens of thousands of genomic features. Until quantum hardware scales by orders of magnitude in qubit count, connectivity, and gate fidelity, the biological problems that can be addressed on real quantum processors will remain far smaller than those routinely solved by classical methods.

Preskill's (2025) paper offers quite an up-to-date snapshot of the current state of the quantum computing hardware field, including fault-tolerant (i.e., error-corrected) quantum computing, which is widely believed to be necessary for demonstrating quantum advantage in practically relevant applications, including bioinformatics. We remark that while error correction is a general hurdle for all quantum computers, different quantum technologies progress at different rates in tackling other hurdles. For example, superconducting quantum hardware is currently at the forefront of qubit count and error correction (Caune et al., 2026), but still suffers from limited connectivity. Approaches are being developed to address this problem (Escofet et al., 2025). On the other hand, neutral atom-based quantum systems offer “natural” all-to-all qubit connectivity and excellent error correction, but still lag behind with respect to qubit count (Bluvstein et al., 2026).

More specifically, practical quantum advantage in bioinformatics will require several milestones to be met simultaneously: sufficiently large numbers of high-fidelity logical qubits to encode biologically relevant problem instances; fault-tolerant operations with logical error rates compatible with the circuit depths required by the target algorithms; connectivity and routing overheads that do not eliminate the theoretical computational benefit; and end-to-end runtimes that outperform optimized classical or HPC baselines at comparable accuracy.

The consistently positive results reported across all comparative studies are encouraging, but they should be interpreted with caution. Publication bias could play a significant role in this trend. The lack of negative findings makes it hard to accurately evaluate the true performance of quantum vs. classical approaches. However, in our work, we have not found evidence of publication bias, so it is currently unclear whether the problem affects quantum bioinformatics as well.

Reproducibility remains a major concern. Only 34% of all reviewed studies provide publicly available source code, and only 30% make both code and data simultaneously accessible. The challenges related to quantum reproducibility are further complicated by other issues: restricted and expensive access to QPUs, temporal variability in hardware calibration, the probabilistic nature of quantum measurements, the black-box character of vendor platforms, and high sensitivity to hyperparameters. When the majority of published results cannot be independently verified, the credibility of reported performance claims is weakened, and the ability to build upon existing work is severely restricted.

The software ecosystem for quantum bioinformatics is fragmented and lacks standardization. Our analysis identified many distinct platforms across the simulation and quantum-inspired studies alone, with nearly half of all such studies using custom or unspecified implementations. Unlike classical bioinformatics, which benefits from mature, community-adopted frameworks and tools, quantum bioinformatics lacks a comparable infrastructure. This fragmentation prevents interoperability, complicates benchmarking, and raises the barrier to entry for new researchers.

Lastly, the boundary between truly quantum methods and quantum-inspired classical heuristics remains blurred in the literature. Quantum-inspired and simulation-based algorithms, which borrow concepts from quantum mechanics but execute entirely on classical hardware, constitute the majority of reviewed studies. While these methods have produced valuable results, their relationship to genuine quantum advantage is indirect. The terminological overlap between quantum-inspired and truly quantum methods can create confusion about the nature of the computational contribution. It may also inflate perceptions of quantum readiness in the field.

7.3 Recommendations for future work

Based on the patterns identified in this review, we offer several recommendations for future work.

First, the community should prioritize developing and adopting standardized benchmark datasets and evaluation protocols for quantum bioinformatics. Such benchmarks should span multiple problem types and scales, include both toy instances suitable for current hardware and larger instances for simulation-based evaluation, and define clear metrics for comparing quantum and classical approaches. Dedicated platforms for negative results would help compensate for publication bias and provide a more balanced empirical picture.

Second, authors should commit to full computational reproducibility by depositing complete source code, hardware configuration parameters, and experimental data in persistent public repositories. The field cannot mature if the majority of published results remain unverifiable.

Third, future studies should explicitly characterize the scaling behavior of their proposed methods, rather than reporting performance only at the small scales currently accessible on quantum hardware. Understanding how solution quality and computational cost evolve as problem size increases is essential for assessing whether a quantum approach will retain its advantages at practically relevant scales.

Fourth, the development of domain-specific quantum software libraries for bioinformatics would significantly lower the barrier to entry, improve reproducibility, and facilitate fair comparisons across studies. Such libraries should abstract away hardware-specific details while exposing the parameters necessary for reproducible experimentation.

Fifth, researchers should maintain clear terminological distinctions between quantum hardware experiments, quantum circuit simulations, and quantum-inspired classical heuristics. Each category makes fundamentally different claims about computational advantage, and mixing them obscures the true state of progress toward practical quantum benefit in bioinformatics.

8 Conclusion

Quantum computing holds substantial promise for transforming computational biology and biomedicine. The intrinsic quantum-mechanical nature of molecular interactions, the combinatorial complexity of genomic optimization problems, and the exponential dimensionality of biomedical feature spaces all point to a natural and potentially revolutionizing role for quantum computation in the life sciences. This systematic review has investigated 53 experimental studies that collectively demonstrate the extent of this emerging field, spanning genome assembly, sequence alignment, protein folding, molecular simulation, drug discovery, phylogenetics, and machine learning for clinical and multi-omics data.

The current state of the field can be represented by credible prototypes. Across all four thematic groups and all four quantum paradigms, researchers have produced working implementations that match or exceed classical baselines on the problem instances tested. Real quantum hardware has been successfully applied to biologically meaningful tasks, demonstrating that NISQ-era devices can already contribute to bioinformatics, but on a limited scale. Quantum-inspired methods, while not directly using quantum hardware, have demonstrated competitive performance on realistically sized datasets and introduced algorithmic innovations.

Looking ahead, the trajectory of quantum hardware development provides grounds for optimism. Qubit counts, gate fidelities, coherence times, and connectivity are improving steadily across all major hardware platforms. As these improvements continue, the problem sizes accessible to quantum processors will grow, slowly closing the gap between the toy-scale demonstrations that dominate the current literature and the large-scale problems that define practical bioinformatics. Equally important, the maturation of the quantum software ecosystem will lower the barrier to entry and improve the reliability and reproducibility of quantum bioinformatics research.

The path from prototype to practice will require interdisciplinary collaboration between quantum physicists, computer scientists, and domain experts in biology and medicine. It will also require the community to adopt rigorous standards for benchmarking, reproducibility, and transparent reporting that match the methodological expectations of both the quantum computing and bioinformatics communities. If these challenges are met, quantum computing has the potential to become not merely an alternative to classical computation for biological problems, but a new lens through which the complexity of living systems can be understood and harnessed for human benefit.

Generative AI statement

The author(s) declared that Generative AI was used in the creation of this manuscript. We acknowledge the use of Grammarly for assisting in improving the grammar, wording, and overall sentence structure of the original version of this manuscript. The authors carefully reviewed all suggestions from Grammarly, ensuring that the enhancements aligned with their intended message.

Any alternative text (alt text) provided alongside figures in this article has been generated by Frontiers with the support of artificial intelligence and reasonable efforts have been made to ensure accuracy, including review by the authors wherever possible. If you identify any issues, please contact us.

Publisher's note

All claims expressed in this article are solely those of the authors and do not necessarily represent those of their affiliated organizations, or those of the publisher, the editors and the reviewers. Any product that may be evaluated in this article, or claim that may be made by its manufacturer, is not guaranteed or endorsed by the publisher.

Supplementary material

The Supplementary Material for this article can be found online at: https://www.frontiersin.org/articles/10.3389/fhpcp.2026.1923360/full#supplementary-material

Supplementary material Supplementary material at the source

Supplementary material Supplementary material at the source

Footnotes

Edited by: Umberto Ferraro Petrillo, Sapienza University of Rome, Italy

Reviewed by: Daniele Loco, Qubit Pharmaceuticals, France

Amjed Abbas Ahmed, Imam Al-Kadhum College, Iraq