HybridQC: Bioinformatics & Quantum Workflows

• HybridQC is a term covering two domains: an R package for scRNA-seq quality control using threshold filtering with Isolation Forests, and a descriptor for hybrid quantum-classical workflows.
• In bioinformatics, HybridQC processes Seurat objects by combining fixed threshold rules with multivariate anomaly detection for effective quality filtering of single-cell RNA-seq data.
• In quantum computing and machine learning, HybridQC represents architectures that partition tasks between classical optimization and quantum evaluation to boost computational efficiency.

HybridQC denotes at least two distinct constructs in current literature. In bioinformatics, it is an R package for single-cell RNA-seq quality control that combines threshold-based filtering with Isolation Forest-based anomaly detection in Seurat workflows (Lai, 10 Jul 2025). In quantum computing, the same label and closely related forms such as HCQ, HQC, and hybrid quantum-classical optimization are used for architectures that partition computation across classical processors and quantum resources, usually under NISQ-era constraints and with explicit classical orchestration, optimization, or post-processing (Lamża et al., 2024). The term is therefore context-dependent rather than canonical.

1. Terminological scope

In the cited literature, “HybridQC” is not univocal. One usage is a named software package for omics preprocessing; the other is a broader descriptor for hybrid quantum-classical systems, software stacks, and algorithms.

Usage	Domain	Defining feature
HybridQC	scRNA-seq QC	Threshold filtering plus Isolation Forest on Seurat objects
HybridQC / HCQ / HQC	Quantum computing	Classical–quantum partitioning of optimization, learning, or execution

The scRNA-seq usage is explicit in the paper introducing the R package HybridQC (Lai, 10 Jul 2025). The quantum-computing usage appears across work on hybrid quantum-classical optimization, system integration, and software migration, where the emphasis is not a single framework but a class of workflows in which classical and quantum components are deliberately co-designed (Maio et al., 14 Mar 2025). This suggests that HybridQC is best treated as a family of architectures rather than a single standardized method.

2. HybridQC as single-cell RNA-seq quality control

In single-cell transcriptomics, HybridQC is a lightweight R package for quality control of scRNA-seq data. Its core design is a two-stage workflow: compute standard per-cell QC metrics, run Isolation Forest on those QC features, and filter cells flagged by either criterion (Lai, 10 Jul 2025). The standard metrics explicitly listed are nFeature_RNA, nCount_RNA, and percent.mt. The package is designed for Seurat objects, accessed from R, and uses reticulate plus Python pyod for the machine-learning backend.

The defining decision rule is an OR rule. If $T_i$ denotes failure of threshold-based QC and $M_i$ denotes Isolation Forest outlier status, then the package’s hybrid filter is reconstructed as

$H_i = T_i \lor M_i,$

so a cell is removed if it fails threshold QC or is flagged by the anomaly detector (Lai, 10 Jul 2025). This makes the package simultaneously interpretable and adaptive: fixed cutoffs remain visible, but multivariate outliers that look acceptable on any single metric can still be removed.

The intended workflow is operationally simple. A Seurat object is loaded, run_basic_qc() computes the standard metrics, run_isolation_forest_qc() assigns anomaly scores, and filter_cells() removes cells flagged by either route. The paper also describes UMAP visualization of Isolation Forest scores via Seurat’s plotting stack, notably FeaturePlot(), and demonstrates the package on a synthetic 10x Genomics-style PBMC dataset containing 2,000 cells and 1,000 genes (Lai, 10 Jul 2025).

Its stated use cases are noisy or shallow-depth datasets, non-model organisms, rare samples, and pilot studies. At the same time, the paper is explicit about limitations: it does not provide rigorous quantitative benchmarks, default threshold rules are not fully documented, Isolation Forest hyperparameters are not specified, and the OR rule creates a risk of over-filtering unusual but genuine cell states (Lai, 10 Jul 2025). In this sense, HybridQC in bioinformatics is a workflow packaging contribution rather than a new anomaly-detection theory.

3. HybridQC in quantum optimization workflows

In quantum-computing research, HybridQC commonly refers to staged or nested workflows in which classical optimization and quantum evaluation are interleaved. QHyper is exemplary: it is an integration library for hybrid quantum-classical optimization that organizes experiments around four modules—problem definition, converter, solvers, and optimizers—and supports variational gate-based algorithms, D-Wave annealing or hybrid services, and classical reference solvers such as Gurobi (Lamża et al., 2024). Its internal Polynomial representation decouples problem modeling from solver-specific forms such as QUBO, CQM, and DQM, while local and global optimizers can be nested around a solver to tune circuit parameters, penalty weights, or initialization.

The resulting pattern is explicitly hierarchical. In the notation reconstructed in the paper, one may optimize hyperparameters $h$ outside a variational loop over $\theta$,

$$h^*=\arg\min_{h\in\mathcal{H}} F(h), \qquad F(h)=\min_\theta f(\theta;h),$$

which captures a common HybridQC structure: classical global search over encodings or penalties, classical local search over variational parameters, and quantum or hybrid evaluation inside the loop (Lamża et al., 2024).

A different HybridQC design appears in the QCQP solver of “A hybrid algorithm for quadratically constrained quadratic optimization problems,” where the quantum device evaluates objective and derivative information for a variationally encoded state, while a classical primal-dual interior-point method enforces the quadratic constraints (Zhou et al., 2023). The unified optimization form is

$$\min_{\boldsymbol{\vartheta}} F_0(\boldsymbol{\vartheta}) \quad \text{s.t.} \quad F_i(\boldsymbol{\vartheta}) \le 0,$$

with $\boldsymbol{\vartheta}=(\eta,\boldsymbol{\theta})$. The stated advantage is logarithmic qubit scaling in the variable dimension under amplitude encoding, while the classical side retains mature constrained-optimization machinery (Zhou et al., 2023).

Resource-aware hybridization is even more explicit in hqQUBO for multiple sequence alignment. There, Hybrid Query Encoding keeps only gap/occupancy decisions on qubits and offloads residue-identity lookup to classical query functions, yielding a hybrid objective

$$L(x;p)=\sum_{i\neq j}\sum_k w^{(i,j)}_{f_i(k),f_j(k)}x_{i,k}x_{j,k} +p\sum_i\left(\sum_k x_{i,k}-l_i\right)^2.$$

The paper’s main systems claim is that this reduces qubit usage to $\mathcal{O}(NL)$ rather than the $M_i$0 associated with more explicit encodings (Chen et al., 2 Jun 2025).

A more workflow-level formulation is HSQC, or Hybrid Sequential Quantum Computing, which treats hybridization as an ordered search pipeline rather than an inner variational loop. The two reported instantiations are SA $M_i$1 BF-DCQO $M_i$2 MTS and SA $M_i$3 BF-DCQO $M_i$4 SA. On 156-qubit cubic HUBO instances executed on IBM heavy-hex hardware, the authors report estimated runtime speedups of up to $M_i$5 over standalone simulated annealing and up to $M_i$6 over memetic tabu search, with the quantum stage framed as a basin-transfer or refinement phase rather than a complete solver (Chandarana et al., 7 Oct 2025).

4. HybridQC in machine learning architectures

A major research line uses HybridQC to denote hybrid learning architectures in which a small quantum module is embedded inside an otherwise classical model. The “Hybrid classical-quantum Autoencoder for Anomaly Detection” paper places a 4-qubit PQC in the bottleneck of a classical autoencoder, measures Pauli-$M_i$7 expectation values, and then applies Isolation Forest in the resulting latent space (Sakhnenko et al., 2021). On the reported public benchmarks, HAE improves F1 on all three datasets and improves recall on all three; for example, on Satellite, AE reaches F1 $M_i$8 and HAE reaches F1 $M_i$9 (Sakhnenko et al., 2021).

A related pattern appears in the tensor-network-plus-VQC classifier for MNIST 3-vs-6. There, a trainable MPS compresses 784-dimensional input to a 4-dimensional feature vector for a 4-qubit VQC, with end-to-end training across the classical and quantum boundary. The reported result is 99.44% test accuracy for MPS-VQC with $H_i = T_i \lor M_i,$0, compared with 87.34% for PCA-VQC under the same 4-qubit downstream model (Chen et al., 2020). The paper’s significance for HybridQC is architectural: the classical compression front-end is not fixed preprocessing but part of the jointly trained pipeline.

Other examples keep the quantum component even more tightly scoped. The histopathology paper uses pretrained CNN backbones such as ResNet18 and replaces the final fully connected layer with a VQC-based QNN, yielding 88.5% accuracy for Hybrid ResNet18 on 1,000 images versus 88.0% for Classical ResNet18, although the classical model still has better AUC in that comparison (Majumdar et al., 2023). The Alzheimer’s HCQ pipeline uses a supervised 3D $H_i = T_i \lor M_i,$1-VAE to compress MRI volumes to a 64-dimensional latent code, PLS reduces this to six components, and a six-qubit ZZ quantum feature map feeds a precomputed-kernel SVM; the baseline reaches 67.2% accuracy and 0.759 AUC, while the stability-enhanced variant reaches 72.1% accuracy and 0.799 AUC (Tiwari et al., 12 Jun 2026).

HybridQC has also been applied to quantum data analysis itself. The hybrid quantum autoencoder for clustering and enhanced classification maps pure quantum states to a classical latent vector via encoder expectation values, then applies Gaussian mixture models or classical classifiers in that latent space. The reported downstream results include 85.1% clustering accuracy and up to 0.95 classification accuracy with an SVM on the learned latent representation (Srikumar et al., 2021). In reinforcement-learning form, HCQA couples a classical DQN with a quantum action-selection circuit for generating two-qubit quantum sensor circuits; the reported output reaches normalized QFI $H_i = T_i \lor M_i,$2 and outperforms the compared classical DQN, QRA, and GAQA baselines on the stated task (Alomari et al., 28 Aug 2025).

Across these papers, the recurring design pattern is classical bulk computation plus a small quantum subroutine inserted at an information bottleneck, kernel interface, or policy-selection stage. This suggests that, in present practice, HybridQC in machine learning is predominantly an interface-engineering problem rather than a replacement of classical deep models by large quantum networks.

5. Systems, simulation, and HPC-oriented HybridQC

At system scale, HybridQC often refers to the co-design of orchestration, compilation, runtime, and hardware access. ScaleQC is a hybrid execution framework based on circuit cutting. It partitions a circuit into subcircuits feasible for smaller QPUs, then reconstructs outputs classically via tensor-network contraction. Its reconstruction formula is

$H_i = T_i \lor M_i,$3

and its states-merging strategy is used to search concentrated-output circuits without materializing the full state vector. The paper reports benchmarks up to 1,000 qubits and frames the central challenge as balancing QPU limits against classical post-processing cost (Tang et al., 2022).

QMIO addresses a different bottleneck: rapid repeated classical–quantum iteration in an HPC center. It integrates a classical HPC cluster, a 32-qubit Oxford Quantum Circuits superconducting processor, and a 34-qubit classical emulator in one on-premises environment, with a middleware layer that uses ZeroMQ between a gateway node and the quantum control node (Cacheiro et al., 25 May 2025). The architectural lesson is concrete: exposing the QCN directly to SLURM introduced 1–3 seconds of overhead per quantum job, which was unacceptable for iterative hybrid algorithms, so the deployed design separates coarse-grained scheduler allocation from low-latency message-based circuit dispatch (Cacheiro et al., 25 May 2025).

HybridQ addresses hybridization at the simulation-software level. It provides one framework for state-vector evolution, tensor contraction, Clifford expansion, and noisy-circuit simulation across CPUs, GPUs, TPUs, and MPI-enabled HPC systems (Mandrà et al., 2021). Its design is explicitly method-hybrid and hardware-hybrid rather than quantum-hardware hybrid in the deployment sense of QMIO. For HybridQC practice, this matters because much algorithm development still depends on classical simulation stacks capable of switching representations and hardware targets without rewriting circuits.

The software-engineering dimension is analyzed directly in “The Road to Hybrid Quantum Programs,” which treats hybridization as evolution from an existing classical program to a functionally equivalent hybrid quantum program. The process is characterized by task decomposition, quantum-candidate identification, data encoding, quantum implementation, and integration back into a classical–quantum loop, with the original classical computation retained as an oracle for differential evaluation (Maio et al., 14 Mar 2025). This is a systems view of HybridQC in which the critical challenge is not only algorithm design but also code migration, variant management, and interdisciplinary coordination.

6. Limitations, misconceptions, and research trajectory

The literature is consistent on one point: HybridQC is technically productive but methodologically uneven. Many papers emphasize workflow or architectural novelty while leaving important details under-specified. The scRNA-seq HybridQC package does not provide rigorous benchmarks or exact default hyperparameters (Lai, 10 Jul 2025). The Alzheimer’s HCQ pipeline does not include classical-kernel baselines on the same six-dimensional latent representation, so the marginal contribution of the quantum kernel is not isolated (Tiwari et al., 12 Jun 2026). The selective-state-space hybrid Mamba model reports only four training epochs on reshaped MNIST and leaves key circuit details unspecified (Ebrahimi et al., 11 Nov 2025). The histopathology HQC study shows only marginal gains in one small-data setting and stronger classical performance on larger data (Majumdar et al., 2023). These are not contradictions; they indicate that HybridQC remains dominated by proof-of-concept studies and system-level design exploration.

A separate misconception concerns the paper “Toward a Unified Hybrid HPCQC Toolchain” (Seitz et al., 2023). The supplied document is not a substantive contribution on HybridQC or HPCQC; it is an IEEE conference LaTeX template populated with placeholder text, with no architecture, algorithms, implementation, or evaluation relevant to hybrid quantum-classical computing. It functions only as a negative case: title-level terminology is not sufficient evidence of technical content.

The forward trajectory in the cited work is therefore less about a single mature paradigm than about convergence on a set of engineering themes. These include principled fusion between classical and quantum representations rather than naive concatenation, as argued in the multimodal feature-fusion paper (Alavi et al., 22 Dec 2025); explicit orchestration layers for nested optimization and heterogeneous backends, as in QHyper (Lamża et al., 2024); coarse-grained classical scheduling paired with low-latency quantum dispatch, as in QMIO (Cacheiro et al., 25 May 2025); and semi-automatic identification of quantum-ready code fragments in legacy software, as proposed by the software-evolution study (Maio et al., 14 Mar 2025). This suggests that the future of HybridQC will depend at least as much on interfaces, intermediate representations, and evaluation protocols as on the isolated performance of any single quantum subroutine.