REVIEW 4 major objections 5 minor 17 references
Alz-QNet: A Quantum Regression Network for Studying Alzheimer's Gene Interactions
T0 review · 4 major / 5 minor · reviewed 2026-08-06 · deepseek-v4-flash
Pith's one-line read Alz-QNet claims a variational quantum circuit infers Alzheimer's gene-gene interactions from binarized entorhinal-cortex data using half the controlled-rotation gates of the standard quantum gene regulatory network.
desk verdict The paper's main claim—halving QGRN gates because θx,y = θy,x—doesn't survive contact with Sec 5.2: the symmetry is imposed by copying the upper triangle, not observed, so the gate reduction and the inferred interactions are fitted artifacts. read the letter →
The pith
A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.
The reading
What carries the argument
The central object is the variational quantum circuit Alz-QNet: an encoding layer converts each nucleus's 8-bit activation state into qubit rotations (initialized with $\theta = 2\arcsin(\sqrt{a_k})$ so each qubit's $|1\rangle$ probability matches the gene's observed activation frequency), and a parameterized layer of CRY gates couples every pair of gene-qubits. Each fitted angle $\theta_{x,y}$ is read as the strength of the interaction between genes $x$ and $y$. The load-bearing identity is $\theta_{x,y} = \theta_{y,x}$, which lets the paper use one CRY gate per unordered pair rather than two per ordered pair; the authors argue the resulting amplitude structure is unchanged, so the identica
What would settle it
Fit the same eight-gene data with a circuit that keeps two independent angles per gene pair ($\theta_{x,y}$ and $\theta_{y,x}$); if the fitted asymmetries exceed the noise level for any pair, the symmetric one-gate-per-pair parametrization is omitting directional information and the claimed 50% gate reduction no longer represents the same interaction model.
Extended reading notes
Core claim
Alz-QNet is a variational quantum circuit that takes, for each nucleus, an 8-bit binarized expression profile of APP, SREBF2, EGR1, YY1, PLD3, GAS7, FGF14, and AKT3 and is trained so that its output distribution matches the observed frequency distribution of activation patterns across 1,041 AD-patient nuclei. The interaction between genes $x$ and $y$ is encoded by a controlled-Y rotation angle $\theta_{x,y}$; the paper's central observation is that $\theta_{x,y} = \theta_{y,x}$, so the lower triangle of the interaction matrix can be filled by copying the fitted upper triangle, cutting the number of CRY gates from $n(n-1)$ to $n(n-1)/2$. With this reduction the model reproduces the QGRN basel
Load-bearing premise
The load-bearing premise is that gene-gene interactions are pairwise symmetric, because the paper builds the lower triangle of the interaction matrix by copying the fitted upper-triangle values, and if the true regulatory relationships carry direction, the 50% gate reduction and the inferred edges would not follow.
Editorial extensions
If this is right
- If Alz-QNet is right, quantum regression can recover the structure of a gene network from a few thousand binarized nuclei at half the circuit cost of the standard QGRN.
- The predicted edges, YY1 repressing PLD3, PLD3 repressing APP, and EGR1 activating APP/SREBF2/GAS7, become concrete hypotheses that can be tested by knockdown or overexpression experiments.
- The 50% reduction in controlled-rotation gates lowers the gate count for $n$ genes from $n(n-1)$ to $n(n-1)/2$, directly improving the scalability of quantum GRN inference on near-term hardware.
- The alignment between predicted $\theta$-values and independently documented mechanisms, such as APP activating SREBF2 and FGF14 repressing SREBF2, argues that the fitted angles carry biological signal rather than only fitting noise.
Reading between the lines
- The paper leaves implicit that, because the lower triangle is copied from the upper triangle, the fitted matrix is symmetric by construction; the circuit cannot by itself distinguish 'gene A regulates gene B' from 'gene B regulates gene A', so directed statements like 'YY1 represses PLD3' lean on prior biology, not on the quantum fit alone.
- A natural extension is to compare Alz-QNet's symmetric interaction strengths with a classical asymmetric GRN reconstruction on the same 1,041 nuclei; large discrepancies would localize where the symmetry assumption distorts the network.
- The gate-halving argument is essentially a parameter-sharing claim and should transfer to any pairwise symmetric interaction model in quantum machine learning, not just gene regulation; the paper's heatmap analysis is a concrete instance of that general principle.
Signed reviews
Editorial analysis
A structured set of objections, weighed in public.
Referee Report
Summary. The paper proposes Alz-QNet, a variational quantum circuit (VQC) that regresses a binarized gene-expression distribution derived from snRNA-seq entorhinal-cortex samples (GSE138852) for eight AD-related genes. The circuit uses one CRY gate per gene pair after imposing θx,y = θy,x, halving the n(n−1) gates of the earlier QGRN. The authors report a gene-interaction graph with activation/repression edges and interpret several edges (e.g., YY1 repressing PLD3) as biologically meaningful. Simulations in Qiskit show observed vs output frequency distributions for Alz-QNet and QGRN.
Significance. If valid, a quantum GRN method with halved circuit depth would be a useful step for NISQ-era QML in genomics, and the application to AD is timely. The paper is transparent about its proof-of-concept nature and gives a reasoned binarization protocol and a literature-support table. However, its central quantitative claim is not supported: the θ symmetry is imposed by construction, not learned, and the biological “predictions” are in-sample fitted parameters without error bars, baselines, or out-of-sample validation. The contribution is therefore currently more a circuit-architecture proposal than an established inference result.
major comments (4)
- [Sec. 5.2 / Contribution 2] The claimed observation θx,y = θy,x is not an empirical finding. The text states that the upper triangular matrix is the optimized theta and that the “lower triangular matrix was constructed from the equal values of θx,y = θy,x.” Thus the symmetry is imposed by copying, and the 50% gate reduction is true by construction. To support the scalability claim, the authors must fit the full QGRN with n(n−1) independent CRY angles on the same data and show (i) a comparable output distribution and (ii) that the symmetric optimum is not a local artifact. Without this, the assertion that half the gates “suffice” without altering the final quantum state is unsubstantiated.
- [Sec. 6.1 / Sec. 6.3] The reported “predictions” (e.g., YY1 represses PLD3, θ = −0.1129) are the fitted variational parameters, not predictions validated on unseen data. There is no train/test split, cross-validation, bootstrap, or uncertainty quantification. As a result, the biological conclusions are in-sample descriptions of the trained circuit. To claim the model “recovers biologically meaningful regulatory circuits,” the authors should evaluate on held-out nuclei or synthetic ground-truth networks and compare with classical GRN inference baselines (e.g., Pearson correlation, GENIE3, PIDC) using the same binarized data.
- [Sec. 5.2 / Fig. 4] “Similar probability distribution” is assessed only visually. No quantitative divergence or fit statistic is reported for Alz-QNet vs observed or QGRN vs observed. Given n = 8 qubits, the output space has 256 states and the observed distribution is sparse; visual agreement can be misleading. Report e.g. KL/JS divergence, total variation distance, or R² and include per-state residuals.
- [Sec. 5.2 / Fig. 3 and symmetric ansatz] Gene regulation is biologically directed. Imposing θx,y = θy,x forces every interaction to be symmetric, so the graph in Fig. 3 cannot represent directed regulation (activation/repression by a regulator of a target) without additional orientation information. The authors need to either justify the symmetry biologically or show that the fitted model is not missing directionality. Relatedly, a CRY gate is not symmetric under exchange of control and target; the matrix symmetry reported in Sec. 5.2 does not by itself imply the circuit structure is symmetric.
minor comments (5)
- [Abstract] “CE microenvironment” should be “EC microenvironment”; gene names such as FGF14 and SREBF2 are inconsistently spaced throughout (e.g., “FGF 14”, “S REBF 2”).
- [Fig. 5 caption] The caption says θ (X-axis) and ϕ1 (Y-axis) vary continuously, but each panel is for a different fixed value of ϕ1. Please clarify what is plotted in each heatmap and how the claimed symmetry is read from the figure.
- [Sec. 5.1] The choice of binarization at a Pearson residual threshold of zero is reasonable, but its effect on the inferred network is not tested. A sensitivity analysis over nearby thresholds would strengthen the results.
- [Throughout] No code or data availability statement is provided. Since the paper relies on Qiskit simulations and a specific filtering pipeline, releasing code would materially improve reproducibility.
- [Fig. 3 caption] The caption uses green/red edges for up/down regulation but the graph edges have no arrows. Please state explicitly whether edges are directed or undirected and how the sign is assigned.
Circularity Check
The θx,y=θy,x symmetry is imposed by copying the fitted upper triangle into the lower triangle, so the 50% gate reduction is true by construction; the 'predicted' gene interactions are the fitted CRY angles re-labeled as predictions.
-
self definitional
[Sec. 1, Contribution 2; Sec. 5.2, theta-matrix paragraph]
"We observed that the value of θx,y = θy,x. It enhances the scalability potential of our proposed Alz-QNet. ... The black upper triangular matrix is the optimized value of theta obtained, and the lower triangular matrix was constructed from the equal values of θx,y = θy,x."
The equality used to justify replacing n(n−1) CRY gates with n(n−1)/2 is not an empirical result of an unconstrained fit: the lower triangular matrix is filled by copying the fitted upper triangle. Thus θx,y = θy,x holds by construction, and the gate reduction simply restates the symmetric parameterization chosen by the authors. The paper gives no quantitative comparison showing that this reduced circuit reproduces the unconstrained QGRN output; the similar distribution in Fig. 4 is only visual. Contribution 2 therefore reduces to the definition of Alz-QNet's ansatz.
-
fitted input called prediction
[Sec. 6.1 and Sec. 6.3, 'YY1 and PLD3 Interaction' and 'Alz-QNet Model Predictions']
"Our results show a negative interaction between YY1 and PLD3 (−0.112907)... Our quantum model, Alz-QNet, predicts a robust network of six core gene interactions... The strong alignment between our predicted θ-values and established mechanisms..."
The 'predicted' interactions are exactly the optimized CRY angles of the variational circuit (e.g., θYY1,PLD3 = −0.112907), which were fitted to the observed binarized expression distribution. Calling these fitted parameters 'predictions' renames the regression output; there is no held-out set or independent biological target being forecast. The literature checks in Supplementary Table S1 are post-hoc comparisons on the same fitted θ-values, so they do not turn fitted inputs into predictions.
full rationale
The central load-bearing step is the claimed observation θx,y = θy,x. The paper itself states that the lower triangular matrix 'was constructed from the equal values of θx,y = θy,x' after the upper triangle was optimized, so the symmetry is imposed by the authors' construction rather than discovered from a full QGRN fit. Consequently, the 50% reduction in CRY gates is a property of the chosen symmetric ansatz, and the claim that it 'saves computations' without losing information is not independently validated; Fig. 4(c) is a visual similarity check with no quantitative fit metric. The biological 'predictions' (e.g., YY1 represses PLD3 with θ = −0.1129) are the fitted variational parameters themselves, so they are not predictions in an out-of-sample sense; they are the model output. This is a partial but substantial circularity: the paper's headline contribution (gate reduction) is true by definition, and its predicted interactions are fitted inputs relabeled. I do not flag the numerous self-citations of basic gate definitions or QGRN, because those are not load-bearing and are standard external references. The proof-of-concept caveat in Sec. 5.1 is noted but does not cure the construction-by-copy issue.
Assumptions & free parameters
free parameters (2)
- Binarization threshold =
0 (Pearson residual)
- Pairwise CRY interaction angles theta_x,y =
e.g., theta(YY1,PLD3) = -0.1129; 28 values for 8 genes
assumptions (5)
- domain assumption Gene regulation can be represented by binary switching states (expressed/not expressed) derived from Pearson residual sign.
- ad hoc to paper The interaction between any two genes is symmetric, so theta_x,y = theta_y,x and one CRY gate per pair suffices.
- ad hoc to paper The sign of the fitted theta determines activation vs repression (green/red edges).
- domain assumption The variational quantum circuit with only pairwise CRY gates can approximate the empirical joint distribution over 8 genes.
- standard math Standard quantum computing formalism (qubits, gates, unitary evolution).
Cite this review
Pith. "Pith review of Alz-QNet: A Quantum Regression Network for Studying Alzheimer's Gene Interactions." pith.science (2026). https://pith.science/paper/67NY2F3J
@misc{pith2026250804743,
author = {Pith},
title = {Pith review of: Alz-QNet: A Quantum Regression Network for Studying Alzheimer's Gene Interactions},
year = {2026},
howpublished = {\url{https://pith.science/paper/67NY2F3J}},
note = {Machine review of arXiv:2508.04743}
}
abstract
Understanding the molecular-level mechanisms underpinning Alzheimer's disease (AD) by studying crucial genes associated with the disease remains a challenge. Alzheimer's, being a multifactorial disease, requires understanding the gene-gene interactions underlying it for theranostics and progress. In this article, a novel attempt has been made using a quantum regression to decode how some crucial genes in the AD Amyloid Beta Precursor Protein ($APP$), Sterol regulatory element binding transcription factor 14 ($FGF14$), Yin Yang 1 ($YY1$), and Phospholipase D Family Member 3 ($PLD3$) etc. become influenced by other prominent switching genes during disease progression, which may help in gene expression-based therapy for AD. Our proposed Quantum Regression Network (Alz-QNet) introduces a pioneering approach with insights from the state-of-the-art Quantum Gene Regulatory Networks (QGRN) to unravel the gene interactions involved in AD pathology, particularly within the Entorhinal Cortex (EC), where early pathological changes occur. Using the proposed Alz-QNet framework, we explore the interactions between key genes ($APP$, $FGF14$, $YY1$, $EGR1$, $GAS7$, $AKT3$, $SREBF2$, and $PLD3$) within the CE microenvironment of AD patients, studying genetic samples from the database $GSE138852$, all of which are believed to play a crucial role in the progression of AD. Our investigation uncovers intricate gene-gene interactions, shedding light on the potential regulatory mechanisms that underlie the pathogenesis of AD, which help us to find potential gene inhibitors or regulators for theranostics.
Reference graph
Works this paper leans on
-
[1]
Introduction Alzheimer’s disease (AD) presents a formidable challenge in healthcare, characterized by progressive cognitive decline and neurodegeneration (Grubman et al., 2019). The accumulation of peptides of Amyloid Beta ( Aβ) from APP is a crucial factor in the pathology of AD (Wang et al., 2018). Despite extensive research efforts over decades, the in...
-
[2]
In this study, we optimized Quantum Gene Regulatory Networks (QGRN) (Roman-Vicharra and Cai, 2023) to address the computational complexity and expense asso- ciated with the construction of variational quantum cir- cuits for larger datasets. Our proposed Alz-QNet requires n(n−1) 2 controlled rotation gates over n(n−1) for the QGRN model, demonstrating that...
work page 2023
-
[3]
We observed that the value ofθx,y =θy,x
In our Alz-QNet circuit, the reduction of CRY gates by half compared to the traditional QGRN (Roman-Vicharra and Cai, 2023), optimizes the number of variational parame- ters,θx,y between two qubits x and y. We observed that the value ofθx,y =θy,x. It enhances the scalability potential of our proposed Alz-QNet
work page 2023
-
[4]
The study also aimed to uncover how APP interacts with other key genes in AD, potentially offering information on control-based therapies for the gene expression of the dis- ease
-
[5]
Furthermore, beyond APP, the research used single nu- cleus RNA sequencing of the entorhinal cortex to explore less explored genes such asYY 1, S REBF 2, PLD3, GAS 7, and EGR1. Motivation: QML research in general has demonstrated better performance in term of generalization, expressivity, privacy and robustness (Konar et al., 2023c,a; Caro et al., 2022). ...
work page 2022
-
[6]
Various Gene Types with Alzheimer’s Diseases Comprehending the regulatory connections among pivotal genes associated with AD, such as APP, S REBF 2, and other relevant genes, is essential for deciphering the molecular mech- anisms driving AD pathogenesis. These genes are fundamen- tal in synaptic function, lipid metabolism, and neuronal viabil- ity, and t...
work page 2021
-
[7]
Nuclei Isolation: Tissue samples are dissociated to release individual nuclei while preserving RNA integrity
-
[8]
Library Preparation: RNA is extracted from isolated nu- clei, and cDNA libraries are generated by reverse tran- scription
Show all 17 references
-
[9]
Sequencing: cDNA libraries are sequenced using high- throughput sequencing platforms, generating millions of short reads corresponding to RNA transcripts
-
[10]
Data Analysis: Bioinformatics tools are used to align se- quencing reads to a reference genome, quantify gene ex- pression levels, and perform downstream analyzes, such as identifying cell types and characterizing gene regula- tory networks
-
[11]
Studying Gene Expression: snRNA− seq provides infor- mation on gene expression heterogeneity within cell pop- ulations and allows researchers to identify rare cell types and transcriptomic changes associated with various bio- logical processes and disease states. By profiling ...
2022
-
[12]
Qubits and Basis States In quantum computing, the basic unit is qubit
Quantum Computing Theory 3.1. Qubits and Basis States In quantum computing, the basic unit is qubit. A qubit can exist in a superposition of the states |0⟩ and|1⟩, represented as (Rieffel and Polak, 2000): |ψ⟩ =α|0⟩ +β|1⟩, (1) whereα andβ are complex numbers such that |α|2 +|β...
2000
-
[13]
(9) • Controlled-RY Gate: The controlled- RY (CRY ) gate rotates around the Y-axis on the target qubit if the control qubit is in the state |1⟩
cos( θ 2) ! . (9) • Controlled-RY Gate: The controlled- RY (CRY ) gate rotates around the Y-axis on the target qubit if the control qubit is in the state |1⟩. This can be constructed using a CX gate and RY gates as follows (Konar et al., 2023e): CRY (θ) = 1 0 0 ...
-
[14]
− sin(θ 2) 0 0 sin( θ
-
[15]
(10) This operation can be decomposed into a sequence involv- ing a CX gate and RYgates: CRY (θ) = (I⊗ RY(θ 2))· CX· (I⊗ RY(−θ 2))· CX
cos( θ 2) . (10) This operation can be decomposed into a sequence involv- ing a CX gate and RYgates: CRY (θ) = (I⊗ RY(θ 2))· CX· (I⊗ RY(−θ 2))· CX . (11) Here, I is the identity matrix. The standard basis state table for controlled rotation gates is provided in ...
-
[16]
Simulations Our work is integrated with Qiskit, an open-source quantum computing library that simulates a noisy quantum circuit using the Aer Simulator backend with default parameters. 5.1. Dataset Processing and Parameter Initialization To construct a Gene Regulatory Network ...
2016
-
[17]
This supports the study conducted by Grub- man et al
Discussions In this article, integrating transcriptomic data, our Alz- QNet o ffers insights into how regulatory mechanisms con- tribute to AD. This supports the study conducted by Grub- man et al. (Grubman et al., 2019) with an emphasis on consid- ering transcriptional and ep...
2019
Reviewed August 6, 2026 · model on record in the stance chip above.
Discussion (0). Continue with ORCID to comment.