{"id":"07d4de46-0976-4f0f-a19f-482dbaa035c2","arxiv_id":"2607.02749","paper_version":1,"verdict":"CONDITIONAL","confidence":"HIGH","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":5,"one_line_summary":"Hamiltonian-aware Louvain partitioning plus block-restricted full-objective VQE and CUDA-Q multi-QPU batching yields lower lattice-protein energies than fixed-ansatz baselines while remaining competitive on reconstructed RMSD and docking affinity.","lead":"Graph-VQE partitions protein Hamiltonians with Louvain community detection so VQE can optimize weakly coupled qubit blocks in parallel on CUDA-Q multi-QPU simulation, then refines globally. It reports lower energies than standard VQE baselines on small lattice peptides and competitive RMSD/affinity versus AlphaFold3 and IBM hardware runs.","discovery_kind":"new_method","skeptic_critique":{"model":"grok-4.5","headline":"Energy gains rest on a lattice Hamiltonian whose low-energy states may not map to the all-atom structures used for RMSD/affinity claims.","rationale":"The reader already flags the lattice-to-all-atom transfer as the weakest assumption. The stress test simply makes that assumption operational: without a direct check that the reported low-energy bitstrings are both penalty-free and geometrically close to the experimental frames, the energy superiority in Table I cannot be taken as evidence of physically competitive structures. The proposed shot-level penalty audit plus RMSD recomputation is a single, fully determined computation that either confirms or falsifies the mapping. Because the engineering contribution (Hamiltonian-aware multi-QPU parallelism) remains intact regardless of the outcome, the verdict stays CONDITIONAL rather than moving to REJECT; the same conditions the reader already listed (code release, error bars, larger systems) continue to apply.","tokens_in":19869,"tokens_out":604,"duration_ms":5589,"concrete_test":"For each of the ten sequences, decode the highest-probability bitstring of the Graph-VQE optimum, recompute H_penalty alone, and count the fraction of 8192 shots that yield H_penalty=0. Simultaneously compute the Kabsch RMSD of that decoded lattice backbone against the same PDBbind Cα reference used in Table III. If more than ~20 % of shots retain residual penalty energy, or if the zero-penalty conformations still show RMSD >2.5 Å on the CI set, the energy-to-structure link is broken and the strongest claim weakens.","verdict_should_be":"CONDITIONAL","load_bearing_attack":"The central claim equates Graph-VQE’s lower lattice energies (Table I, 54/60 wins) with physically meaningful conformations that are competitive with AlphaFold3 and IBM hardware VQE (Tables II–III). That equivalence rests on the tetrahedral-lattice model of §IV-A: N=2(L-1) qubits, MJ contacts, and uniform penalties λ_chiral=λ_back=λ_overlap=10.0. Under these weights a single geometric violation produces large positive energy, so any optimizer that merely avoids penalties can report large negative energies without recovering native-like packing. Louvain partitioning (Eqs. 9–10) is performed on absolute Pauli coefficients of this same Hamiltonian; the resulting blocks therefore optimize a coarse-grained objective whose ground-state manifold need not coincide with the crystallographic Cα frames used for Kabsch RMSD or the Modeller-relaxed full-atom models used for AutoDock Vina. The paper never reports the fraction of sampled bitstrings that remain penalty-free, nor the energy gap between the reported minima and the true lattice ground state. Consequently the “competitive RMSD/affinity” numbers may reflect post-hoc reconstruction rather than genuine recovery of native contacts by the partitioned VQE.","agreement_with_reader":"agree"},"referee_report":{"model":"grok-4.5","summary":"Graph-VQE extends NVIDIA CUDA-Q with optimization-level parallelism for lattice protein-folding VQE. From the tetrahedral-lattice Hamiltonian (MJ contacts plus uniform geometric penalties), the authors build a weighted qubit interaction graph, partition it by Louvain modularity into weakly coupled blocks, optimize each block by restricted updates on the full-Hamiltonian objective while other parameters are frozen, then perform a short global refinement with CUDA-Q Hamiltonian batching. A Qiskit-to-CUDA-Q translation layer maps EfficientSU2 and operators onto JIT kernels. On ten 5–6-residue peptides, under six entanglement patterns, Graph-VQE (and its no-global ablation) reports the lowest energy in 54/60 settings versus SeqVQE, HBatch, EQC, CVaR-VQE and ADAPT-VQE; multi-GPU scaling preserves energy quality; reconstructed structures yield competitive Kabsch RMSD and AutoDock Vina affinities versus AlphaFold3 and IBM hardware baselines.","tokens_in":20240,"tokens_out":851,"duration_ms":7974,"significance":"If the energy gains and structure metrics hold under tighter validation, the work supplies a practical, Hamiltonian-aware route to optimization-level multi-QPU parallelism that CUDA-Q currently lacks, together with a usable Qiskit bridge. Strengths that should be credited include the systematic multi-entanglement evaluation (Table I), the no-global ablation isolating the refinement stage, the shot-sensitivity study (Fig. 6), the GPU scaling that preserves accuracy (Fig. 8), and the explicit hierarchical schedule (block-restricted full-H updates + batched global refinement). These are concrete engineering contributions for hybrid biomolecular simulation on multi-QPU platforms, even though the present instances remain small lattice peptides.","major_comments":[{"comment":"§IV-A / Eqs. (1)–(3) and Tables II–III: The central claim equates lower lattice energies (Table I) with physically meaningful conformations competitive with AlphaFold3 and IBM hardware VQE. Under uniform penalties λ=10 a single geometric violation produces large positive energy, so large negative energies can be obtained by merely avoiding penalties without recovering native packing. The manuscript never reports the fraction of sampled bitstrings that remain penalty-free, nor the gap of the reported minima to the true lattice ground state. Without those diagnostics the RMSD/affinity numbers (which further rely on Modeller relaxation of a Cα trace) cannot be taken as evidence that the partitioned VQE recovered native contacts rather than post-hoc reconstruction of any valid lattice fold.","section":null},{"comment":"§V-A and Table I: All results are exact state-vector simulations on 5–6-residue peptides (N=8–10 qubits). The abstract and conclusion present Graph-VQE as a path to high-fidelity biomolecular simulations and multi-QPU scaling. The energy advantage is real under the stated protocol, but the load-bearing claim that the Louvain + restricted-update schedule will continue to help once circuit depth, shot noise, and inter-block correlations grow with system size is not yet supported. At minimum the paper should either (i) demonstrate a larger instance or (ii) clearly bound the claim to the lattice-peptide regime and relegate the drug-discovery framing to future work.","section":null},{"comment":"§IV-C / Eqs. (9)–(10) and the free-parameter list: Louvain modularity on absolute Pauli coefficients is presented as biologically informed, yet the only free parameters that control block quality (resolution, S_min merge threshold) and the early-stopping / epoch-patience schedule are fixed without sensitivity analysis. Because the no-global ablation already wins 9/60 cases and loses under certain entanglement topologies, it is unclear how robust the hierarchical schedule is to these choices. A short ablation on S_min / resolution (or an explicit statement that defaults were used throughout) is needed before the partitioning strategy can be treated as a reliable contribution.","section":null}],"minor_comments":[],"recommendation":"major_revision","confidential_remarks":null},"author_rebuttal":null,"desk_editor":{"model":"grok-4.5","letter":"Graph-VQE is a practical CUDA-Q extension that actually improves optimization quality, not just measurement speed. The core idea is simple and new as a package: build the weighted qubit interaction graph from the MJ-derived Pauli coefficients, Louvain-partition into weakly coupled blocks, run concurrent restricted updates of each block’s parameters against the full Hamiltonian, then do a short joint refinement with Hamiltonian batching. The Qiskit-CUDA-Q bridge is a useful engineering extra.\n\nWhat it does well is clear from the tables. Across six entanglement patterns it takes the lowest energy in 54 of 60 settings, the no-global ablation shows the refinement step usually helps, GPU scaling from 1 to 6 devices cuts wall time while holding energy, and the shot-sensitivity curve is reported. It also beats ADAPT-VQE, CVaR-VQE, EQC and plain HBatch under the same shot budget. Negative energies correctly flag valid (penalty-free) lattice conformations, and the reconstructed RMSD/affinity numbers sit in the same ballpark as AlphaFold3 and the IBM hardware baselines after Modeller relaxation.\n\nSoft spots are real but proportionate. Everything is 5–6-residue tetrahedral lattice under exact state-vector simulation; the large energy gaps partly reflect better avoidance of the λ=10 penalties rather than finer native packing. The paper never reports the fraction of penalty-free bitstrings or the gap to the true lattice ground state, so the “competitive with AF3” claim rests on the post-processing pipeline. The abstract’s “highly practical path toward high-fidelity biomolecular simulations” over-reaches. No code is released. Those are the limits of a solid systems paper, not load-bearing flaws in the schedule itself.\n\nThis is for people building multi-QPU VQE stacks or hybrid bio-simulation pipelines who need a concrete, Hamiltonian-aware optimization schedule that demonstrably moves the energy needle. The math and citation pattern are clean; the empirical gains are reproducible from the description. I would send it to peer review, ask for code and dialed-back claims, and expect it to improve. Worth reading if you work on scaling VQE.","headline":"Clean multi-QPU VQE engineering that reliably beats baselines on lattice-peptide energies; structure comparisons are secondary and the systems remain tiny.","tokens_in":20838,"tokens_out":539,"would_cite":false,"duration_ms":19826,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"grok-4.5","headline":"Partitioning a protein Hamiltonian by its interaction graph lets multi-GPU VQE reach lower folding energies than standard methods.","keywords":["Variational Quantum Eigensolver","protein folding","Hamiltonian partitioning","Louvain community detection","multi-QPU parallelization","CUDA-Q","hybrid quantum-classical computing"],"falsifier":"Run the same Graph-VQE pipeline on a larger or all-atom peptide set where the ground-state energy and crystallographic structure are known; if the method no longer reaches lower energy than the listed baselines or yields systematically worse RMSD/affinity than AlphaFold3 under identical shot budgets, the central claim fails.","tokens_in":20772,"feed_emoji":"🧬","tokens_out":655,"duration_ms":6521,"temperature":0.7,"pith_summary":"Graph-VQE claims that protein-folding VQE can be made both more accurate and more parallel by building a weighted qubit interaction graph from the Hamiltonian, detecting weakly coupled communities with Louvain, and optimizing each community’s parameters on the full energy while other parameters stay fixed, then finishing with a short global refinement that batches Hamiltonian terms across GPUs. The authors argue this biologically informed split overcomes the sequential bottleneck of classical optimizers and the barren landscapes that appear when every parameter is updated together. On ten short lattice peptides the method reports the lowest energy in the large majority of settings against SeqVQE, Hamiltonian batching, EQC, CVaR-VQE and ADAPT-VQE, and the reconstructed backbones yield RMSD and docking affinities competitive with AlphaFold3 and IBM hardware baselines. The practical payoff is a Qiskit-compatible, multi-GPU path that isolates algorithmic performance from device noise and scales wall-clock time without degrading solution quality.","feed_headline":"Graph split lets multi-GPU VQE fold peptides to lower energy","feed_subtitle":"Louvain blocks on the interaction graph beat standard VQE and match AlphaFold3 RMSD on lattice proteins","key_machinery":"Hamiltonian-aware Louvain partitioning of the weighted qubit interaction graph (edges weighted by absolute Pauli coefficients derived from Miyazawa–Jernigan contacts), followed by concurrent restricted full-objective block updates and a short global refinement stage that reuses CUDA-Q multi-QPU Hamiltonian batching.","core_discovery":"When the protein Hamiltonian’s interaction graph is partitioned into Louvain communities and each community’s variational parameters are updated under the full Hamiltonian objective while the rest remain frozen, followed by a brief joint refinement with Hamiltonian batching, the optimizer reaches substantially lower final energies than global VQE baselines on the same lattice protein-folding instances, while reconstructed structures remain competitive with AlphaFold3 and hardware baselines.","pith_inferences":[],"forward_implications":[],"fun_headline_variants":["Louvain communities cut VQE energies on protein Hamiltonians","Graph-VQE freezes clusters then refines for lower fold energies","Multi-GPU Louvain split beats global VQE on lattice peptides","Hamiltonian graph partitions deliver lower protein fold energies","CUDA-Q Graph-VQE matches AlphaFold3 RMSD at reduced energy"],"cache_read_input_tokens":16512,"weakest_assumption_plain":"That a coarse tetrahedral-lattice model with fixed contact energies and uniform geometric penalties, once split by community detection on absolute Pauli weights, still produces blocks whose restricted optimization recovers physically meaningful low-energy folds comparable to all-atom and hardware results.","fun_headline_variants_meta":{"raw":{"variants":["Louvain communities cut VQE energies on protein Hamiltonians","Graph-VQE freezes clusters then refines for lower fold energies","Multi-GPU Louvain split beats global VQE on lattice peptides","Hamiltonian graph partitions deliver lower protein fold energies","CUDA-Q Graph-VQE matches AlphaFold3 RMSD at reduced energy"]},"model":"grok-4.5","effort":"low","cost_usd":0.003544,"raw_usage":{"total_tokens":1156,"prompt_tokens":752,"num_sources_used":0,"completion_tokens":71,"cost_in_usd_ticks":35440000,"prompt_tokens_details":{"text_tokens":752,"audio_tokens":0,"image_tokens":0,"cached_tokens":256},"completion_tokens_details":{"audio_tokens":0,"reasoning_tokens":333,"accepted_prediction_tokens":0,"rejected_prediction_tokens":0}},"tokens_in":752,"tokens_out":71,"duration_ms":3342,"temperature":1.0,"reasoning_tokens":333,"cache_read_input_tokens":256,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-07-12T07:17:24.813135+00:00","model_set":{"reader":"grok-4.5"},"falsifier":"Run the same Graph-VQE pipeline on a larger or all-atom peptide set where the ground-state energy and crystallographic structure are known; if the method no longer reaches lower energy than the listed baselines or yields systematically worse RMSD/affinity than AlphaFold3 under identical shot budgets, the central claim fails.","supporting_citations":[],"review_version":1}