srijahn/moderna-challenge

Purpose and execution model

This is a script-driven RNA secondary-structure research project for the WISER × Moderna challenge. It converts short RNA sequences into constrained QUBOs, solves them with simulated QAOA and CVaR-VQE, and evaluates decoded structures against ViennaRNA MFE references. There is no application server; entry points are executable Python workflows, notably statistical_benchmark.py, generate_final_results.py, and compare_to_vienna.py.

Sequence sets and classical reference boundary

Curated folding-positive workloads define the active experiments: benchmark_sequences.py supplies eight 8–14 nt sequences, while test_sequences.py supports smaller validation cases. ViennaRNA is the authoritative thermodynamic reference and candidate evaluator:

• test_vienna.py

• structure_metrics.py

• official_example_benchmark.py

structure_metrics.py calls ViennaRNA MFE folding and evaluates each candidate dot-bracket structure in comparable kcal/mol units, plus base-pair precision/recall/F1, pair distance, and Hamming distance. The 44-nt official example remains classical-only because its QUBO is infeasible for the simulated quantum solvers.

Default pair-indicator QUBO compilation

The central model is rna_to_qubo_full.py, which maps each feasible base pair to a binary variable/qubit. It includes Watson-Crick and G-U wobble candidates, minimum hairpin-loop constraints, and penalties for shared-nucleotide conflicts and crossing pairs (therefore excluding pseudoknots). It also provides brute-force validation for small instances.

• rna_to_qubo_full.py

• rna_to_qubo.py

• qubo_energy.py

This compiler is the boundary between RNA/domain constraints and generic binary optimization. Its simple pair-count-style objective is intentionally distinct from ViennaRNA thermodynamics, so a QUBO optimum may not be an MFE-optimal biological structure.

Quantum solver layer

Two independent PennyLane-based solvers consume the real QUBO, optimize circuit parameters, sample/decode binary assignments, and return structures/energies:

• qaoa_rna_solver.py

• cvar_vqe_rna_solver.py

Both use derivative-free SciPy COBYLA-style optimization rather than statevector backpropagation. QAOA is less scalable/reliable in batch use and is capped near 11 qubits for quality experiments; CVaR-VQE uses a two-local ansatz and is the more broadly run method. Statevector simulation, not quantum hardware, is the material execution runtime.

Evaluation and reproducible result production

The main flow is: curated sequence → QUBO compilation → solver trial(s) → assignment-to-dot-bracket decoding → ViennaRNA metric evaluation → CSV result table under results/.

• statistical_benchmark.py

• batch_accuracy.py

• generate_final_results.py

statistical_benchmark.py is the principal rigorous benchmark: repeated CVaR-VQE and QAOA trials produce trial-level and mean/std datasets. final_summary.py consumes final result and scaling tables for a human-readable aggregate report. results/ is the sole generated-output boundary.

Scaling, noise, and feasibility studies

Separate workflows measure actual simulator resource behavior and robustness rather than extrapolated formulas:

• scaling_analysis_real.py

• cvar_vqe_scaling_analysis.py

• noise_simulation.py

They record qubits, depth/gates, forward runtime, and—only below optimization cutoffs—quality metrics. Noise experiments use PennyLane’s mixed-state simulator with depolarizing channels. Plot scripts consume these CSVs but do not alter solver logic.

Alternative encoding and key design decision

rna_to_qubo_onehot.py implements directed, per-position one-hot pairing variables with consistency penalties. compare_qubo_encodings.py verifies it against the default encoding and ViennaRNA where brute force is feasible. The one-hot scheme uses exactly twice the qubits and more penalty terms while yielding equivalent optima on tested instances, so the pair-indicator encoding remains the production path for solvers and scaling studies.