QuantumConoMix: Benchmarking Shallow VQCs for Resource-Constrained Peptide Classification

Quantum computing holds the promise of detecting subtle biological signals that classical approaches may overlook. However, practical quantum methods remain limited by hardware constraints. We introduce QuantumConoMix, a dataset of conotoxin peptide sequences enriched with physicochemical descriptors and pretrained ESM embeddings, to evaluate machine learning and quantum classification methods under resource constraints. Conotoxins are short, venom-derived peptides with therapeutic relevance, but they pose challenges for automated classification. We benchmark shallow variational quantum circuits (VQCs) against classical baselines to assess near-term feasibility. Classical models achieve up to 0.97 accuracy, while two-qubit VQCs reach ~0.69 accuracy when combined with embeddings and optimized feature maps. Ablation studies highlight hydrophobicity and charge as key physicochemical features, while embeddings significantly improve performance when included. These results establish QuantumConoMix as a case study in feature design for classical–quantum tradeoffs and provide a benchmark for quantum-enhanced peptide classification.