Inspiration

When a novel pathogen emerges, genomic surveillance uncovers thousands of new viral mutations every day, but deep-mutational-scanning (DMS) at the bench takes two to six weeks per variant. Labs are severely bottlenecked not by finding mutations, but by deciding which ones to test first. Existing alignment-based tools (like EVmutation or TranceptEVE) fail in this scenario because a brand-new virus doesn't yet have the deep sequence history needed for alignment. On the other hand, massive protein language models (PLMs) like Meta's ESM have hundreds of millions of parameters but lack an interpretable mechanism to provide a ranked triage shortlist that a lab can act on. We wanted to build a lean, verifiable tool that solves this exact triage problem the day a pathogen appears.

What it does

Q-VIRA is a web platform and API that performs alignment-free viral mutation triage using a Variational Quantum Circuit (VQC). Instead of relying on sequence alignments, it scores amino-acid substitutions based on a handful of residue-level descriptors (like volume, hydropathy, and structural burial). The platform serves two models: SARS-CoV-2 Specialist: A 2-qubit, 41-parameter circuit targeting RBD to ACE-2 binding. Cross-Virus Generalist: A 3-qubit, 63-parameter circuit that spans 17 proteins across 12 viral organisms (like influenza, HIV-1, and dengue).

How we built it

We architected Q-VIRA with a modern, decoupled stack: Frontend: Built with React/Vite and hosted on Vercel. Backend & API: A FastAPI application deployed on Render that serves the model weights and handles data orchestration. Data Pipeline: We pull viral sequences from NCBI, align them to reference genomes using Smith-Waterman, and extract features using Biopython and AlphaFold structures. Quantum Scoring Engine: We used a data-reuploading VQC built in PennyLane. The circuit embeds six features onto three qubits using dense Ry & Rz rotations, feeding into Strongly Entangling Layers. Chemistry Validation: We also built a Variational Quantum Eigensolver (VQE) for molecular hydrogen (H2) running on IBM Qiskit Aer to validate our quantum stack against exact classical methods.

Challenges we ran into

One of our biggest hurdles was avoiding the "leaky split" trap that inflates the metrics of many machine learning models in biology. If you split training and test data randomly, a model can simply memorize a mutation site during training. We had to strictly enforce protein-grouped splits (for the generalist model) and site-grouped splits (for the specialist model) to simulate the true deployment condition of encountering a novel virus. We also ran into feature engineering challenges on small datasets. While adding structural features like AlphaFold residue burial (contacts within 10 Å) massively improved the cross-virus generalist model, adding those exact same features to the smaller SARS-CoV-2 specialist dataset caused immediate overfitting, dropping our AUC from 0.742 to 0.679.

Accomplishments that we're proud of

Beating Meta's ESM: Our 63-parameter, 3-qubit circuit outperformed every variant of Meta's 650-million-parameter ESM family on an alignment-free task, achieving a +0.125 mean AUC advantage over ESM-1v across four held-out proteins Unmatched Clean-Prefix Depth: Because labs work top-down from a triage list, calibration matters more than raw AUC. We achieved a "clean prefix" of 22—meaning 22 correct top-ranked calls before a single false positive—which is 2.4x deeper than the best classical baselines. Quantum Chemistry Verification: Our $H_2$ VQE matched the exact (FCI) energy to < 2 x 10^(-7) Ha, correctly capturing electron correlation where classical Hartree-Fock missed by up to 0.22 Ha (ie approximately equal to 139 kcal/mol) at stretched geometries.

What we learned

We learned that in real-world triage, a well-calibrated output distribution is the most valuable property a model can have, which is exactly where our VQC shined. We also learned the immense predictive power of basic structural descriptors; substituting protein-agnostic AlphaFold residue burial in place of virus-specific distance metrics improved our cross-virus model by +0.0265 AUC across the board. Finally, we learned that honesty in model evaluation—reporting every random seed and utilizing strict, zero-leakage splits—builds far more trust than artificially inflated benchmark numbers.

What's next for BioQubit Labs

Moving forward, we want to scale the model's capabilities while maintaining its alignment-free mandate. Feature Evolution: We plan to upgrade our residue-level descriptors by exploring learned representations, specifically testing ESM-2 embeddings as inputs to the quantum circuit. Complex Interactions: The current pipeline evaluates single-point mutations; our next major milestone is modeling epistasis to understand how multiple concurrent mutations interact. Expanding the Database: We aim to expand the cross-virus generalist model beyond the 17 structure-verified proteins currently supported to create a truly universal first-responder triage tool.

Built With

Share this project:

Updates

Submission history