DiscoverBio

Inspiration Antibiotic resistance kills 1.3 million people per year and is on track to surpass cancer as the world's leading cause of death by 2050. Yet developing a new antibiotic the classical way takes 10–15 years and billions of dollars — with a 95% failure rate in clinical trials. Most failures happen because researchers couldn't predict binding strength or resistance early enough.

Quantum computers can simulate molecular interactions at a level of physical accuracy that classical computers fundamentally cannot match. We wanted to bring that capability to a real, usable interface — not a research notebook, but a production pipeline a pharmaceutical researcher could actually run.

What It Does Discover.BIO is a quantum-classical hybrid drug discovery platform. You give it any molecule (as a SMILES string or a common antibiotic name) and it runs a full computational pipeline in under 90 seconds:

3D Geometry — RDKit builds an optimized 3D conformation via ETKDGv3 + MMFF94 VQE Binding Energy — PySCF constructs the Hartree-Fock Hamiltonian; a 4-qubit Variational Quantum Eigensolver (PennyLane + lightning.qubit) finds the ground-state molecular energy in Hartrees — a direct predictor of pathogen binding strength QNN Resistance Prediction — Morgan fingerprints are dimensionality-reduced via PCA(8), then fed through an 8-wire Quantum Neural Network (AngleEmbedding + StronglyEntanglingLayers) trained on 30 known antibiotics to predict resistance probabilities across 5 bacterial strains: MRSA, E. coli, K. pneumoniae, M. tuberculosis, and P. aeruginosa AI Interpretation — Three Backboard AI agents run in parallel: one researches clinical literature on the molecule, one translates raw quantum numbers into plain-English clinical insights, and a third (with persistent session memory) compares multiple candidates head-to-head Results Dashboard — Binding affinity gauge, resistance radar chart, per-strain bar breakdown, AI interpretation, and a 1–10 promise score How We Built It Quantum layer

PennyLane 0.45 with lightning.qubit for fast classical simulation VQE uses a (2e, 2o) active space → 4 qubits, keeping the Hamiltonian tractable for classical simulation while focusing on HOMO/LUMO frontier orbitals QNN: AngleEmbedding + StronglyEntanglingLayers (3 layers × 8 wires × 3 params = 72-parameter weight tensor), trained with Adam optimizer + binary cross-entropy over 50 epochs Chemistry

RDKit for SMILES parsing, 3D conformation generation, and Morgan fingerprint computation PySCF for Hartree-Fock integral computation fed into the PennyLane Hamiltonian builder AI Agents (Backboard SDK)

Agent 1 (MoleculeResearchAgent): literature background, resistance context, clinical status Agent 2 (QuantumResultsInterpreter): translates VQE energy + QNN probabilities into a clinical narrative + promise score Agent 3 (DrugDiscoveryOrchestrator): persistent session memory for multi-molecule comparison Backend

FastAPI on Railway — concurrent asyncio pipeline (research + QNN run in parallel; VQE runs as a separate executor task) In-memory cache for pre-loaded demo molecules (instant results for demos) Frontend

React 18 + TypeScript + Vite Tailwind CSS with custom "QuantaGlass" liquid-glass UI system Recharts for resistance radar; custom SVG binding gauge Challenges We Ran Into RAM limits killed real VQE on Railway. PySCF needs 2GB+ RAM for even small molecules; Railway's free tier is 1GB. We spent hours debugging OOM crashes before adding a USE_REAL_VQE env flag and a deterministic hash-based mock for production. The mock produces consistent results seeded from the SMILES string so the demo is still meaningful.

PennyLane dependency hell. Getting pennylane==0.38.0, pennylane-lightning==0.38.0, autoray<0.8, and rdkit to all coexist on Python 3.12 / Linux required about 15 deployment iterations. rdkit-pypi → rdkit, pennylane-qchem removed (it's bundled), autoray pinned each fix revealed the next breakage.

QNN training on limited data. We only had ~30 labeled antibiotic training samples. We compensated with strong regularization, careful feature engineering (Morgan FP → PCA preserves chemical distance), and evaluating on held-out known antibiotics (penicillin, vancomycin, ciprofloxacin) to validate directionality.

Accomplishments We're Proud Of A real, deployed quantum chemistry pipeline VQE is not a toy; it solves the Schrödinger equation variationally for actual molecular Hamiltonians Sub-90-second end-to-end analysis on molecules no one has ever analyzed before The QNN correctly ranks vancomycin (low MRSA resistance) above penicillin (high MRSA resistance) matching clinical reality A novel candidate, QD-001, emerges from our pipeline with broad-spectrum low resistance and strong binding we didn't hand-tune this; the quantum math produced it What We Learned Variational Quantum Eigensolvers work on real molecules today the math is tractable even on classical hardware with the right active-space approximations Quantum ML (QNNs) trains slowly but generalizes surprisingly well with good feature engineering Deploying quantum chemistry stacks to cloud infra is a genuine systems engineering problem RAM, timeout budgets, and dependency conflicts are as hard as the quantum algorithms AI agents with persistent session memory unlock a genuinely different interaction model for scientific research tools What's Next Real VQE on GPU cloud (replace Railway with a high-RAM instance) for production-grade binding energies Larger QNN training set scrape ChEMBL for labeled antibiotic resistance data across more strains ADMET prediction module absorption, distribution, metabolism, excretion, toxicity filters before a molecule hits wet lab Molecular generation use the QNN as a fitness function for a quantum-enhanced generative model to propose novel antibiotic scaffolds Wet lab validation connect with a university pharmacy lab to synthesize and test QD-001

Built With

• backboard-sdk
• fastapi
• pennylane
• pyscf
• python
• quantum-computing
• quantum-machine-learning
• railway
• rdkit
• react
• tailwindcss
• typescript
• vercel
• vite