Research map / Probabilistic and scientific modeling / Variational and generative sampling
QVAE-Mole: The Quantum VAE with Spherical Latent Variable Learning for 3-D Molecule Generation
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Proposes QVAE-Mole, a fully quantum VAE with amplitude-encoded 3-D molecules and a von Mises-Fisher latent space, plus a conditional variant QCVAE-Mole, evaluated by simulation on QM9.
Problem
Generating 3-D molecules with quantum machine learning models on NISQ hardware; prior quantum molecule generators (QGAN-HG, SQ-VAE) work on molecular graphs or use hybrid classical parameters and normal latent spaces that clash with quantum state normalization.
Approach
Amplitude-encode normalized 3-D atom coordinates, atom types and an auxiliary value into a quantum state; use a fully quantum encoder/decoder with partial trace as latent bottleneck and a vMF hyperspherical latent space matching the unit norm of quantum states; encode property conditions via angle-encoded condition qubits and condition layers.
Key results
- QVAE-Mole reaches 78.13% validity, 27.39% Unique×Valid and 57.38% Novel×Valid on 10K random QM9 molecules, outperforming quantum/hybrid baselines SQ-VAE and QGAN-HG.
- QVAE-Mole uses 224 quantum parameters and 0 classic parameters versus 3.1M for G-SphereNet and 5.3M for EDM, while trailing EDM (91.93% validity).
- In single-condition generation, QCVAE-Mole raises the fraction of molecules matching the target logP=1.0 condition from 2.6% to 45.6%.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| use QVAE-Mole instead of MLP-VAE | QM9 3-D random molecule generation · QM9 | Valid: 51.26% ± 1.1 -> 78.13% ± 0.6 | Helped |
| use QVAE-Mole instead of EDM | QM9 3-D random molecule generation · QM9 | Valid: 91.93% -> 78.13% ± 0.6 | Hurt |
| use QVAE-Mole instead of SQ-VAE | QM9 molecular generation · QM9 | Valid: 44.23% ± 1.0 -> 78.13% ± 0.6 | Helped |
| use QVAE-Mole instead of QGAN-HG | QM9 molecular generation · QM9 | Novel×Valid: 18.48% ± 1.0 -> 57.38% ± 0.9 | Helped |
| add conditional property encoding | QVAE-Mole · QM9 single-condition generation | condition-match percentage: 0.4: 52.5 -> 75.2 | Helped |
| add conditional property encoding | QVAE-Mole · QM9 single-condition generation | condition-match percentage: logP=1.0: 2.6 -> 45.6 | Helped |
| add conditional property encoding | QVAE-Mole · QM9 multi-condition generation | property distribution alignment | Helped |
| replace vMF latent distribution with a normal distribution (ablation) | QVAE-Mole · QM9 random generation | Valid | Hurt |
| replace vMF latent distribution with a normal distribution (ablation) | QVAE-Mole · QM9 random generation | Unique×Valid | Mixed |
| replace normal distribution with vMF distribution (ablation) | MLP-VAE · QM9 random generation | generation performance | Hurt |
| use fidelity loss instead of classic reconstruction loss (ablation) | QVAE-Mole · QM9 random generation | Novel×Valid: 57.38% ± 0.9 -> 31.50% ± 0.7 | Hurt |
| use fidelity loss instead of classic reconstruction loss (ablation) | QVAE-Mole · QM9 random generation | Valid: 78.13% ± 0.6 -> 74.39% ± 0.8 | Hurt |
Surprising findings
- Normal-distribution latent space performs better in classic VAE than vMF, contrary to the expectation that vMF suits the quantum setting.
- Smaller vMF concentration kappa improves validity but reduces uniqueness and novelty, a validity-diversity trade-off.
Limitations
- x
Experimental setup
Datasets: QM9
Benchmarks: QM9
Metrics: Valid Unique×Valid Novel×Valid Avg. MMD of bond lengths condition match percentage
Baselines: MLP-VAE E-NFs G-SchNet G-SphereNet EDM SQ-VAE QGAN-HG P2-QGAN-HG
Builds on: VAE SQ-VAE QGAN-HG EDM
Closest papers by meaning
- MolSpectra: Pre-training 3D Molecular Representation with Multi-modal Energy Spectra (ICLR 2025 · similarity 0.84)
- TITAN: A Trajectory-Informed Technique for Adaptive Parameter Freezing in Large-Scale VQE (NeurIPS 2025 · similarity 0.83)
- Learning Disentangled Equivariant Representation for Explicitly Controllable 3D Molecule Generation (AAAI 2025 · similarity 0.83)
- Symphony: Symmetry-Equivariant Point-Centered Spherical Harmonics for 3D Molecule Generation (ICLR 2024 · similarity 0.83)
- PALQO: Physics-informed model for Accelerating Large-scale Quantum Optimization (NeurIPS 2025 · similarity 0.83)
- Accelerating 3D Molecule Generation via Jointly Geometric Optimal Transport (ICLR 2025 · similarity 0.83)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.