Research map / Probabilistic and scientific modeling / Bayesian optimization
De Novo Molecular and Crystal Design with Latent Space Bayesian Optimization
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.
Doctoral consortium thesis summary proposing LCA-LSBO, a latent space Bayesian optimization method that augments VAE training with latent points from promising regions and a latent consistency loss, applied to de novo molecule and crystal design.
Problem
De novo molecule and crystal design via black-box optimization is expensive and high-dimensional; VAEs trained on data are biased toward dense training regions, mismatching BO's need to explore sparse regions, and retraining with queried points needs many costly evaluations.
Approach
Augment VAE retraining with latent points sampled near promising regions (variance below 1) and add a latent consistency loss that penalizes re-encoding deviations, so the VAE learns from synthetic promising points without many extra real evaluations.
Key results
- LCA-LSBO identified a molecule with docking score better than 50,000 randomly sampled molecules within under 350 iterations.
- Crystal-LSBO discovered novel crystal structures for formation energy that reportedly outperform existing VAE and diffusion-based methods (no numbers given in this text).
- A wet-lab synthesis of an optimized molecule is reported as achieved in ongoing work (no details given).
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use LCA-LSBO instead of random sampling | Molecular design · Ion-channel protein docking | docking score | Helped |
| Use Crystal-LSBO with latent consistency-aware optimization | Crystal design · De novo crystal design | formation energy | Helped |
Limitations
- Synthesizability of generated de novo molecules and crystals is hard for experts to assess.
- Crystal and molecule results are summarized without quantitative detail in this extract.
Experimental setup
Benchmarks: docking score for ion-channel protein (molecules) formation energy (crystals)
Metrics: docking score sample efficiency (number of iterations) formation energy
Baselines: VAE-based LSBO weighted retraining diffusion model-based approaches
Builds on: Latent space optimization (Gómez-Bombarelli et al. 2018) weighted retraining (Tripp et al. 2020) metric learning LSBO (Grosnit et al. 2021)
Closest papers by meaning
- A Sober Look at LLMs for Material Discovery: Are They Actually Good for Bayesian Optimization Over Molecules? (ICML 2024 · similarity 0.88)
- Return of the Latent Space COWBOYS: Re-thinking the use of VAEs for Bayesian Optimisation of Structured Spaces (ICML 2025 · similarity 0.86)
- Hybrid Generative AI for De Novo Design of Co-Crystals with Enhanced Tabletability (NeurIPS 2024 · similarity 0.85)
- Multi-Objective Molecular Design Through Learning Latent Pareto Set (AAAI 2025 · similarity 0.85)
- MACS: Multi-Agent Reinforcement Learning for Optimization of Crystal Structures (NeurIPS 2025 · similarity 0.85)
- Navigating the Design Space of Equivariant Diffusion-Based Generative Models for De Novo 3D Molecule Generation (ICLR 2024 · similarity 0.85)
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.