Research map / Probabilistic and scientific modeling / Molecules and proteins

Return of the Latent Space COWBOYS: Re-thinking the use of VAEs for Bayesian Optimisation of Structured Spaces

ICML 2025 · Probabilistic and scientific modeling › Molecules and proteins · approach: Generative latent models for single-cell data · 0 citations (OpenAlex)

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 COWBOYS, a decoupled latent-space-free Bayesian optimisation method that samples from a pretrained VAE conditioned on a structure-space Tanimoto GP via MCMC (pCN) on the latent prior.

Problem

Latent space Bayesian optimisation over molecules suffers from poor GP surrogate alignment with VAE latent spaces (stochastic decoder, offline VAE training) and ill-suited box-shaped search spaces that ignore the Gaussian annulus concentration of VAE latent prior.

Approach

Decouple generative and surrogate models: fit a GP with a Tanimoto kernel directly on molecular structures, and sample new molecules from the pretrained VAE's prior conditioned on the GP probability of improvement over the best observed value, using a deterministic decoder approximation and preconditioned Crank-Nicolson MCMC, with no latent-space GP and no search-space box.

Key results

• On the 25-task PMO-style benchmark after 300 evaluations, COWBOYS achieves the best average score on most tasks, e.g. isomer_c7h8n2o2 1.000±0.00 vs 0.88 for TURBO and 0.864 for RandomLineBO.
• In the low-data SELFIES VAE setting (100 BO steps, batch 5), COWBOYS achieves substantially better sample efficiency than LSBO baselines, including those that fine-tune the VAE, with only InvBO sometimes matching it.
• Ablation shows COWBOYS is robust to MCMC chains/steps unless very small; replacing the structure-space Tanimoto GP with a latent-space GP degrades performance on most tasks.

Measured effects reported in the paper

Surprising findings

• Sampling from the full stochastic decoder instead of the deterministic most-likely decoding significantly hurt performance on 10 of 25 tasks.
• Random-walk-style concerns aside, the Gaussian annulus argument implies the box-clipped latent search space used by standard LSBO excludes the high-prior shell where VAE encodings concentrate.

Experimental setup

Datasets: PMO benchmark tasks (Gao et al.) GAUCHE molecular benchmark tasks (Moss et al.)

Benchmarks: PMO discrete BO benchmark (25 molecular tasks) GAUCHE/SELFIES VAE low-budget molecular design tasks (six tasks)

Metrics: best objective value after evaluations (mean ± s.d./s.e.) sample efficiency

Baselines: LOL-BO CoBO W-LBO InvBO PG-LBO LSBO TURBO-L Graph-GA hill-climbing CMAES GA RandomLineBO

Builds on: latent space bayesian optimisation TURBO GAUCHE preconditioned Crank-Nicolson MCMC

Closest papers by meaning

• Molecule Design by Latent Prompt Transformer (NeurIPS 2024 · similarity 0.87)
• Navigating Chemical Space with Latent Flows (NeurIPS 2024 · similarity 0.86)
• A Sober Look at LLMs for Material Discovery: Are They Actually Good for Bayesian Optimization Over Molecules? (ICML 2024 · similarity 0.86)
• De Novo Molecular and Crystal Design with Latent Space Bayesian Optimization (AAAI 2025 · similarity 0.86)
• DecompOpt: Controllable and Decomposed Diffusion Models for Structure-based Molecular Optimization (ICLR 2024 · similarity 0.86)
• Towards Unified and Lossless Latent Space for 3D Molecular Latent Diffusion Modeling (NeurIPS 2025 · similarity 0.86)