[Paper Review] GAUCHE: A Library for Gaussian Processes in Chemistry

Ryan‐Rhys Griffiths, Leo Klarner+25 more|arXiv (Cornell University)|Dec 6, 2022

Mass Spectrometry Techniques and Applications33 citations

TL;DR

GAUCHE provides GPU-enabled Gaussian process kernels for molecular representations (fingerprints, strings, graphs) and supports regression, uncertainty quantification, and Bayesian optimization in chemistry.

ABSTRACT

We introduce GAUCHE, a library for GAUssian processes in CHEmistry. Gaussian processes have long been a cornerstone of probabilistic machine learning, affording particular advantages for uncertainty quantification and Bayesian optimisation. Extending Gaussian processes to chemical representations, however, is nontrivial, necessitating kernels defined over structured inputs such as graphs, strings and bit vectors. By defining such kernels in GAUCHE, we seek to open the door to powerful tools for uncertainty quantification and Bayesian optimisation in chemistry. Motivated by scenarios frequently encountered in experimental chemistry, we showcase applications for GAUCHE in molecular discovery and chemical reaction optimisation. The codebase is made available at https://github.com/leojklarner/gauche

Motivation & Objective

• Enable probabilistic modeling for molecules and chemical reactions using Gaussian processes.
• Provide GPU-backed kernels for structured molecular inputs (graphs, strings, fingerprints).
• Interface GP/BO tooling with existing libraries (GPyTorch, BoTorch, RDKit) to facilitate uncertainty quantification and optimization in chemistry.
• Evaluate regression, UQ, and Bayesian optimization performance across diverse molecular datasets.

Proposed method

• Implement GPU-based, kernel-based GP regression for molecular representations (fingerprints, fragments, SMILES/SELFIES strings, graphs).
• Incorporate string, fingerprint, and graph kernels, including Tanimoto, SMILES string kernel, and GraKel graph kernels, with learnable hyperparameters via marginal likelihood.
• Integrate with GPyTorch and BoTorch for scalable GP training and Bayesian optimization workflows.
• Benchmark on molecular property prediction datasets (Photoswitch, ESOL, FreeSolv, Lipophilicity) and reaction yields (Buchwald–Hartwig, Suzuki–Miyaura).
• Assess uncertainty quantification using NLPD and related metrics and compare against Bayesian NNs and deep ensembles.

Experimental results

Research questions

• RQ1 Which molecular representations (fingerprints, strings, graphs) and corresponding kernels yield the best regression accuracy across diverse chemistry tasks?
• RQ2 How well do GAUCHE's kernels provide calibrated uncertainty estimates for molecular data?
• RQ3 Can GP-based Bayesian optimization using GAUCHE kernels outperform random search on ultra-low data molecular discovery tasks?
• RQ4 How do graph, string, and fingerprint kernels compare in BO performance for photoswitches and reaction yields?
• RQ5 What is the effect of representation-task coupling on BO and UQ performance in chemical spaces?

Key findings

• GP regression performance is task-dependent across datasets and representations.
• String and graph kernels can outperform fingerprint-based kernels in certain property prediction tasks.
• Uncertainty estimates (NLPD) generally correlate with regression accuracy across datasets.
• Bayesian optimization using GAUCHE kernels outperforms random search in ultra-low data regimes (e.g., photoswitch dataset).
• BO performance aligns with regression/UQ quality: better regression/UQ can lead to stronger BO performance on targeted chemistry tasks.