Calibration and generalizability of probabilistic models on low-data chemical datasets with DIONYSUS

Abstract

A toolkit for the study of the calibration, performance, and generalizability of probabilistic models and molecular featurizations for low-data chemical datasets.

Multi-fidelity Bayesian optimization of covalent organic frameworks for xenon/krypton separations

“…Empirically, GPs tend to be effective surrogate models for Bayesian optimization of molecules in the small-data regime. 109…”

“…, require a small number of examples to learn to make accurate predictions), and (iii) express well-calibrated uncertainty. 109,113,114 The acquisition function for experimental planning must balance exploration, exploitation, and cost. The surrogate model and acquisition function must be cheap to train and evaluate, respectively, relative to the simulations/experiments to evaluate the material property.…”

Real-World Molecular Out-Of-Distribution: Specification and Investigation

“…Severity of the Distribution Shifts. The ultimate severity measure of the training to deployment covariate shift is the gap in performance and uncertainty calibration 17 (see Section 4.6.1). Due to the cyclic nature of the drug discovery process, uncertainty calibration is specifically important during deployment to effectively balance between exploration and exploitation.…”

Bayesian optimisation for additive screening and yield improvements – beyond one-hot encoding

“…Existing research, however, suggests that the application of BO can still help reach promising results even in those scenarios. 49 Despite these challenges, we demonstrate that augmenting BO with adequate reaction representations, initialisation schemes and appropriate surrogate models results in an efficient search towards the best-performing additives in less than 100 evaluations while using as little as ten initialisation reactions.…”