Linear inductive bias boosts extrapolative molecular discovery by Bayesian active learning in the small-data limit
- 1Department of Chemical Engineering
- 2Kyung Hee University
Abstract
Bayesian active learning is widely used for molecular discovery, yet surrogate selection remains challenging when scarce initial data must guide the search beyond the observed property range. Here we benchmark diverse Gaussian process (GP) configurations across ten small-molecule and polymer datasets, focusing on the extrapolative regime where the initial labeled samples are drawn exclusively from below the median of the target property distribution. We identify explicit linear structure as an effective inductive bias for data-scarce extrapolative discovery: linear-kernel GPs provide a strong baseline, while adding linear components to a stationary kernel helps mitigate poor outcomes in the tail of the performance distribution. However, this advantage weakens under later-stage interpolative regime, where adaptive surrogate selection can further improve the discovery performance. These findings support incorporating linear inductive biases into data-driven materials discovery, where data scarcity and limited coverage of the design space are the norm rather than the exception.
Related
- Docking, Molecular Dynamics, QSAR, and Machine Learning in Alzheimer's Small-Molecule Drug Discovery: Current Utility, Key Limitations, and Reproducibility Challenges
- Bgolearn: a unified Bayesian optimization framework for accelerating materials discovery
- Cheminformatics as a core driver of molecular discovery and design
- Integrating artificial intelligence and molecular simulation for peptide discovery: limitations, validation and decision-centered workflows
- A Scoping Review of Diffusion-Based De Novo Molecular Generation: Architectures, Evaluation, and Open Challenges for Computational Drug Discovery
BibTeX
@misc{ha2026linear,
title = {{Linear inductive bias boosts extrapolative molecular discovery by Bayesian active learning in the small-data limit}},
author = {Min Young Ha},
howpublished = {ChemRxiv},
year = {2026},
month = oct,
doi = {10.26434/chemrxiv.15010454/v1},
url = {https://chemrxiv.org/doi/full/10.26434/chemrxiv.15010454/v1},
note = {CC BY},
}
RIS
TY - UNPB
TI - Linear inductive bias boosts extrapolative molecular discovery by Bayesian active learning in the small-data limit
AU - Min Young Ha
PY - 2026
DA - 2026/10/14
PB - ChemRxiv
DO - 10.26434/chemrxiv.15010454/v1
UR - https://chemrxiv.org/doi/full/10.26434/chemrxiv.15010454/v1
AB - Bayesian active learning is widely used for molecular discovery, yet surrogate selection remains challenging when scarce initial data must guide the search beyond the observed property range. Here we benchmark diverse Gaussian process (GP) configurations across ten small-molecule and polymer datasets, focusing on the extrapolative regime where the initial labeled samples are drawn exclusively from below the median of the target property distribution. We identify explicit linear structure as an effective inductive bias for data-scarce extrapolative discovery: linear-kernel GPs provide a strong baseline, while adding linear components to a stationary kernel helps mitigate poor outcomes in the tail of the performance distribution. However, this advantage weakens under later-stage interpolative regime, where adaptive surrogate selection can further improve the discovery performance. These findings support incorporating linear inductive biases into data-driven materials discovery, where data scarcity and limited coverage of the design space are the norm rather than the exception.
ER -