Variational low-rank adaptation for uncertainty quantification in molecular property prediction
File(s)
Submitted version
Author(s)
Jayasekera, Shavindra
Sieg, Jochen
Mathea, Miriam
Li, Yingzhen
Type
Journal Article
Abstract
Reliable molecular property prediction is essential for various scientific applications, such as drug discovery and material design. In recent years, machine learning models, such as those based on transformer architectures, have shown promise in accurately predicting molecular properties. However, these models often lack robust uncertainty quantification, in particular in regression settings, which is crucial for informed decision-making in high-stakes scenarios. In this work, we propose Variational LoRA, a novel method for uncertainty quantification in transformer-based molecular property prediction models. Variational LoRA combines the parameter-efficient finetuning technique, Low-Rank Adaptation (LoRA), with a Bayesian framework using variational inference to learn
a posterior distribution over the LoRA weights for transformer-based chemical foundation models. We evaluate our method on several benchmark regression datasets for molecular property prediction and demonstrate that Variational LoRA provides well-calibrated uncertainty estimates while maintaining competitive predictive performance.
a posterior distribution over the LoRA weights for transformer-based chemical foundation models. We evaluate our method on several benchmark regression datasets for molecular property prediction and demonstrate that Variational LoRA provides well-calibrated uncertainty estimates while maintaining competitive predictive performance.
Date Acceptance
2026-08-02
Citation
Journal of Cheminformatics
ISSN
1758-2946
Publisher
BMC
Journal / Book Title
Journal of Cheminformatics
Copyright Statement
Copyright This paper is embargoed until publication. Once published the Version of Record (VoR) will be available on immediate open access.
License URL
Publication Status
Accepted