Molecular Active Learning: How can LLMs Help?
Keywords: Active learning, Bayesian optimization, large language model
Abstract: Drug discovery, and molecular discovery more broadly, can be framed as a sequential active learning problem ---facing a candidate pool, strategies are designed to sequentially acquire molecules to assay, aiming to find the best molecule within the fewest rounds of trial and error.
To automate this process, Bayesian optimization (BO) methods can mimic the approach of human medicinal chemists by constructing \textit{representations} from existing knowledge, quantifying \textit{uncertainty} for the predictions, and designing \textit{acquisition} experiments that balance exploitation and exploration.
Traditionally, these three stages are implemented using building blocks such as graph neural networks (GNN) as representations, variational inference (VI) or Gaussian process (GP) for uncertainty quantification, and analytical expressions as acquisition functions.
To facilitate the integration of both domain-specific and general knowledge into various stages of this process, in this paper, we investigate which parts of this workflow can be augmented or replaced by large language models (LLM).
To this end, we present \textbf{COLT}, a software library for \textbf{C}hemical \textbf{O}ptimization with \textbf{L}anguage- and \textbf{T}opology-based modules, and thoroughly benchmark the combination thereof.
We found that \textit{none} of the LLMs, no matter incorporated at what stage, can outperform the simple and fast Bayesian baseline with GNN and GP.
As a remedy, we offer a new tuning recipe with direct preference optimization (DPO), where the optimization of synthetic properties can be used to increase the efficiency of the acquisition in real-world tasks.
Primary Area: foundation or frontier models, including LLMs
Code Of Ethics: I acknowledge that I and all co-authors of this work have read and commit to adhering to the ICLR Code of Ethics.
Submission Guidelines: I certify that this submission complies with the submission instructions as described on https://iclr.cc/Conferences/2025/AuthorGuide.
Reciprocal Reviewing: I understand the reciprocal reviewing requirement as described on https://iclr.cc/Conferences/2025/CallForPapers. If none of the authors are registered as a reviewer, it may result in a desk rejection at the discretion of the program chairs. To request an exception, please complete this form at https://forms.gle/Huojr6VjkFxiQsUp6.
Anonymous Url: I certify that there is no URL (e.g., github page) that could be used to find authors’ identity.
No Acknowledgement Section: I certify that there is no acknowledgement section in this submission for double blind review.
Submission Number: 13070
Loading