![[2602.16684] Retrieval-Augmented Foundation Models for Matched Molecular Pair Transformations to Recapitulate Medicinal Chemistry Intuition](https://arxiv.org/static/browse/0.3.4/images/arxiv-logo-fb.png){kind=logo}
[2602.16684] Retrieval-Augmented Foundation Models for Matched Molecular Pair Transformations to Recapitulate Medicinal Chemistry Intuition
Summary
This article presents a novel approach using retrieval-augmented foundation models for matched molecular pair transformations, enhancing medicinal chemistry intuition through improved analog generation.
Why It Matters
The research addresses limitations in current machine learning methods for medicinal chemistry by introducing a framework that enhances the controllability and diversity of molecular transformations. This advancement could significantly impact drug discovery and design processes, making it relevant for both researchers and practitioners in the field.
Key Takeaways
- Introduces a variable-to-variable formulation for analog generation.
- Develops prompting mechanisms for user-specified transformation patterns.
- Implements a retrieval-augmented framework for contextual guidance.
- Demonstrates improved diversity and novelty in molecular structures.
- Validates effectiveness through experiments on chemical corpora and patents.
Computer Science > Machine Learning arXiv:2602.16684 (cs) [Submitted on 18 Feb 2026] Title:Retrieval-Augmented Foundation Models for Matched Molecular Pair Transformations to Recapitulate Medicinal Chemistry Intuition Authors:Bo Pan, Peter Zhiping Zhang, Hao-Wei Pang, Alex Zhu, Xiang Yu, Liying Zhang, Liang Zhao View a PDF of the paper titled Retrieval-Augmented Foundation Models for Matched Molecular Pair Transformations to Recapitulate Medicinal Chemistry Intuition, by Bo Pan and 6 other authors View PDF HTML (experimental) Abstract:Matched molecular pairs (MMPs) capture the local chemical edits that medicinal chemists routinely use to design analogs, but existing ML approaches either operate at the whole-molecule level with limited edit controllability or learn MMP-style edits from restricted settings and small models. We propose a variable-to-variable formulation of analog generation and train a foundation model on large-scale MMP transformations (MMPTs) to generate diverse variables conditioned on an input variable. To enable practical control, we develop prompting mechanisms that let the users specify preferred transformation patterns during generation. We further introduce MMPT-RAG, a retrieval-augmented framework that uses external reference analogs as contextual guidance to steer generation and generalize from project-specific series. Experiments on general chemical corpora and patent-specific datasets demonstrate improved diversity, novelty, and controllability, ...
