ReAlignFit: Chemical Induced Fit for Molecular Relational Learning
A new method called Representational Alignment with Chemical Induced Fit (ReAlignFit) has been proposed to improve stability in Molecular Relational Learning (MRL). MRL predicts relationships between molecular pairs by extracting structural features, but aligning substructure representations via attention mechanisms lacks chemical knowledge, causing unstable performance on shifted data. ReAlignFit introduces an Induced Fit-based inductive bias to dynamically align substructure representations. A Bias Correction Function based on substructure edge reconstruction is designed to align representations. The method is theoretically justified and aims to enhance MRL stability across chemical spaces such as functional groups and scaffolds. The paper is available on arXiv with ID 2502.07027.
Key facts
- ReAlignFit stands for Representational Alignment with Chemical Induced Fit.
- It addresses instability in Molecular Relational Learning (MRL).
- MRL predicts relationships between molecular pairs.
- Existing attention mechanisms lack chemical knowledge guidance.
- ReAlignFit introduces Induced Fit-based inductive bias.
- A Bias Correction Function uses substructure edge reconstruction.
- The method targets chemical space shifted data (e.g., functional groups, scaffolds).
- The paper is on arXiv: 2502.07027.
Entities
Institutions
- arXiv