PRISM: Prior Relational Information for Self-supervised Modeling to Enhance Solubility OOD Generalization

Key points

• Solubility depends on intermolecular interactions between solutes and solvents, and is central to drug discovery, chemical and materials process engineering.1
• Learning chemically meaningful relational information without labels can alleviate data scarcity and improve OOD generalization, but remains challenging.1
• PRISM is a label-free pretraining framework that converts expert-defined chemical priors into relational teachers.1
• For solubility prediction, it trains four encoders to preserve the population-level relational information from priors capturing hydrophobic, electrostatic, polarizability, and substructural effects, using Kullback-Leibler (KL) divergence to match distance-based neighborhood distributions during pretraining.1

Sources (1)

• [1]PRISM: Prior Relational Information for Self-supervised Modeling to Enhance Solubility OOD Generalization
  arXiv (AI, ML, NLP, CV, robotics, multi-agent) · Oct 19, 12:18 AM

  We introduce Prior Relational Information for Self-supervised Modeling (PRISM) and apply it to solubility OOD generalization.

  Solubility depends on intermolecular interactions between solutes and solvents, and is central to drug discovery, chemical and materials process engineering.

Extractive summary: sentences quoted from the sources.