Property-driven localization and characterization in deep molecular representations.

Area of Science:

• Computational Chemistry
• Machine Learning
• Cheminformatics

Background:

• Deep learning models excel at learning molecular representations for structure-property relationships.
• Interpreting these complex representations remains a challenge in molecular science.
• Understanding learned features is crucial for reliable molecular property prediction and design.

Purpose of the Study:

• To develop an unsupervised method for localizing and characterizing property-driven elements within pre-trained molecular models.
• To enhance the interpretability of deep molecular representations.
• To evaluate the utility of extracted features for downstream tasks.

Main Methods:

• Utilized non-parametric property-driven subset scanning (PDSS) to analyze representations from chemical language and graph generative models.
• Assessed detection capabilities on diverse molecular benchmarks including ZINC-250K, MOSES, MoleculeNet, FlavorDB, and M2OR.
• Investigated representation evolution during domain adaptation and evaluated extracted elements for dimensionality reduction.

Main Results:

• Discovered significant information condensation in pre-trained embeddings after task-specific fine-tuning.
• Identified a high degree of task-specific unique property-driven elements within molecular embeddings.
• Demonstrated that extracted property-driven elements can serve as effective lower-dimension representations, maintaining or improving performance on new tasks.

Conclusions:

• The proposed PDSS method effectively localizes and characterizes informative features in deep molecular representations.
• Task-specific fine-tuning leads to specialized feature localization in molecular embeddings.
• Using discovered property-driven elements as features offers a powerful dimensionality reduction strategy for molecular property prediction without requiring model retraining.