This document proposes a conditional graph information bottleneck (CGIB) approach for molecular relational learning tasks. CGIB aims to extract important subgraphs that maximize the preserved information about a molecule's properties while considering the conditional information from another interacting molecule. The method is tested on datasets for molecular interaction prediction tasks like chromophore-solvent absorption and drug-drug interactions. Results show CGIB outperforms baselines at these tasks while providing interpretable important subgraph detections.

1. Van Thuy Hoang
Network Science Lab
Dept. of Artificial Intelligence
The Catholic University of Korea
E-mail: hoangvanthuy90@gmail.com
2023-10-23
Namkyeong Lee, Dongmin Hyun, Gyoung S. Na,
Sungwon Kim, Junseok Lee, Chanyoung Park

2. 2
X
 X

3. 3
CONTENTS
 Background
 • Molecular Relational Learning
 • Functional Group
 • Information Bottleneck
 Problems
 Conditional Graph Information Bottleneck
 Experiments
 Conclusion

4. 4
MOLECULAR RELATIONAL LEARNING
 Molecular Relational Learning Learning the interaction behavior between a pair
of molecules
 Examples
 Predicting optical properties when a Chromophore and Solvent react
 Predicting solubility when a solute and solvent react
 Predicting side effects when taking two types of drugs simultaneously

5. 5
FUNCTIONAL GROUP
 Functional Group
 Specific atomic groups that play an important role in determining the
chemical reactivity of organic compounds
 Compounds with the same functional group generally have similar
properties and undergo similar chemical reactions
 The hydroxyl group structure has the characteristic of increasing the pol
arity of the molecule

6. 6
FUNCTIONAL GROUP
 The hydroxyl group structure has the characteristic of increasing the pol
arity of the molecule
 Molecule can be represented as a graph
 Functional group can be represented as a subgrap
Recently, information theory-based approaches have be
en proposed to detect important subgraph

7. 7
Information Bottleneck Theory
 Information Bottleneck Theory
 A theoretical approach to trade-off between information compression and
preservation

8. 8
GRAPH INFORMATION BOTTLENECK
 Information Bottleneck Graph
 Subgraph that maximally preserves the property of the original graph
 Motif in ordinary graphs
 Functional group in molecules

9. 9
GRAPH INFORMATION BOTTLENECK
 Extract a subgraph in terms of edges
 Model an edge based Bernoulli distribution to perform graph compression

10. 10
GRAPH INFORMATION BOTTLENECK
 Extract a subgraph in terms of nodes
 Extract a subgraph in terms of nodes Inject noise into node embeddings to
perform graph compression

11. 11
CONDITIONAL GRAPH INFORMATION BOTTLENECK
 X

12. 12
CONDITIONAL GRAPH INFORMATION BOTTLENECK
 X

13. 13
EXPERIMENTS DATASET
 Chromophore dataset
 Absorption max, Emission max, Lifetime
 Solvation Free Energy dataset
 MNSol
 FreeSolv
 CompSol
 Abraham
 CombiSolv
 Drug-Drug Interaction dataset
 ZhangDDI
 ChChMiner

14. 14
EXPERIMENTS MAIN TABLE
 Observations
 Outperforms baselines on both Molecular Interaction / Drug-Drug
Interaction tasks
Performance on Molecular Interaction

15. 15
CONCLUSION
 proposed a method for tackling relation learning tasks, which are crucial for
scientific discovery
 Based on Conditional Information Bottleneck
 It is crucial to consider Graph 2 (Solvent) when detecting the important
subgraph from Graph 1 (Chromophore)
 CGIB has interpretability, which makes it highly practical