Kim Hyun Woo
Network Science Lab
The Catholic University of Korea
E-mail :
kimwoohyun0622@gmail.com
Conditional Graph Information
Bottleneck for Molecular
Relational Learning
2
 Introduction
• Recap: Graph Convolutional Networks (GCNs)
• Recap: Graph data as molecules
• Recap: Information bottleneck principle
• Molecular Relational Learning
 Method
• Overview
• Architecture
• Iterative Substructure Extractor
• Interactive Graph Information Bottleneck
 Experiment
 Conclusion
3
Recap: Graph Convolutional Networks (GCNs)
• Key Idea: Each node aggregates information from its neighborhood to get
contextualized node embedding.
• Limitation: Most GNNs focus on homogeneous graph.
Neural
Transformation
Aggregate neighbor’s
information
GNN
4
Recap: Graph data as molecules
• Molecules can be naturally represented as graphs with their atoms as nodes and
chemical bonds as edges.Graph data such as molecules and polymers are found to have
attractive properties in drug and material discovery
• Molecules as graphs
Molecular Learning
5
Recap: Information bottleneck principle
• In perspective of Information bottleneck
principle, Deep learning model can be
defined by extraction and compression
phase.
• early phase of epoch, DL extract
information that related with target Y
from input X
• Late phase of epoch, DL compress
information that is not related with Y
from input X
• It can be summurized
• “DL learning is process that removing
un-related information with target Y
from input X”
Information Bottleneck
6
Molecular Relational Learning
• When molecule interact with other molecule. Not all of the chemical structure participate
in a reaction. but Few substructure make a reaction and their own property. And we call
this substructure, Functional group.
• But previous work, a lots of paper utilizing all of molecule structure to predict reaction.
Because of that, previous model have poor generalizablilty and performance.
Molecular Relational Learning
7
Molecular Relational Learning
Rethink what is the “Condition”
8
Molecular Relational Learning
Rethink what is the “Condition”
9
Molecular Relational Learning
Summurization of the Condition
• We mentioned three Category of the condition(Interaction)
• 1. No Condition
• 2. Condition as entire molecular structure
• 3. Condition as sub-structure
• The Proposed method state “when only a substructure interact, then we should make
sub-strucrue condition”
10
Overview
Method
11
Molecular Graph Encoder
Method
• To modeling Molecular representation, they choose
CGIB manner, which represent feature interaction.
• Process follow three steps
• First, the encode Molecule by GNN.
• Sencond, calculate similarity score between Atoms
in each molecule
• Third, based on Similarity score, make intracted
Representation
12
Iterative Substructure Extractor
Method
• The Proposed method state “when only a substructure interact, then
we should make substructure condition”
• In here they Suggest ISE, finding process Sub-structure Condition.
• The key point is T time interaction.
• First, they calculate similarity Score between Full Graph(H_2j)
and Substructure Condition(previous Step Substructure)
• Second, Based on Representation they calculate Preserve
probablity
13
Interactive Graph Information Bottleneck
Method
• The Proposed method state “when only a substructure
interact, then we should make substructure condition”
• In here, there are Three IB term
14
Interactive Graph Information Bottleneck
Method
15
Performance
Experiment
16
Hyper-Parameter Sensitivity
Experiment
17
Interpreterbility
Experiment
18
QA

251229_HW_LabSeminar[Conditional Graph Information Bottleneck for Molecular Relational Learning].pptx

  • 1.
    Kim Hyun Woo NetworkScience Lab The Catholic University of Korea E-mail : kimwoohyun0622@gmail.com Conditional Graph Information Bottleneck for Molecular Relational Learning
  • 2.
    2  Introduction • Recap:Graph Convolutional Networks (GCNs) • Recap: Graph data as molecules • Recap: Information bottleneck principle • Molecular Relational Learning  Method • Overview • Architecture • Iterative Substructure Extractor • Interactive Graph Information Bottleneck  Experiment  Conclusion
  • 3.
    3 Recap: Graph ConvolutionalNetworks (GCNs) • Key Idea: Each node aggregates information from its neighborhood to get contextualized node embedding. • Limitation: Most GNNs focus on homogeneous graph. Neural Transformation Aggregate neighbor’s information GNN
  • 4.
    4 Recap: Graph dataas molecules • Molecules can be naturally represented as graphs with their atoms as nodes and chemical bonds as edges.Graph data such as molecules and polymers are found to have attractive properties in drug and material discovery • Molecules as graphs Molecular Learning
  • 5.
    5 Recap: Information bottleneckprinciple • In perspective of Information bottleneck principle, Deep learning model can be defined by extraction and compression phase. • early phase of epoch, DL extract information that related with target Y from input X • Late phase of epoch, DL compress information that is not related with Y from input X • It can be summurized • “DL learning is process that removing un-related information with target Y from input X” Information Bottleneck
  • 6.
    6 Molecular Relational Learning •When molecule interact with other molecule. Not all of the chemical structure participate in a reaction. but Few substructure make a reaction and their own property. And we call this substructure, Functional group. • But previous work, a lots of paper utilizing all of molecule structure to predict reaction. Because of that, previous model have poor generalizablilty and performance. Molecular Relational Learning
  • 7.
    7 Molecular Relational Learning Rethinkwhat is the “Condition”
  • 8.
    8 Molecular Relational Learning Rethinkwhat is the “Condition”
  • 9.
    9 Molecular Relational Learning Summurizationof the Condition • We mentioned three Category of the condition(Interaction) • 1. No Condition • 2. Condition as entire molecular structure • 3. Condition as sub-structure • The Proposed method state “when only a substructure interact, then we should make sub-strucrue condition”
  • 10.
  • 11.
    11 Molecular Graph Encoder Method •To modeling Molecular representation, they choose CGIB manner, which represent feature interaction. • Process follow three steps • First, the encode Molecule by GNN. • Sencond, calculate similarity score between Atoms in each molecule • Third, based on Similarity score, make intracted Representation
  • 12.
    12 Iterative Substructure Extractor Method •The Proposed method state “when only a substructure interact, then we should make substructure condition” • In here they Suggest ISE, finding process Sub-structure Condition. • The key point is T time interaction. • First, they calculate similarity Score between Full Graph(H_2j) and Substructure Condition(previous Step Substructure) • Second, Based on Representation they calculate Preserve probablity
  • 13.
    13 Interactive Graph InformationBottleneck Method • The Proposed method state “when only a substructure interact, then we should make substructure condition” • In here, there are Three IB term
  • 14.
  • 15.
  • 16.
  • 17.
  • 18.