Virtual screening 5 - GCN

Graph Convolutional Networks

- 비정형화된 구조에 사용가능.

 

System

- Structure : Representation, Computation.

  - Entity : element, size, mass, ...

  - Relation : property btw entities.

  - Rule : relational inductive biases, structure를 바탕으로 DL 설계.

 

Graph representation

$$ Graph = G(X,A) $$

X : Node, Vertex

- Atoms in a molecule

A : Adjacency matrix

- Edges of a graph

- Connectivity, relationship

 

Molecular graphs

X = [ Atome type , # of Hs , # Valence , Aromaticity ]^T

$$ H_i^{l+1} = \sigma (AH^l W^l + b^l)$$

 

Permutation invariant

- Node-wise summation

$$ Z_G = \tau \left( \sum_{i\in G}MLP(H_i^l) \right) $$

- Graph gathering

$$ Z_G = \tau \left(\sum \sigma (MLP_1 (H_i^L , H_i^0)) \odot MLP_2(H_i^L) \right) $$

 

𝜏 : ReLU activation

σ : sigmoid activation

⊙ : element-wise matrix multiplication (Hadamard product0)

 

Attention, α

- Deeply learning molecular structure-property relationships using attention- and gate-augmented graph convolutional network, https://arxiv.org/abs/1805.10988

- Relation의 강도를 다르게 하기 위해 weight를 주는 것.

$$ H^{l+1} = \sigma\left( \sum_{j\in N(i)} \alpha_{ij} H_j^l W^l \right) $$

- GCN에 attention을 더해줌으로써 local feature를 더 잘 얻어낼 수 있다.

 

Classification (active vs. inactive)

- Predicting Drug–Target Interaction Using a Novel Graph Neural Network with 3D Structure-Embedded Graph Representation, https://doi.org/10.1021/acs.jcim.9b00387

- Pose prediction : PDBbind v.2018

- Virtual screening : DUD-E & MUV

 

Regression (binding affinity scoring)

- PotentialNet for Molecular Property Prediction, https://doi.org/10.1021/acscentsci.8b00507

- Gated recurrent unit

- PDBbind v.2007

- Refined train, Core test

 

Pros

- Protein과 ligand atom의 bond를 explicit하게 표현할 수 있음.

- Protein과 ligand를 compact하게 표현할 수 있음. (atom이 있냐, 이것들이 연결되어있는지 표현가능)

- Rotational 및 translational invariant함.

 

Cons

- 3차원 구조를 직관적으로 표현하기 어려움.

 

 

 

Reference

- GCN for virtual screening l 김우연, https://youtu.be/eCr1gIy2KDs