MoonNote

GeomGCL: Geometric Graph Contrastive Learning for Molecular Property Prediction 본문

Study/Paper Review

GeomGCL: Geometric Graph Contrastive Learning for Molecular Property Prediction

Kisung Moon 2021. 12. 15. 12:07

3줄 요약

- 2D molecular graph와 3D graph를 활용하여 contrastive learning

- distance와 angle 정보를 2D, 3D 관점에서 모두 활용

- 기존 geometric GNNs들과 성능 차이가 꽤 많이 남(SGCN, MAT, DimeNet 등)

 

Methods

1. Geometric Embedding

- local distance(2D), 2D angle, global distance(3D), 3D angle을 구하기 위해 선행 연구에서 사용된 radial basis function(RBF)를 사용하였다.

 

2. Adaptive Geometric Message Passing

(i) Node -> Edge Message Passing

- edge를 구성하는 pairwise node의 embedding과 bond feature(2D), global distance(3D)[둘다 edge 정보]를 concat

 

(ii) Edge -> Edge Message Passing

- (2D) target edge의 embedding과 neighbors edge와의 2D angle embedding을 element-wise 연산 후 sum

- (3D) target edge의 embedding과 neighbors edge와의 3D angle embedding을 element-wise 연산 후 sum

- (3D) 여러 개의 angle domain을 적용한 결과를 concat후 hierachical aggregation

 

(iii) Edge -> Node Message Passing

- target node의 neighbors edge embedding과 distance embedding 각각을 element-wise 연산 후 sum

 

(iv) Node -> Graph Attentive Pooling

 

3. Geometric Contrastive Optimization

- local한 spatial correlation을 반영하기 위해 spatial regularizaion을 사용

- spatial loss + contrastive loss를 함께 사용

 

 

4. Downstream

- 2d encoder와 3d encoder의 결과를 sum

 

Result

- 기존 방법들과 꽤 많은 성능 차이가 남

원문

https://arxiv.org/abs/2109.11730

 

 

Comments