【GNN报告】腾讯AI lab 徐挺洋：图生成模型及其在分子生成中的应用

目录

1、简介

2、An overview of Graph Generative Models and Their Applications on Molecular Generation

背景

 图生成学习范式

从VAE角度的工作

VGAE

 GraphVAE

 JTVAE​编辑

​编辑 Scaffold Hopping骨架跃迁

 GAN方法角度工作

 MoIGAN​编辑

 Flow角度工作

 MoFlow

Glow

GraphDF 

Diffusion角度工作

 GraphEBM

 EDM

小结

3、参考

---

1、简介

报告嘉宾：徐挺洋（腾讯AI lab）

报告题目

An overview of Graph Generative Models and Their Applications on Molecular Generation（图生成模型及其在分子生成中的应用）

报告摘要

Generating novel graph structures is a fundamental problem for studies in drug discovery, engineering, and social networks. Although a set of different generative models, such as VAE, GAN, Normalizing Flow, and diffusion models, have achieved great successes in image processing, it is still hard to naively apply such generative models directly to graph structures due to the complex distribution, non-unique nature of graph and non-local dependencies between nodes and edges in a graph. In this report, we will review how above generative models are employed into graph generations and their proper applications on the task of molecular generation. We will introduce VGAE, JTVAE, and GraphGMVAE as VAE-based methods; MolGAN and MolAICal as GAN-based methods; MoFlow and GraphDF as Flow-based methods. We will also discuss the current challenges for diffusion models in graph generation.

报告人简介

Tingyang Xu is a Senior researcher of Machine Learning Center in Tencent AI Lab. He obtained the Ph.D. degree from The University of Connecticut in 2017 and the bachelor’s degree from Shanghai Jiaotong University. In Tencent AI Lab, he is working on deep graph learning, graph generations and applying the deep graph learning model to various applications, such as molecular generation and rumor detection. His main research interests include social network analysis, graph neural networks, and graph generations, with particular focus on designing deep and complex graph learning models for molecular generations. He has published several papers on data mining, machine learning top conferences KDD, WWW, NeurIPS, ICLR, CVPR, ICML, etc.

2、An overview of Graph Generative Models and Their Applications on Molecular Generation

背景

 在图像和语音上生成模型很成熟

 GAN图生成意义和学习过程

在分子图、控制流和知识图谱上比较有意义

 在CV领域上生成模型上经典方法

GAN：对抗角度，将真假数据一起放入辨别器，之和用生成器生成假数据

VAE：编码解码的变分角度，拟合后重构损失，一般假设高斯分布

Flow:双射函数fx

Diffusion：通过先验分布，逐步加Noise，之和仔逆向推回之前无噪声的x_0

 在图上哪类方法更适合？

前两种显然适合

 图因为有邻接矩阵，对阵，又是01矩阵，后两种方法难以直接使用

 图生成学习范式

1、得到一个隐空间z后，通过一个网络直接将整个A的矩阵生成出来

2、像生成序列一样逐步生成

 

从VAE角度的工作

VGAE

思想：将隐空间z和其转置相乘可以得到的非01的矩阵，后sigmoid，最后重构损失

不足：不满足置换不变

 GraphVAE

不是简单的计算损失，使用了了一个相似度矩阵

 

 分子生成

衡量标准：

1）合法性

2）唯一性

3）新颖性

常见数据集：

1）QM9

2）ZINC

 JTVAE

 先来了解另一篇工作

step by step做的

要不要加点？加点后要不要再该店上加边？是否需要额外再加一个点？在两个点的基础上这两个点间要不要加边

 JTVAE不同于一个一个点做，而变成一个结构一个结构的生成

 树结构

 框架

虽然耗时，但是合法性和唯一性以及创新性都很棒

 Scaffold Hopping骨架跃迁

在隐空间在做不同假设

 高斯混合分布

 在红色部分也就是簇附近附近采骨架点，甚至更远 区域采

 效果对比

活性和结构相似度这两块分析

 GAN方法角度工作

因为效果不是很好，就只介绍MoIGAN

 MoIGAN

结果对比

合法性高，唯一性和新颖性低

 Flow角度工作

 MoFlow

重构上有优势

 

Glow

 MoFlow

边矩阵计算

 实验结果对比

GraphDF 

 

 就是加点/边的过程

训练过程

 

实验对比

Diffusion角度工作

 GraphEBM

逐渐加噪声，而后去噪的一个学习过程，基本在高斯噪声

 EDM

 

实验对比

 

小结

 

 

3、参考

LOGS 第2022/07/23期 || 腾讯AI lab 徐挺洋：图生成模型及其在分子生成中的应用_哔哩哔哩_bilibili

 LOGS 第2022/07/23期 || 腾讯AI lab 徐挺洋：图生成模型及其在分子生成中的应用 (qq.com)

参考文献

1. Yu, Y., Xu, T., Li, J., Qiu, Y., Rong, Y., Gong, Z., ... & Huang, J. (2021). A novel scalarized scaffold hopping algorithm with graph-based variational autoencoder for discovery of JAK1 inhibitors.ACS omega, 6(35), 22945-22954.

2. Bai, Q., Tan, S., Xu, T., Liu, H., Huang, J., & Yao, X. (2021). MolAICal: a soft tool for 3D drug design of protein targets by artificial intelligence and classical algorithm.Briefings in bioinformatics, 22(3), bbaa161.

3. Bai, Q., Liu, S., Tian, Y., Xu, T., Banegas‐Luna, A. J., Pérez‐Sánchez, H., ... & Yao, X. (2022). Application advances of deep learning methods for de novo drug design and molecular dynamics simulation. Wiley Interdisciplinary Reviews: Computational Molecular Science, 12(3), e1581.