ICML 2022 | 3dlinker: e (3) equal variation self encoder for molecular link design

Paper title: ICML2022|3DLinker: An E(3) Equivariant Variational Autoencoder for Molecular Linker Design

Paper link: https://arxiv.org/abs/2205.07309

Paper code: https://github.com/GraphPKU/3DLinker

Publication venue: ICML 2022 Long Presentation (118/5630)

Institution: Beijing Institute for General Artificial Intelligence, Tsinghua University, Peking University

Authors: Yinan Huang, Xingang Peng, Jianzhu Ma, Muhan Zhang

 

Research background and motivation

Machine learning plays an increasingly important role in drug design . among , Link design (linker design) It is a widely used drug design method . In link design , We have two called fragment The molecules of , these two items. fragment Generally, it has specific chemical functions ( such as , One fragment It can bind to the target protein , Another one fragment It can attract enzymes used to decompose target proteins ). Our goal is to generate a linker Molecules to link these two fragments, Making it a drug with multiple chemical functions at the same time .

Link design

The link design problem can be regarded as a conditional generation model ： Given two fragments, Construct a ” completion “fragments The probability model of . here drug By inputting fragments And generated linker constitute . Once you have this generation model , Enter any two interested fragments, The model can sample a new sample , In this way, automatic 、 Efficient link design .

fragments and linker It is usually seen as a graph (Graph)： Atoms as vertices , Chemical bonds act as edges . Thus, link design can be equivalent to training the conditional generation model of a graph . however , Retaining only the graph structure will lose 3D geometric information , This may lead to unrealistic 、 Physically unstable linker, Or it can't be directly used in downstream tasks that need three-dimensional coordinates . Previous work [1, 2] Can only be used very roughly fragment Three dimensional information ( For example, manually add distance Information or use an additional convolution network ) And cannot generate linker The space coordinates of .

Link design containing 3D geometric information

In this work , The author proposes a self coding model that can generate three-dimensional coordinates and graphs at the same time ,3DLinker, Used to model the following probability distribution ：

here G Representative diagram ,X Represents spatial coordinates . The most important ,3DLinker It is equivalent to the transformation of coordinate system ： When you type fragments Coordinates undergo any coordinate system transformation  , Entire output drug The coordinates of will change accordingly  , Simultaneous output drug The graph of remains unchanged  . This means that the model can be trained and generalized in any coordinate system . in addition , This task can also be regarded as the pre training of the graph with three-dimensional coordinates ： Randomly mask out some (linker) Graph and coordinates of , Learn how to learn from the rest (fragments) To reconstruct the original three-dimensional image .

 

Method

3DLinker The core of is an equivariant messaging (message passing) modular , be called MF-MP (Mixed-Features Message Passing). For a vertex in the graph  ( In position ), It has invariable characteristics ( Unchanged under the transformation of coordinate system ), It also has equivariant characteristics  ( Transform with the coordinate system ), Where and are channel Count . Every one of channel Is a number independent of the coordinate system , And every one of channel Is the three components of a three-dimensional vector ( Imagine an arrow with a direction ), Will change as the coordinate system changes . In the process of message transmission , These features will aggregate the messages of adjacent vertices while maintaining the invariance and equivariability of the features . Here's one MF-MP The process of message transmission . The detailed calculation process is shown in the paper Methodology part .