Learning Gradient Fields for Molecular Conformation Generation

By Chence Shi, Shitong Luo, Minkai Xu, Jian Tang. 2021

261 citations recorded by the source corpus.

We study a fundamental problem in computational chemistry known as molecular conformation generation, trying to predict stable 3D structures from 2D molecular graphs. Existing machine learning approaches usually first predict distances between atoms and then generate a 3D structure satisfying the distances, where noise in predicted distances may induce extra errors during 3D coordinate generation. Inspired by the traditional force field methods for molecular dynamics simulation, in this paper, we propose a novel approach called ConfGF by directly estimating the gradient fields of the log density of atomic coordinates. The estimated gradient fields allow directly generating stable conformations via Langevin dynamics. However, the problem is very challenging as the gradient fields are roto-translation equivariant. We notice that estimating the gradient fields of atomic coordinates can be translated to estimating the gradient fields of interatomic distances, and hence develop a novel algorithm based on recent score-based generative models to effectively estimate these gradients. Experimental results across multiple tasks show that ConfGF outperforms previous state-of-the-art baselines by a significant margin.

Subfields: Materials Chemistry, Molecular Biology, Spectroscopy, Generative Models, Computational Chemistry & Drug Discovery

How researchers see this paper

The researchers whose trusted networks rank it highest, strongest advocate first. Every researcher gets a different order.

In Materials Chemistry

In Molecular Biology

In Spectroscopy

In Generative Models

In Computational Chemistry & Drug Discovery

Citations in the corpus

TANKBind: Trigonometry-Aware Neural NetworKs for Drug-Protein Binding Structure Prediction (2022, background): “…structure modeling Jumper et al. [2021], Baek et al. [2021], Jing et al. [2021], Ganea et al. [2021], molecular low-energy generation prediction Shi et al. [2021], Xu et al. [2022], Méndez-Lucio et al. [2021], property/function prediction Schütt et al. [2017], Somnath et al. [2021] and…”
EquiBind: Geometric Deep Learning for Drug Binding Structure Prediction (2022, background): “However, very recent DL models have produced a breakthrough in computational protein folding (Jumper et al., 2021b; Baek et al., 2021) and fast generation of small molecule low-energy conformation ensembles (Ganea et al., 2021b; Luo et al., 2021; Xu et al., 2020; Shi et al., 2021; Du et al., 2022).”
GeoDiff: a Geometric Diffusion Model for Molecular Conformation Generation (2022, methodology): “2, our observation is in line with previous studies (Shi et al., 2021) that the state-of-the-art ML models (shown in Tab.”
Torsional Diffusion for Molecular Conformer Generation (2022, methodology): “A number of machine learning methods for conformer generation has been developed [Xu et al., 2021a,b, Shi et al., 2021, Luo et al., 2021], the most recent and advanced of which are GeoMol [Ganea et al., 2021] and GeoDiff [Xu et al., 2022].”
Predicting Molecular Conformation via Dynamic Graph Score Matching (2021, methodology): “The state-of-the-art method CONFGF [34] estimates pseudo-forces acting on atoms and generates conformations via Langevin MCMC [42], which bypasses the two-stage fashion in previous works and enhances the performance significantly.”
Diffusion Models: A Comprehensive Survey of Methods and Applications (2022, methodology): “Interdisciplinary Applications Molecular Graph Modeling [107],[94],[3],[259],[233], [247],[210],[152] Material Design [256],[153] Medical Image Reconstruction [224],[36],[37],[38],[178],[257]”
MolGenSurvey: A Systematic Survey in Machine Learning Models for Molecule Design (2022, methodology): “ConfGF [168] 2D-Graph 3D-Geometry Diffusion EVFN [55] 2D-Graph 3D-Geometry Diffusion DGSM [169] 2D-Graph 3D-Geometry Diffusion GeoDiff [170] 2D-Graph 3D-Geometry Diffusion”
Uni-Mol: A Universal 3D Molecular Representation Learning Framework (2023, methodology): “Following the settings in previous works [47, 55], we use GEOM-QM9 and GEOM-Drugs [56] dataset in this task.”
Antigen-Specific Antibody Design and Optimization with Diffusion-Based Generative Models for Protein Structures (2022, background): “Recently, progress has been made in developing equivariant diffusion models for molecular 3D structures [51, 19, 42].”
Equivariant Diffusion for Molecule Generation in 3D (2022, background): “In recent works (Shi et al., 2021; Luo et al., 2021; Xu et al., 2022) it was shown that score-based and diffusion models are effective at coordinate prediction, especially when the underlying neural network respects the symmetries of the data.”
SE(3)-Equivariant Energy-based Models for End-to-End Protein Folding (2021, background): “[47, 31, 27] only focus on small molecules consist of tens of atoms, while proteins are macro-molecules and often involve over thousands of atoms.”
TorchDrug: A Powerful and Flexible Machine Learning Platform for Drug Discovery (2022, background): “Dai et al. (2019); Shi et al. (2021) develop two delicate packages for retrosynthesis prediction and 3D molecular conformation prediction, respectively.”
Equivariant 3D-conditional diffusion model for molecular linker design (2022, unknown): “More details on datasets, baselines and metrics can be found in Methods.”
Generating Molecular Conformer Fields (2023, methodology): “To avoid the accumulative errors from two-stage generation, (Shi et al., 2021) implement score-based generative model to directly model the gradient of logarithm density of atomic coordinates.”
Crystal Diffusion Variational Autoencoder for Periodic Material Generation (2021, methodology): “These vectors are used to predict the scores, and we follow Song & Ermon (2019); Shi et al. (2021) to parameterize the score network with noise scaling: s X ( ˜ M | z ; σ A , σ X ) = s X ( ˜ M | z ) /σ X .”