Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

REBIND: Enhancing ground-state molecular conformation via force-based graph rewiring

Taewon Kim, Hyunjin Seo, Sungsoo Ahn, Eunho Yang

TL;DR

REBIND is a novel framework that rewires molecular graphs by adding edges based on the Lennard-Jones potential to capture non-bonded interactions for low-degree atoms, achieving up to a 20\% reduction in prediction error.

Abstract

Predicting the ground-state 3D molecular conformations from 2D molecular graphs is critical in computational chemistry due to its profound impact on molecular properties. Deep learning (DL) approaches have recently emerged as promising alternatives to computationally-heavy classical methods such as density functional theory (DFT). However, we discover that existing DL methods inadequately model inter-atomic forces, particularly for non-bonded atomic pairs, due to their naive usage of bonds and pairwise distances. Consequently, significant prediction errors occur for atoms with low degree (i.e., low coordination numbers) whose conformations are primarily influenced by non-bonded interactions. To address this, we propose REBIND, a novel framework that rewires molecular graphs by adding edges based on the Lennard-Jones potential to capture non-bonded interactions for low-degree atoms. Experimental results demonstrate that REBIND significantly outperforms state-of-the-art methods across various molecular sizes, achieving up to a 20\% reduction in prediction error.

REBIND: Enhancing ground-state molecular conformation via force-based graph rewiring

TL;DR

REBIND is a novel framework that rewires molecular graphs by adding edges based on the Lennard-Jones potential to capture non-bonded interactions for low-degree atoms, achieving up to a 20\% reduction in prediction error.

Abstract

Predicting the ground-state 3D molecular conformations from 2D molecular graphs is critical in computational chemistry due to its profound impact on molecular properties. Deep learning (DL) approaches have recently emerged as promising alternatives to computationally-heavy classical methods such as density functional theory (DFT). However, we discover that existing DL methods inadequately model inter-atomic forces, particularly for non-bonded atomic pairs, due to their naive usage of bonds and pairwise distances. Consequently, significant prediction errors occur for atoms with low degree (i.e., low coordination numbers) whose conformations are primarily influenced by non-bonded interactions. To address this, we propose REBIND, a novel framework that rewires molecular graphs by adding edges based on the Lennard-Jones potential to capture non-bonded interactions for low-degree atoms. Experimental results demonstrate that REBIND significantly outperforms state-of-the-art methods across various molecular sizes, achieving up to a 20\% reduction in prediction error.

Paper Structure

This paper contains 39 sections, 18 equations, 4 figures, 5 tables.

Table of Contents

  1. Introduction
  2. Preliminaries
  3. Problem definition.
  4. Multi-head self-attention.
  5. Inter-atomic interaction modeling in prior works.
  6. Limitation of inter-atomic force modeling in prior studies
  7. Limitations of prior works for non-bonded interactions.
  8. Experimental verification.
  9. REBIND: Enhancing ground-state molecular conformation via force-based graph rewiring
  10. Overview
  11. Force-aware graph rewiring
  12. Force modeling with Lennard-Jones potential.
  13. Construction of force adjacency matrix.
  14. Integration with multi-head self-attention
  15. Objective function
  16. ...and 24 more sections

Figures (4)

  • Figure 1: Atom-wise error analysis with respect to the relative atom degree on varing architectures in the QM9 dataset. (a) represents the average atom-wise RMSD while (b) shows the average atom-wise MAE for each bin of low-degree and high-degree atoms. $\Delta$ denotes the gap between the errors of the low-degree and high-degree groups.
  • Figure 2: Overview of the ReBind framework.
  • Figure 3: An illustration of edge augmentation in ReBind using the LJ potential. Non-bonded atomic pairs with the largest force magnitudes are added as edges to the molecular graph in a degree-compensating manner. The augmented edges are treated according to the nature of the forces, distinguishing between attractive and repulsive interactions.
  • Figure 4: Atom-wise E-RMSD analysis with respect to the relative atom degree on ReBind and baselines. $\Delta$ denotes the gap between the errors of the low-degree and high-degree bins.