HydroNet: Benchmark Tasks for Preserving Intermolecular Interactions and Structural Motifs in Predictive and Generative Models for Molecular Data

Sutanay Choudhury, Jenna A. Bilbrey, Logan T. Ward, Sotiris S. Xantheas, Ian T. Foster, Joseph P. Heindel, Ben Blaiszik, Marcus E. SchwartingarXiv 2020 · OpenReview

Introduction

HydroNet is a comprehensive open-source dataset comprising 4.95 million water clusters, designed to benchmark machine learning (ML) models on their ability to capture complex, long-range intermolecular interactions. By providing both 3D coordinates and graph-based representations, HydroNet challenges the community to move beyond simple covalent bonding and master the many-body effects inherent in hydrogen bonding.

Problem Statement

In the field of molecular machine learning, most models excel at understanding covalent bonds—the strong, short-range links that hold atoms together within a molecule. However, many of nature’s most important phenomena, from protein folding to electrolyte transport in batteries, are driven by non-covalent interactions, such as hydrogen bonding.

These interactions are notoriously difficult for ML to model because they are:

  1. Long-range: They act over larger distances than covalent bonds.
  2. Many-body: The interaction between two water molecules is influenced by the presence of a third or fourth nearby.
  3. Structurally Diverse: Even with the same number of molecules, water can form a dizzying array of "isomers" (different shapes) with nearly identical energies.

HydroNet fills this gap by providing a playground for developing models that can identify the structural motifs (like cages and rings) that emerge from these subtle forces.

Methods

The researchers generated a dataset of 4.95 million low-energy water clusters ranging from 3 to 30 water molecules each. Potential energies were calculated using the TTM2.1-F potential, a highly accurate model based on first-principles physics.

Data representations in HydroNet: XYZ coordinates, atomic graphs, and coarse-grained molecular graphs.

To support different ML approaches, the data is provided in three formats:

  • XYZ Coordinates: Raw spatial data for 3D-aware models.
  • Atomic Interaction Graphs: Nodes are atoms; edges represent both covalent and hydrogen bonds.
  • Coarse Graphs: Nodes are entire water molecules; edges represent only the intermolecular hydrogen bonds.

Datasets

The core contribution is the HydroNet dataset, which contains 4.95 million samples of water clusters. This is significantly larger and more focused on intermolecular patterns than standard datasets like QM9, which primarily focus on small organic molecules and covalent properties. The data is distributed in JSON and TensorFlow Protobuf formats containing coordinates, energies, and connectivity.

Key Insights and Ideas

The Geometry Gap

A striking finding is the performance difference between models that "see" 3D space and those that only "see" connectivity. Models using 3D coordinates (like SchNet) are roughly 100 times more accurate than standard graph-based models (MPNNs). This suggests that current graph neural networks are missing critical spatial information necessary to understand hydrogen bonding.

Structural Evolution

The paper highlights how water clusters change "personality" as they grow. Small clusters (N=3-5) are quasi-planar; medium clusters (N=7-16) become cubic; and large clusters (N=17+) form complex, cage-like structures. A successful generative model must capture this progression.

Graph-theoretic analysis showing the evolution of cluster complexity (degree, path length, and cycles) with increasing size.

The "Coarse" Advantage

While graph models currently lag behind, the authors found that models trained on "coarse" graphs (treating the whole water molecule as a single node) actually generalize better to unseen cluster sizes than those trained on raw atomic graphs.

Key Results

  • SchNet (Geometry-to-Energy): Achieved a Mean Absolute Error (MAE) of 0.0427 kcal/mol. This is highly accurate, though the model required ~65 hours of training on 4 NVIDIA V100 GPUs.
  • MPNN (Graph-to-Energy): Achieved an MAE of ~0.5 kcal/mol. While this is "chemical accuracy" in some contexts, it is two orders of magnitude worse than the geometry-based approach, highlighting a major area for architectural improvement in graph ML.
  • Generalization: The models were tested on cluster sizes (N=10 and N=30) outside the primary training range (N=11-29) to ensure they weren't just memorizing specific shapes.

Limitations and Future Work

Conventional Message Passing Neural Networks (MPNNs) struggle with long-range interactions. Future work needs to explore new network designs—perhaps using global attention or higher-order graph structures—to bridge the 100x accuracy gap. Additionally, while the paper defines the generative task, generating 3D structures that satisfy both graph-theoretic constraints and energy minimization remains an open challenge. The ultimate goal is to use these small cluster models to build better simulations for bulk water and ice.

Summary updated

Abstract

Intermolecular and long-range interactions are central to phenomena as diverse as gene regulation, topological states of quantum materials, electrolyte transport in batteries, and the universal solvation properties of water. We present a set of challenge problems for preserving intermolecular interactions and structural motifs in machine-learning approaches to chemical problems, through the use of a recently published dataset of 4.95 million water clusters held together by hydrogen bonding interactions and resulting in longer range structural patterns. The dataset provides spatial coordinates as well as two types of graph representations, to accommodate a variety of machine-learning practices.
BibTeX
@misc{choudhury2020hydronet,
  author = {Sutanay Choudhury and Jenna A. Bilbrey and Logan T. Ward and Sotiris S. Xantheas and Ian T. Foster and Joseph P. Heindel and Ben Blaiszik and Marcus E. Schwarting},
  title = {{HydroNet: Benchmark Tasks for Preserving Intermolecular Interactions and Structural Motifs in Predictive and Generative Models for Molecular Data}},
  year = {2020},
  eprint = {2012.00131},
  archivePrefix = {arXiv},
  url = {https://arxiv.org/abs/2012.00131},
}
Figure