High-throughput electronic property prediction of cyclic molecules with 3D-enhanced machine learning

Peikun Zheng, Olexandr Isayev

TL;DR

This paper introduces "Ring Vault," a large and diverse dataset of cyclic molecules, and demonstrates that a 3D-aware machine learning model (AIMNet2) significantly outperforms 2D models in predicting key electronic properties. This advancement promises to accelerate the discovery of new functional materials and drug candidates by enabling more accurate and efficient virtual screening.

What Problem Does This Paper Address?

Designing new organic molecules for applications like pharmaceuticals, organic electronics, and energy materials often relies on understanding their electronic properties. Cyclic structures are fundamental building blocks in many such molecules, influencing their shape, flexibility, and ultimately, their function. However, existing datasets used to train machine learning (ML) models for predicting these properties often lack sufficient structural diversity, particularly for complex cyclic systems. This limitation hinders the generalizability and accuracy of ML models, making it challenging to explore new chemical spaces effectively. The paper addresses this by creating a more comprehensive dataset and developing a more sophisticated ML approach that leverages 3D structural information.

Methods

The researchers employed a multi-pronged approach:

  • Dataset Construction: They built "Ring Vault," a large dataset of over 200,000 cyclic molecules, encompassing monocyclic, bicyclic, and tricyclic structures with various ring sizes and elements. This dataset was curated from commercial sources and literature.
  • Quantum Mechanical (QM) Calculations: To generate ground truth data for training, they performed QM calculations (DFT) on a subset of 36,000 molecules. These calculations determined five key electronic properties: HOMO-LUMO gap, ionization potential (IP), electron affinity (EA), and oxidation/reduction potentials (Eox, Ered).
  • Machine Learning Model Training: They trained and evaluated three ML models:
  • Graph Attention Network (GAT): A 2D graph-based model.
  • Chemprop: Another established 2D graph-based model.
  • AIMNet2: A model that was fine-tuned to incorporate 3D conformational information. This model was initially used to generate low-energy 3D conformers for the QM calculations.
  • Principal Component Analysis (PCA): PCA was used on the AIMNet2 model's learned molecular embeddings to explore the relationships between molecular structure and predicted properties.

Datasets

  • Ring Vault: A newly constructed dataset containing 201,546 cyclic molecules, including monocyclic, bicyclic, and tricyclic systems, spanning 11 non-metallic elements. This dataset aims to provide greater structural diversity than existing benchmarks.
  • QM-calculated subset: A subset of 36,000 molecules from Ring Vault was used for QM calculations to generate training and validation data.

Key Insights and Ideas

  • 3D is Crucial for Electronic Properties: The study strongly suggests that incorporating 3D conformational information is vital for accurately predicting electronic properties, outperforming 2D-based models. This highlights that the spatial arrangement of atoms, not just their connectivity, plays a significant role.
  • Dataset Diversity Matters: The underrepresentation of monocyclic molecules in the training set led to higher prediction errors for this class, underscoring the importance of balanced and diverse datasets for robust ML model generalization.
  • AIMNet2's Power: The fine-tuned AIMNet2 model, leveraging 3D data, demonstrated superior performance, achieving high accuracy (R² > 0.95) and significantly reducing errors compared to 2D models.
  • Structure-Property Correlations Revealed: PCA analysis of AIMNet2 embeddings revealed clear correlations between electronic properties (like HOMO-LUMO gap, Eox, Ered) and structural features (e.g., conjugation extent, functional groups), offering valuable insights for rational molecular design.
  • Redox Potential Nuances: The study showed strong linear correlations between IP/Eox and EA/Ered, validating the model and providing theoretical insights into redox behavior. It also highlighted how specific functional groups (electron-donating vs. withdrawing) and their positions can tune redox potentials.

Key Results

  • The fine-tuned AIMNet2 model achieved R² values exceeding 0.95 and reduced Mean Absolute Errors (MAEs) by over 30% compared to 2D models (GAT and Chemprop) on the test set.
  • All models achieved R² values above 0.9, demonstrating good predictive capabilities.
  • The HOMO-LUMO gap was identified as the most challenging property to predict accurately, likely due to its sensitivity to electronic structure and conformation.
  • PCA analysis showed distinct trends for different properties along the principal components, linking them to structural features like conjugation and functional groups.
  • The study demonstrated that molecular structural modifications (e.g., adding electron-withdrawing groups, altering conjugation) can effectively tune electronic properties like Ered and HOMO-LUMO gap, with AIMNet2 predictions closely matching DFT results.

Limitations and Future Work

  • Monocyclic Molecule Underrepresentation: The authors acknowledge that the underrepresentation of monocyclic molecules in the training set impacted their prediction accuracy. Augmenting the training data for these structures could improve performance.
  • Fine Energetic Discrimination: For applications requiring extremely high precision (e.g., <0.05 eV) or direct experimental comparison, further high-level validation is recommended.
  • Outlier Structures: Molecules with unusual structural characteristics (highly strained rings, irregular conjugation, silicon heterocycles) showed larger prediction errors, indicating the need for broader training set coverage to capture such cases.
  • Solvent Effects: While aqueous solution was used for redox potential calculations, the authors note that organic solvents are often more relevant for practical applications. Future work could extend predictions to specific organic solvents.
  • Further Descriptors: Incorporating quantitative physical organic chemistry descriptors (like Hammett parameters) could provide deeper mechanistic insights.
  • Expanding Scope: Future directions include expanding the dataset to more complex polycyclic systems and additional heteroatoms, integrating active learning for more efficient exploration, and predicting other properties like optical or thermal characteristics.

Summary updated

Abstract

Complex organic molecules play a pivotal role in bioactive compounds and organic functional materials, yet existing molecular datasets lack structural diversity for such systems, limiting the generalizability of machine learning (ML) models. This study introduces a high-quality dataset, Ring Vault, comprising 201 546 cyclic molecules, including monocyclic, bicyclic, and tricyclic systems, spanning 11 non-metallic elements. This dataset covers a wide chemical space and provides a robust foundation for molecular property prediction. Leveraging quantum mechanical (QM) calculations on a subset (36 000 molecules), we trained three ML models (Graph Attention Network, Chemprop, and AIMNet2) to predict five key electronic properties: HOMO-LUMO gap, ionization potential (IP), electron affinity (EA), and redox potentials (<i>E</i> <sub>ox</sub>, <i>E</i> <sub>red</sub>). The fine-tuned AIMNet2 model, incorporating 3D conformational information, outperformed 2D-based models, achieving <i>R</i> <sup>2</sup> values exceeding 0.95 and reducing mean absolute errors (MAEs) by over 30%. Principal component analysis (PCA) of AIMNet2 embeddings revealed intrinsic correlations between electronic properties and structural features, such as conjugation extent and functional group effects. This work establishes a robust framework for high-throughput screening and rational design of cyclic molecules, with applications spanning drug discovery, organic electronics, and energy materials. The dataset and methodology provide a foundation for exploring complex structure-property relationships and accelerating functional molecule discovery.
BibTeX
@inproceedings{zheng2025high,
  author = {Peikun Zheng and Olexandr Isayev},
  title = {{High-throughput electronic property prediction of cyclic molecules with 3D-enhanced machine learning}},
  booktitle = {Chemical},
  year = {2025},
  doi = {10.1039/d5sc04079e},
}
Figure