Source-linked AI summary
Molecular Property Prediction: A Multilevel Quantum Interactions Modeling Perspective
Chengqiang Lu, Qi Liu, Chao Wang, Zhenya Huang, Peize Lin, Lixin He
TL;DR
Molecular property prediction is important but is constrained by DFT’s computational cost, limited labeled data, and underexplored quantum interactions. MGCN uses molecular graphs and hierarchical interaction layers to model these interactions, with experiments reporting generalizability and transferability across molecule sizes and datasets.
Problem
Molecular property prediction requires efficient modeling of complex quantum interactions despite DFT’s high cost and limited labeled data, especially for large molecules.
Method
MGCN represents molecules as graphs and uses hierarchical interaction layers to learn atom-wise, pair-wise, triple-wise, and higher-level representations for prediction.
Results
MGCN is reported to be generalizable and transferable across equilibrium and off-equilibrium molecular datasets and across molecule sizes.
Takeaways & Limitations
MGCN provides a multilevel molecular-interaction modeling approach for property prediction using overall interaction representations.
Takeaways & Limitations
The authors identify generalizing atom representations as future work because sufficiently large-molecule datasets are difficult to obtain.
Abstract
from arXiv · showhide
Predicting molecular properties (e.g., atomization energy) is an essential issue in quantum chemistry, which could speed up much research progress, such as drug designing and substance discovery. Traditional studies based on density functional theory (DFT) in physics are proved to be time-consuming for predicting large number of molecules. Recently, the machine learning methods, which consider much rule-based information, have also shown potentials for this issue. However, the complex inherent quantum interactions of molecules are still largely underexplored by existing solutions. In this paper, we propose a generalizable and transferable Multilevel Graph Convolutional neural Network (MGCN) for molecular property prediction. Specifically, we represent each molecule as a graph to preserve its internal structure. Moreover, the well-designed hierarchical graph neural network directly extracts features from the conformation and spatial information followed by the multilevel interactions. As a consequence, the multilevel overall representations can be utilized to make the prediction. Extensive experiments on both datasets of equilibrium and off-equilibrium molecules demonstrate the effectiveness of our model. Furthermore, the detailed results also prove that MGCN is generalizable and transferable for the prediction.
Introduction
Molecular property prediction matters for applications such as drug design and material manufacture, but existing approaches face computational and modeling challenges. MGCN addresses these challenges by representing molecules as interaction graphs and modeling quantum interactions hierarchically.
- Molecular property prediction supports substance discovery, drug design, and new-material manufacture.
- DFT models quantum interactions but can take nearly an hour for one 20-atom molecule, limiting large-scale prediction.
- Existing machine-learning approaches often use hand-crafted features or grid-like molecular representations, while complex quantum interactions remain difficult to model.
- Limited labeled data and the scarcity of large-molecule data require models that generalize from smaller molecules and transfer to larger ones.
- MGCN represents each molecule as an interaction graph and uses hierarchical layers to model atom-wise, pair-wise, triple-wise, and higher-level quantum interactions.
Related Work
Prior approaches include DFT, traditional machine learning with hand-crafted descriptors, and deep neural networks. MGCN extends molecular graph modeling by directly representing multilevel quantum interactions and passing interaction information level by level.
- Density Functional Theory: DFT-based molecular property prediction models quantum interactions but has approximate complexity O(N^3), making large-scale prediction time-consuming.
- Traditional Machine Learning Methods: Traditional machine-learning methods use domain-informed hand-crafted features such as bag of bonds, Coulomb matrices, and distance or angle histograms.
- Deep Neural Networks: Deep neural networks offer automatic feature learning and have been applied to molecular property prediction.
- Our Work: MGCN models molecular quantum interactions hierarchically at atom-wise, pair-wise, triple-wise, and higher levels.
- Our Work: MGCN passes interaction information level by level to support generalizability with limited data and transferability across unbalanced molecule sizes.
Multilevel Graph Convolutional Network
MGCN represents molecules as interaction graphs and uses hierarchical interaction layers to model quantum interactions across increasing levels. Its architecture combines atom, bond, and spatial information to support generalization and transferability across molecule sizes.
- Problem formulation: MGCN represents each molecule as a graph whose vertices are atoms and whose edges encode atom interactions, edge types, and spatial information.The graph is treated as complete and undirected, with |E| = N(N − 1)/2.
- Multilevel interaction modeling: Hierarchical interaction layers model atom-wise, pair-wise, triple-wise, and higher-order interactions by repeatedly updating edge and atom representations.Each layer combines node, edge, and spatial information, producing atom representations that reflect progressively higher-order interactions.
- Architecture: The architecture comprises embedding, radial basis function, interaction, and readout layers that transform molecular inputs into predictions.Embedding layers encode atoms and edges, the RBF layer converts distances into tensors, interaction layers learn representations at different levels, and readout produces the final result.
- Readout: MGCN exploits molecular additivity by processing atom- and edge-related representations separately and aggregating them into the predicted property.The edge-related contribution may be ignored when data are scarce because atom-related interactions account for most molecular interactions.
- Spatial encoding: The RBF layer converts coordinate-derived inter-atomic distances into robust distance tensors while removing dependence on the coordinate frame.The resulting tensor D has shape N×N×K, with d_ij representing the distance tensor between atoms i and j.
- Generalizability and transferability: The multilevel design is intended to improve generalizability with limited labeled data and transferability from smaller molecules to larger ones.The paper links this capability to shared atom and edge knowledge and to representations built progressively across interaction levels.
Experiments
Experiments evaluate MGCN on QM9 and ANI-1, examine multilevel interaction modeling, generalizability, transferability, molecule size, and interaction-layer depth. MGCN performs strongly across these evaluations, including varied training-set sizes and transfer from smaller to larger molecules.
- Predictive performance: MGCN gets the best performance in 11 out of 13 QM9 properties, with 11 properties exceeding chemical accuracy.
- Predictive performance: MGCN achieves satisfactory accuracy and outperforms other methods on the off-equilibrium ANI-1 dataset.ANI-1 is more challenging than QM9 because it contains off-equilibrium molecules and is substantially larger.
- Effectiveness of multilevel interactions: MGCN is more accurate and stable than SchNet as molecule size increases, with its advantage becoming more apparent for larger molecules.The experiment evaluates MAE for µ, εHOMO, U, and Cv as a function of atom count.
- Effectiveness of multilevel interactions: A control model blending all interaction levels into one embedding reaches an MAE of 0.03683 for U, worse than MGCN.This comparison supports constructing molecular representations level by level.
- Generalizability: MGCN obtains the lowest MAE across training sets containing 50k, 100k, and 110k samples.All three settings use the same 10k-molecule test set.
- Transferability: Pre-training on small molecules outperforms the other transfer settings when the training set is small, while direct development-model transfer to large molecules also improves as small-data volume increases.The transfer experiments sample 50k small and 50k large molecules from QM9.
- Influence of interaction layers: Four interaction layers produce the best empirical result; fewer layers lack capacity, whereas more than five widen the generalization gap.The layer experiment measures MAE for property U0.
- Summary: MGCN has theoretical time complexity O(N^2) versus O(N^3) for DFT and predicts one molecule in 2.4×10−2 seconds, nearly 1.5×10^5 times faster.The timing comparison uses a single core of a Xeon E5-2660.
Conclusion
MGCN models multilevel quantum interactions for molecular property prediction and is reported as generalizable and transferable. The authors identify improving atom-representation generalization as a future priority because sufficiently large-molecule datasets are difficult to obtain.
- MGCN learns quantum-interaction representations level by level and uses the overall interaction representation for molecular property prediction.
- Experiments on two prevalent datasets demonstrate the competency of the proposed approach.
- The model is reported to be generalizable and transferable.
- Future research should improve atom-representation generalization because sufficiently large-molecule datasets are difficult to obtain.