Molecular Structure Characterization Learning and Property Prediction
Developing equivariant geometry-enhanced graph neural networks and geometric transformers that elegantly extract geometric features and efficiently model molecular structures with low computational costs for molecular property prediction, drug discovery, and molecular dynamics simulation
Geometric deep learning has been revolutionizing the molecular modeling field. Despite state-of-the-art neural network models approaching ab initio accuracy for molecular property prediction, their applications in drug discovery and molecular dynamics simulation have been hindered by insufficient utilization of geometric information and high computational costs. Our research addresses these fundamental challenges through equivariant geometry-enhanced graph neural networks and geometric transformers.
Our key innovations include ViSNet, which elegantly extracts geometric features and efficiently models molecular structures with low computational costs; Geoformer, a novel geometric Transformer with Interatomic Positional Encoding that paves the way for molecular geometric modeling based on Transformer architecture; and Long-Short-Range Message-Passing frameworks that capture non-local interactions for scalable molecular dynamics simulation. These advances enable efficient exploration of conformational space, superior drug-target interaction prediction, and provide reasonable interpretability to map geometric representations to molecular structures.
Research Framework Architecture
Comprehensive framework integrating equivariant geometry-enhanced graph neural networks and geometric transformers for accurate molecular property prediction and structure characterization
The molecular structure characterization framework combines equivariant geometry-enhanced graph neural networks with geometric transformers to achieve superior molecular property prediction and structure understanding.
Key Research Areas
Equivariant geometry-enhanced graph neural networks that elegantly extract geometric features for molecular structure characterization.
- Vector-scalar interactive message passing
- Long-short-range message passing
- Conformational space exploration
Geometric transformers with interatomic positional encoding that parameterize atomic environments for molecular structure modeling.
- Interatomic positional encoding
- Three-way Transformer architecture
- Geometric feature extraction
Physics-informed frameworks for scalable molecular dynamics simulation that capture non-local interactions with low computational costs.
- Long-range interaction modeling
- Scalable MD simulation
- Chemical accuracy preservation
Technical Innovations
We develop cutting-edge graph neural network architectures that push the boundaries of what's possible in molecular modeling:
Multi-Scale GNNs
Hierarchical representations from atoms to complexes
Attention Mechanisms
Learning important molecular interactions dynamically
Equivariant Layers
Preserving molecular symmetries in predictions
Accuracy Improvements
- - 50% reduction in force prediction errors
- - 10x improvement in energy conservation
- - State-of-the-art on QM9 benchmark
- - Superior performance on MD17 dataset
Computational Efficiency
- - 100x faster than DFT calculations
- - Linear scaling with system size
- - GPU-optimized implementations
- - Distributed training capabilities
Applications & Future Prospects
Drug Discovery
- Virtual screening acceleration
- Lead optimization guidance
- ADMET property prediction
- Drug-target interaction modeling
Materials Science
- Catalyst design and optimization
- Battery material discovery
- Polymer property prediction
- Crystal structure prediction
Key Publications
Enhancing geometric representations for molecules with equivariant vector-scalar interactive message passing
Wang, Y., Wang, T., Li, S., He, X., Li, M., Wang, Z., Zheng, N., Shao, B., & Liu, T. Y.
Innovation: Equivariant geometry-enhanced graph neural network that elegantly extracts geometric features and efficiently models molecular structures with low computational costs
Contribution: Outperforms SOTA on multiple MD benchmarks (MD17, revised MD17, MD22) and achieves excellent chemical property prediction on QM9 and Molecule3D datasets
Impact: Efficiently explores conformational space with reasonable interpretability, addressing key challenges in drug discovery and MD simulation applications
Geometric transformer with interatomic positional encoding
Wang, Y., Li, S., Wang, T., Shao, B., Zheng, N., & Liu, T. Y.
Innovation: Novel geometric Transformer with Interatomic Positional Encoding (IPE) that parameterizes atomic environments as Transformer's positional encodings
Contribution: Outperforms SOTA algorithms on QM9 and achieves best performance on Molecule3D for both random and scaffold splits, compared with both Transformers and equivariant GNN models
Impact: Paves the way for molecular geometric modeling based on Transformer architecture, opening new avenues for molecular modeling applications
Long-short-range message-passing: A physics-informed framework to capture non-local interaction for scalable molecular dynamics simulation
Li, Y., Wang, Y., Huang, L., Yang, H., Wei, X., Zhang, J., Wang, T., Wang, Z., Shao, B., & Liu, T. Y.
Innovation: Physics-informed framework generalizing existing equivariant GNNs to incorporate long-range interactions efficiently and effectively, inspired by fragmentation-based methods
Contribution: Demonstrates SOTA results with up to 40% MAE reduction for molecules in MD22 and Chignolin datasets, with consistent improvements to various EGNNs
Impact: Addresses satisfactory description of long-range and many-body interactions, enabling scalable molecular dynamics simulation for chemical and biological systems
SAMF: a self-adaptive protein modeling framework
Ding, W., Xu, Q., Liu, S., Wang, T., Shao, B., Gong, H., & Liu, T. Y.
Innovation: Self-adaptive framework that eliminates redundancy of constraints, resolves conflicts, and folds protein structures iteratively with deep quality analysis system
Contribution: Achieves SOTA performance by exploiting cutting-edge deep learning techniques without complicated domain knowledge and numerous patches as barriers
Impact: Modular design enables easy customization and extension, with superiority amplified over time as input constraint quality improves
Improved drug-target interaction prediction with intermolecular graph transformer
Liu, S., Wang, Y., Deng, Y., He, L., Shao, B., Yin, J., Wang, T., & Liu, T. Y.
Innovation: Dedicated attention mechanism modeling intermolecular information with three-way Transformer-based architecture, addressing topological and spatial limitations
Contribution: Outperforms SOTA by 9.1% and 20.5% for binding activity and pose prediction, with superior generalization to unseen receptor proteins
Impact: Exhibits promising drug screening ability against SARS-CoV-2, identifying 83.1% active drugs validated by wet-lab experiments with near-native binding poses
DSN-DDI: An accurate and generalized framework for drug-drug interaction prediction by dual-view representation learning
Li, Z., Zhu, S., Shao, B., Zeng, X., Wang, T., & Liu, T. Y.
Innovation: Novel dual-view drug representation learning network employing local and global modules iteratively, learning substructures from single drug (intra-view) and drug pair (inter-view)
Contribution: Achieves 13.01% relative improvement and >99% accuracy under transductive setting, with 7.07% relative improvement for unseen drugs
Impact: Exhibits good transferability on synergistic drug combination prediction, serving as generalized framework for drug discovery field applications