IDEA #68OJKC PimNet: Physics-Informed Multiplex Graph Neural Network for Molecular Structures.5/14/2021 17:22; DA.

21A0021 5/14/2021 17:22; DA Shuo Zhang, Graduate Center, Computer Science Lei Xie, Hunter College, Computer Science (primary contact) DESCRIPTION OF SOFTWARE 4. Please provide a comprehensive summary of the software, particularly pointing out program characteristics such as software use and utility, functionality, and anything else that identifies the strengths of the software. Use as much space as necessary. The software contains a novel Graph Neural Network (GNN) algorithm to predict the properties of molecules from their 3D structures. As inspired by molecular mechanics, the software decouples the modeling of local and non-local interactions in 3D molecules with multiplex graphs to reduce the expensive angle-related computations. As a result, the software is efficient regarding time and space when applied on both small molecules and macromolecules (e.g. proteins or RNAs). Besides, the software can predict vectorial properties (e.g. dipole moment) besides scalar properties (e.g. energies) by addressing geometric vectors and quantum mechanics. The performance on benchmark datasets demonstrates the state-of-the-art performance of the software as compared to other baselines. 5. What are the immediate and/or future applications of the invention, and what problem does it solve? The software can be directly used to predict the properties (either scalar or vectorial) of molecules (either micro or macro) given their 3D structures. Since the prediction of molecular properties is fundamental and crucial in molecular discovery process, the software can be used as a building block to learn atom-level or molecule-level embeddings in solving material design and drug discovery problems such as molecule generation, protein design, reaction prediction, retrosynthesis prediction and so on. 6. What are the advantages and novel features of the invention over other current/anticipated technologies? Please describe competing technologies or procedures (include names of products, software, etc.) and their deficiencies, as well as those making or selling such products. The software uniquely decouples the modeling of local and non-local interactions in 3D molecules with multiplex graphs as inspired by molecular mechanics. By doing so, the software greatly reduces the expensive angle-related computations in previous state-of-the-art Graph Neural Networks (DimeNet, DimeNet++, HMGNN). In detail, when modeling the local interactions, the software encodes both angular information and pairwise distance information. As for non-local interactions, the software only incorporates pairwise distance information. To address the relations between local and non-local interactions, the software uses attention mechanism to combine them for final prediction. The software can also predict vectorial properties (e.g. dipole moment) besides scalar properties (e.g. energies) by addressing geometric vectors and quantum mechanics. As a contrast, most of the existing GNNs (e.g. SchNet, PhysNet, MEGNet, MGCN, DeepMoleNet, DimeNet, DimeNet++, HMGNN) only focus on predicting scalar properties. As evaluated on small molecule dataset QM9 and protein-ligand complex dataset PDBBind, the software outperforms the state-of-the-art baselines using significantly less memory, making it applicable to real-world problems. When predicting vectorial dipole moment, the software outperforms the previous best model by 94%. 7. Are there any open-source elements included/embedded in the Source Code. If yes, please list all third party code embedded in or accessed by the disclosed software when such software is run. This list must include, all open source code (free executable code, public domain code, library code and all other executable or source code not written by any of the Inventor(s) listed in Section 3 of this form) directly embedded in the software when it is executed. If the code is not available on the web, please provide copies of any license agreements governing your use of the third party code Python, PyTorch, torchvision, PyTorch Geometric, PyTorch Sparse, PyTorch Scatter, PyTorch Cluster, NumPy, scikit-learn, SciPy, SymPy, RDKit, pytorch-gradual-warmup-lr, tqdm, CUDA Toolkit.