Molecular Representation via Fully Connected Graph Aggregation

The molecule is segmented into atomic-level components, with each atom represented as a node in a graph structure. This segmentation allows the model to capture local atomic features and their relationships, improving prediction accuracy by analyzing molecular structure at the atomic level rather than treating the molecule as a whole.

The patent transforms molecular representation from traditional 1D sequences or 2D images to 3D graph structures with spatial coordinates. By incorporating three-dimensional spatial information and using graph neural networks to process these multi-dimensional atomic relationships, the model achieves more accurate prediction of molecular attributes while maintaining computational efficiency.

The patent performs preliminary encoding of atomic features (such as atom type, charge, and spatial coordinates) into fixed-dimensional vectors before processing. This pre-processing step organizes molecular data into a standardized graph structure with atomic nodes and bond edges, reducing the complexity of subsequent neural network processing while preserving essential molecular information for accurate prediction.

The patent replaces traditional mechanical molecular modeling approaches with graph neural network-based computational methods. By using message-passing mechanisms in GNNs to automatically learn atomic interactions from graph structures, the system achieves high prediction accuracy without requiring complex explicit physical simulations, thereby reducing computational complexity.