Neural networks to identify chemicals

The patent segments the complex drug discovery process into multiple specialized neural network models, each handling specific tasks such as protein folding prediction, ligand binding affinity prediction, and molecular generation. This segmentation allows each model to focus on specific physicochemical interactions without needing to model all interactions simultaneously, thereby improving prediction accuracy while managing computational complexity through modular architecture

The patent introduces neural network models as intermediary components between the protein structure and drug candidate evaluation. These neural networks serve as mediators that predict complex physicochemical interactions without requiring direct calculation of all atomic-level interactions, thus improving prediction accuracy while reducing the computational burden of modeling every interaction explicitly

The patent employs preliminary action by using neural network models to pre-dict protein folding structures and ligand binding affinities before conducting full drug candidate evaluation. This preliminary prediction step filters out unlikely candidates early in the process, allowing comprehensive interaction analysis to be focused only on promising candidates, thereby improving overall prediction accuracy while reducing total computational time

The patent substitutes traditional mechanical computational methods (direct calculation of physicochemical interactions) with neural network-based predictive models. These models learn from training data to predict interaction outcomes without explicitly calculating each interaction, thereby maintaining high prediction accuracy while dramatically reducing computational time and resources required

The patent merges multiple specialized neural network models into an integrated end-to-end computational workflow. The models are combined such that outputs from one model serve as inputs to subsequent models, creating a unified system that leverages the strengths of each individual model while presenting a cohesive interface for drug candidate identification, thereby improving reliability without proportionally increasing operational complexity

The patent designs the neural network models to serve multiple functions within the drug discovery pipeline. For example, the same neural network architecture is used for both protein folding prediction and ligand binding affinity prediction, and models can operate in different modes (generation, evaluation, optimization) depending on the specific task, thereby improving prediction reliability across multiple tasks while reducing the total number of distinct models required