patents

A generative machine-learning model that predicts the 3D structure of a molecular system by iteratively refining a candidate structure. At each time step, the system builds atom embeddings from the current predicted structure, updates them with an equivariant attention operation, and decodes the updated embeddings into a refined 3D structure for the molecular system.

A modular generative machine-learning system for predicting the 3D atomic structure of a complex that includes a protein and a ligand. The architecture composes separate models for the protein, the ligand, and their bound complex to produce accurate predicted protein-ligand docking poses.

A method for training generative models of 3D molecular structure using an alignment objective. At each training step, the generative model proposes several alternative predicted 3D structures for a set of molecules, a respective alignment score is computed for each candidate, and the model parameters are updated to optimize the alignment objective.

A computational platform that automates the generation and screening of large molecular libraries for drug discovery. Candidate molecules are generated through multiple strategies—attaching molecular fragments, simulating chemical reactions with neural networks, and completing masked scaffolds—then filtered with machine-learning models that predict properties such as binding affinity and toxicity, letting researchers explore far larger regions of chemical space in seconds rather than weeks of physical experimentation.

A neural network for video super-resolution that upscales an input video sequence to a higher target resolution. Each input frame is processed by a subnetwork of the network corresponding to the frame’s input resolution, so that a single model can handle a range of possible input and output resolutions.