Unsupervised invertible physics-based vector representation for molecules

Inventors

Byrnes, John J. • Rohwer, Richard J. • Luebke, Kevin J. • Madrid, Peter

Assignees

SRI International Inc

Abstract

The Artificial Intelligence engine can perform one or more operations. A query can be submitted to the Artificial Intelligence engine to search directly for a set of targeted properties for an unnamed molecule having the set of targeted properties. An indication of a structure of one or more candidate molecules found to have the set of targeted properties with the Artificial Intelligence engine is generated by applying one or more machine learning algorithms. The indication of the structure of the one or more candidate molecules found to satisfy the set of targeted properties in 3-dimensional space is supplied to a user in response to the query for the set of targeted properties to the Artificial Intelligence engine.

Core Innovation

An Artificial Intelligence engine is composed of two or more modules that use one or more machine learning models to predict molecular properties of a particular molecule and to inversely output a set of atoms and a geometric structure in three-dimensional space for an unlabeled molecule from targeted properties. A first machine learning model in a molecule property predictor module is trained to learn to predict molecular properties of a particular molecule. A molecule vector module represents a known molecule with its known geometric arrangement in three-dimensional space in vector format, where a portion of a vector value includes coordinate points in the three-dimensional space, and the molecule property predictor module cooperates with the molecule vector module to predict molecular properties from a molecule vector field in three-dimensional space.

An atom vector module correlates an atom in a set of two or more atoms to atom vectors that reflect an identity of the atom itself and an arrangement of a geometric structure of the atom in three-dimensional space, where a portion of a vector value for the atom includes coordinate points in the three-dimensional space. An input module cooperates with the molecule vector module, the molecule property predictor module, and the atom vector module to allow a user to input two or more atoms and obtain output of a set of properties for a resulting molecule formed by the two or more atoms. The same input module also allows a user to supply an input of targeted properties for an unlabeled molecule so that the first machine learning model outputs a set of atoms and the geometric structure in the three-dimensional space for the unlabeled molecule.

The framework further provides for the output of at least one of the set of atoms and the geometric structure in the three-dimensional space for the unlabeled molecule and the set of properties for the resulting molecule to a user on a display screen, in a printout, as an electronic message, or any combination of these. The system uses an Artificial Intelligence engine that ingests information from one or more knowledge sources, including chemistry expert information, as constraints to adapt a machine learning model to find a candidate molecule vector field that agrees with an atom vector field and thereby fit the created constraints. The described vector representations and module cooperation are configured to represent molecular and atomic geometric information in three-dimensional space and use the learned model or models to search for candidate molecule vector fields.

Claims Coverage

The independent claims cover an apparatus and a non-transitory computer-readable medium implementing an Artificial Intelligence engine with cooperating atom vector and molecule vector representations in three-dimensional space, enabling property prediction and inverse prediction of atoms and 3D geometric structure from targeted properties. The inventive features include defining molecule and atom vector modules with coordinate points in three-dimensional space, training a molecule property predictor model to operate on a molecule vector field, and using an input module to support both forward and inverse workflows with specified output channels.

Across the independent claims, the central claim coverage is for an AI apparatus or medium that uses trained machine learning models with atom vectors and molecule vectors defined with coordinate points in three-dimensional space, coupled through an input module to perform both property prediction from input atoms and inverse output of atoms and 3D geometric structure from targeted properties, with defined user-facing output channels. Dependent inventive features further add approximate agreement comparison between atom and molecule vector fields, optional transformer-based mapping into a shared 3D coordinate system, knowledge-source constraints including chemistry expert information, and expanded structural reporting including coordinates, bond angles, and valence bond type.

Stated Advantages

Documented Applications

No documented applications found