System and method for the latent space optimization of generative machine learning models
Inventors
Bucher, Alwin • Kamuntavicius, Gintautas • Prat, Alvaro • Bastas, Orestis • Jocys, Zygimantas • Tal, Roy
Assignees
Interested in licensing this patent?
MTEC can help explore whether this patent might be available for licensing for your application.
Publication Number
US-11610139-B2
Publication Date
2023-03-21
Expiration Date
Abstract
A system and method for optimizing the latent space in generative machine learning models, and applications of the optimizations for use in the de novo generation of molecules for both ligand-based and pocket-based generation. The ligand-based optimizations comprise a tunable reward system based on a multi-property model and further define new measurable metrics: molecular novelty and uniqueness. The pocket-based optimizations comprise an initial multi-property optimization followed up by either a seed-based optimization or a relaxed-based optimization.
Core Innovation
The invention provides a computer system and methods for optimizing the latent space in generative machine learning models for de novo drug discovery. The system trains a plurality of machine learning models, where each machine learning model represents a particular property of a common object, and receives a seed object of the same type as the common object. It optimizes the seed object's properties based on the plurality of machine learning models to form a mixed objective.
The mixed objective is maximized by performing gradient descent on a latent vector encoded from the seed object. The disclosed approach supports latent exploration by using property models and discriminator models that contribute to the mixed objective, thereby guiding optimization in latent space. The latent vector is encoded from the seed object and used as the optimization variable during gradient descent.
In representative implementations, the property and discriminator models include cheminformatics-derived 3D models predicting synthetic accessibility and drug-likeness, as well as docking-based models associated with protein binding pockets and ligand poses, including bioactivity affinity based on IC50 values. Additional models determine whether a conformer’s relative pose is the correct docked pose, and a discriminator model discriminates between generated and real molecules, each contributing to the mixed objective used for latent space optimization.
Claims Coverage
The document includes two independent claims describing a shared latent-space optimization framework for generative machine learning models, using multiple property models to form a mixed objective that is maximized via gradient descent on a latent vector encoded from a seed object. Across the independent claims, the key inventive features are the multi-model property representation, mixed-objective formation, and latent-vector gradient descent maximization.
Mixed-objective latent space optimization via gradient descent
train a plurality of machine learning models, each machine learning model trained to represent a particular property of a common object; receive a seed object, wherein the seed object is the same type of object as the common object; optimize the seed object's properties based on the plurality of machine learning models, wherein the plurality of optimized seed object properties forms a mixed objective; and maximize the mixed objective by performing gradient descent on a latent vector encoded from the seed object.
Property models contributing to the mixed objective
the plurality of machine learning models include models trained to represent particular properties, including at least one 3D model trained using ground truths extracted from cheminformatics software to predict synthetic accessibility; at least one 3D model trained using ground truths extracted from cheminformatics software to predict drug-likeness; a machine-learning model trained on docked ligand poses in protein binding pockets and fine-tuned using public protein databases to output bioactivity affinity based on IC50 values; a machine-learning model trained on ligand docking poses in protein binding pockets to determine whether a conformer’s relative pose to a binding site is the correct docked pose; and a discriminator machine learning model that discriminates between generated and real molecules.
Both independent claims cover computer-implemented systems and corresponding methods that optimize a seed object in generative machine learning by combining multiple property-representing models into a mixed objective, then maximizing that objective using gradient descent on a latent vector encoded from the seed object. Dependent claim sets further specify example property/discriminator models as contributors to the mixed objective.
Stated Advantages
Documented Applications
de novo drug discovery, including latent space optimization in generative machine learning models for molecule generation guided by a mixed objective.
Interested in licensing this patent?