IP Library › Granted Patent US 12694953

Granted Patent B2

US 12694953 · App. 17/427,786 · Granted Jul 28, 2026

Systems, methods, and media for molecule design using machine learning mechanisms

Inventors: Neal W. Woodbury (Tempe, AZ); Alexander T. Taguchi (Cambridge, MA)

Assignee: ARIZONA BOARD OF REGENTS ON BEHALF OF ARIZONA STATE UNIVERSITY

G16C20/50 G16B40/10 G16B40/20

View Patent ↗

Loading inventors, assignments & file history…

Monitor This Case

Get email alerts when status or documents change.

Order Certified Copies

Most orders are placed with the USPTO same day — all within 24 business hours.

Order via The Patent Place →

Pre-filled with this patent's details

Quick Facts

Patent No.: US 12694953
App. No.: 17/427,786
Granted: Jul 28, 2026
Kind: B2

External Links

USPTO Patent Center↗ Full Patent PDF↗

Abstract

Mechanisms for molecule design using machine learning include: forming a first training set for a neural network using, for each of a first plurality of known molecules, a plurality of input values that represent the structure of the known molecule and a plurality of functional property values for the known molecule; training the neural network using the first training set; proposing a first plurality of proposed molecules, and predicting first predicted functional property values of the first plurality of proposed molecules that have the desired function property values; causing the first plurality of proposed molecules to be synthesized to form a first plurality of synthesized molecules; receiving first measured functional property values of the first plurality of synthesized molecules; and adding data regarding the first plurality of synthesized molecules to the first training set to form a second training set and retrain the neural network using the second training set.

Claims (82)

1 . A system for proposing molecules having desired functional property values, comprising:

a computerized synthesizer;

a memory; and

a hardware processor coupled to the computerized synthesizer and to the memory and configured to:

form a first training set for a neural network by:

selecting features of building block molecules to produce selected features and functional properties of a first plurality of proposed molecules that comprise the building block molecules to produce selected functional properties;

accessing a library that identifies known molecules A known made from the building block molecules;

forming a representation of the library in which each known molecule A known is shown as a matrix B known of representations of the building block molecules;

forming a matrix T of the values V of the selected features of all of the building block molecules in the library or all of features of the building block molecules;

multiplying each matrix B known by the matrix T to form a matrix C known of features;

linearizing each matrix C known to form vectors D known ; and

determining, for each vector D known , known values P known for corresponding known molecules A known that comprise the selected functional properties;

train the neural network using at least the vectors D known of the first training set to predict values P known of the selected function properties to produce a trained neural network;

propose a first plurality of proposed molecules A proposed , and predict values P proposed of the selected functional properties of the proposed molecules A proposed using the trained neural network;

synthesize selected molecules of the first plurality of proposed molecules A proposed to form a first plurality of synthesized molecules A synth using the computerized synthesizer;

receive first measured functional property values of the first plurality of synthesized molecules A synth ; and

add data regarding selected molecules of the first plurality of synthesized molecules A synth to the library used to form the first training set to form a second training set and retrain the neural network using the second training set.

2 . The system of claim 1 , wherein the representation of the library is based on one-hot representations of the building block molecules that form the known molecules A known , wherein the one-hot representations are formed using one-hot encoding of the building block molecules.

3 . The system of claim 2 , wherein the building block molecules are amino acids.

4 . The system of claim 1 , wherein the known molecules A known are peptides.

5 . The system of claim 1 , wherein the representation of the library is based on chemical properties of building block molecules that form the known molecules A known .

6 . The system of claim 1 , wherein the neural network includes an encoder layer based on chemical properties of building block molecules that form the known molecules A known .

7 . The system of claim 1 , wherein an iterative process is used to propose the first plurality of proposed molecules.

8 . The system of claim 7 , wherein the iterative process attempts to find a local maximum based on each of the first plurality of proposed molecules.

9 . The system of claim 1 , wherein the hardware processor is further configured to:

propose a second plurality of proposed molecules, and predict second predicted functional property values of the second plurality of proposed molecules that have desired function property values;

synthesize selected molecules of the second plurality of proposed molecules to form a second plurality of synthesized molecules using the computerized synthesizer;

receive second measured functional property values of the second plurality of synthesized molecules; and

determine whether the second measured functional property values a threshold amount different from a first measured functional property values.

10 . A method for proposing molecules having desired functional property values, comprising:

forming a first training set for a neural network by:

accessing a library that identifies known molecules A known made from the building block molecules;

forming a representation of the library in which each known molecule A known is shown as a matrix B known of representations of the building block molecules;

forming a matrix T of the values V of the selected features of all of the building block molecules in the library or all of features of the building block molecules;

multiplying each matrix B known by the matrix T to form a matrix C known of features;

linearizing each matrix C known to form vectors D known ; and

determining, for each vector D known , known values P known for corresponding known molecules A known that comprise the selected functional properties;

training the neural network using at least the vectors D known of the first training set using a hardware processor to predict values P known of the selected function properties to produce a trained neural network;

proposing a first plurality of proposed molecules A proposed , and predict values P proposed of the selected functional properties of the proposed molecules A proposed using the trained neural network;

synthesizing selected molecules of the first plurality of proposed molecules A proposed to form a first plurality of synthesized molecules A synth using a computerized synthesizer;

receiving first measured functional property values of the first plurality of synthesized molecules A synth ; and

adding data regarding selected molecules of the first plurality of synthesized molecules A synth to the library used to form the first training set to form a second training set and retrain the neural network using the second training set.

11 . The method of claim 10 , wherein the representation of the library is based on one-hot representations of the building block molecules that form the known molecules A known , wherein the one-hot representations are formed using one-hot encoding of the building block molecules.

12 . The method of claim 11 , wherein the building block molecules are amino acids.

13 . The method of claim 10 , wherein the known molecules A known are peptides.

14 . The method of claim 10 , wherein the representation of the library is based on chemical properties of building block molecules that form the known molecules A known .

15 . The method of claim 10 , wherein the neural network includes an encoder layer based on chemical properties of building block molecules that form the known molecules A known .

16 . The method of claim 10 , wherein an iterative process is used to propose the first plurality of proposed molecules.

17 . The method of claim 16 , wherein the iterative process attempts to find a local maximum based on each of the first plurality of proposed molecules.

18 . The method of claim 10 , further comprising:

propose a second plurality of proposed molecules, and predict second predicted functional property values of the second plurality of proposed molecules that have desired function property values;

synthesize selected molecules of the second plurality of proposed molecules to form a second plurality of synthesized molecules using the computerized synthesizer;

receive second measured functional property values of the second plurality of synthesized molecules; and

determine whether the second measured functional property values a threshold amount different from a first measured functional property values.

19 . A non-transitory computer-readable medium containing computer executable instructions that, when executed by a processor, cause the processor to perform a method proposing molecules having desired functional property values, the method comprising: