Structure-based, ligand activity prediction using binding mode prediction information

Patent metadata

Field	Value
Publication number	US-12412638-B2
Application number	US-202117166483-A
Country	US
Kind code	B2
Filing date	Feb 3, 2021
Priority date	Feb 3, 2021
Publication date	Sep 9, 2025
Grant date	Sep 9, 2025

Abstract

A system and method for structure-based, small molecule activity prediction using binding mode prediction information. Binding scores between ligands and target molecules, (e.g. proteins, RNA, DNA, lipids, sugars) are first generated using molecular docking. A first machine learned deep neural network (DNN) model is developed using data representing the molecular ligand-target pair 3D structures and docking features to predict binding modes. Using transfer learning, weights of layers learned in the first machine learned model are used as weights in layers of a second machine learned DNN model used to more accurately improve the performance of activity prediction of the second machine learned model. For a target newly paired ligand-target complex, the method further implements a binding mode selector for selecting one or more particular binding poses for input to the activity prediction model for use in activity mode prediction of an activity of the target paired ligand-protein complex.

First claim

What is claimed is:

1. A computer-implemented method of predicting an activity of a ligand against a target molecule, the method comprising:
receiving, at a hardware processor, a representation of a ligand molecule and a target molecule forming a ligand-target molecule pair structure for which an activity is to be determined;
obtaining, at the hardware processor, one or more binding modes corresponding to the received ligand-target molecule pair structure;
determining, using a first neural network running at the hardware processor, a confidence metric characterizing a correctness of each of the obtained one or more binding modes, said first neural network comprising a deep neural network model trained to predict a characterizing confidence metric using labels representing a closeness of binding modes corresponding to training sets of ligand-target molecule pair structures to a reference set of binding modes;
selecting, using the hardware processor, one or more binding modes based on their corresponding characterizing metrics;
inputting, to a second neural network running at the hardware processor, as input features, the selected one or more binding modes, said second neural network comprising a deep neural network of a compatible structure as said first neural network, said second neural network trained to predict an activity based on training sets of ligand-target molecule pair structures and corresponding labels representing their known activities;
determining, using the second neural network, a prediction of an activity for said ligand-target molecule pair structure; and
outputting, by the hardware processor, the activity prediction for said ligand-target molecule pair structure, wherein said output activity prediction is formulated as a classification or a regression with improved accuracy.

2. The method as claimed in claim 1 , wherein the second neural network model comprises:
one or more pre-trained processing layers having weights obtained from the trained first neural network; and
additional task-specific, trainable layers for activity prediction.

3. The method as claimed in claim 1 , wherein the one or more binding modes are obtained using one of: a docking program or a database of crystal structures.

4. The method as claimed in claim 3 , wherein said selecting one or more binding modes based on their corresponding characterizing metric comprises one or more of:
selecting a top-ranked pose predicted by a docking scoring function;
selecting a binding mode having a highest confidence metric as predicted from an output of the first neural network;
selecting a pre-determined number of top binding modes above a confidence metric threshold as determined from the output of the first neural network;
selecting one or more ligand-target molecule binding modes that are, or are close to, experimentally determined structures of same or closely related ligand-target molecule pairs; or
selecting a binding mode based on a consensus achieved by comparison of binding modes generated by other means.

5. The method as claimed in claim 1 , further comprising training said first neural network and second neural network using datasets comprising data corresponding to one or more of: a single protein target, targets within a protein family, or a number of targets sampled across multiple protein families, additional types of molecular targets and a set of associated ligands.

6. The method as claimed in claim 1 , wherein said first neural network deep neural network model is trained using training data constructed as a training set of binding poses in which the binding pose for an active ligand from a crystal co-complex for a target is the same in orthologs or homologs for which that active ligand is also active.

7. The method as claimed in claim 1 , further comprising:
featurizing a selected binding mode pose by representing the binding mode pose as a voxel image or a graph; and
inputting featurized selected binding mode poses as said input features for training said second neural network model.

8. A non-transitory computer readable medium comprising instructions that, when executed by at least one hardware processor, configure the at least one hardware processor to:
receive a representation of a ligand molecule and a target molecule forming a ligand-target molecule pair structure for which an activity is to be determined;
obtain one or more binding modes corresponding to the received ligand-target molecule pair structure;
determine, using a first neural network, a confidence metric characterizing a correctness of each of the obtained one or more binding modes, said first neural network comprising a deep neural network model trained to predict a characterizing confidence metric using labels representing a closeness of binding modes corresponding to training sets of ligand-target molecule pair structures to a reference set of binding modes;
select one or more binding modes based on their corresponding characterizing metrics;
input, to a second neural network, as input features, the selected one or more binding modes, said second neural network comprising a deep neural network of a compatible structure as said first neural network, said second neural network trained to predict an activity based on training sets of ligand-target molecule pair structures and corresponding labels representing their known activities;
determine, using the second neural network, a prediction of an activity for said ligand-target molecule pair structure; and
output the activity prediction for said ligand-target molecule pair structure, wherein said output activity prediction is formulated as a classification or a regression with improved accuracy.

9. The non-transitory computer readable medium as claimed in claim 8 , wherein the second neural network model comprises:
one or more pre-trained processing layers having weights obtained from the trained first neural network; and
additional task-specific, trainable layers for activity prediction.

10. The non-transitory computer readable medium as claimed in claim 8 , wherein the one or more binding modes are obtained using one of: a docking program or a database of crystal structures.

11. The non-transitory computer readable medium as claimed in claim 10 , wherein to select one or more binding modes based on their corresponding characterizing metric, the at least one hardware processor is further configured to:
select a top-ranked pose predicted by a docking scoring function;
select a binding mode having a highest confidence metric as predicted from an output of the first neural network;
select a pre-determined number of top binding modes above a confidence metric threshold as determined from the output of the first neural network;
select one or more ligand-target molecule binding modes that are, or are close to, experimentally determined structures of same or closely related ligand-target molecule pairs; or
select a binding mode based on a consensus achieved by comparison of binding modes generated by other means.

12. The non-transitory computer readable medium as claimed in claim 8 , wherein the at least one hardware processor is further configured to:
train said first neural network and second neural network using datasets comprising data corresponding to one or more of: a single protein target, targets within a protein family, or a number of targets sampled across multiple protein families, additional types of molecular targets and a set of associated ligands.

13. The non-transitory computer readable medium as claimed in claim 8 , wherein the at least one hardware processor is further configured to:
train said fi

Assignees

• Int Business Machines Corportion
• IBM

Inventors

• Morrone Joseph Anthony
• Weber Jeffrey Kurt
• Bagchi Sugato
• Cornell Wendy Dawn

Classifications

• G06N3/09

  Supervised learning · CPC title

• G06N3/0464

  Convolutional networks [CNN, ConvNet] · CPC title

• G06N3/096

  Transfer learning · CPC title

• G06N3/045

  Combinations of networks · CPC title

• G06N3/08

  Learning methods · CPC title