Neural network methods to predict enzyme inhibitor or receptor ligand potency

• US 6,895,396 B2
• Filed: 01/06/2004
• Issued: 05/17/2005
• Est. Priority Date: 06/19/1998
• Status: Expired due to Fees

• Current Assignee: National Institutes of Health (NIH) (Government of the United States of America)
• Original Assignee: Albert Einstein College of Medicine of Yeshiva University
• Inventors: Schwartz, Steven D., Braunheim, Benjamin B., Schramm, Vern L.
• Primary Examiner(s): Knight, Anthony
• Assistant Examiner(s): Hirl, Joseph P.
• Application Number:
  US10/752,259
• Publication Number:
  US 20040148265A1
• Time in Patent Office: 497 Days
• Field of Search: 706/15, 706/21, 706/924, 706/16
• US Class Current: 706/15
• CPC Class Codes: G16B 15/00 ICT specially adapted for analysing two-dimensional or three-dimensional molecular structures, e.g. structural or functional relations or structure alignment
• CPC Class Codes: G16B 15/30 Drug targeting using structural data; Docking or binding prediction
• CPC Class Codes: G16H 50/20 for computer-aided diagnosis, e.g. based on medical expert systems
• CPC Class Codes: Y10S 128/925 Neural network
• CPC Class Codes: Y10S 706/924 Medical

Abstract

A new method to analyze and predict the binding energy for enzyme-transition state inhibitor interactions is presented. Computational neural networks are employed to discovery quantum mechanical features of transition states and putative inhibitors necessary for binding. The method is able to generate its own relationship between the quantum mechanical structure of the inhibitor and the strength of binding. Feed-forward neural networks with back propagation of error can be trained to recognize the quantum mechanical electrostatic potential at the entire van der Waals surface, rather than a collapsed representation, of a group of training inhibitors and to predict the strength of interactions between the enzyme and a group of novel inhibitors. The experimental results show that the neural networks can predict with quantitative accuracy the binding strength of new inhibitors. The method is in fact able to predict the large binding free energy of the transition state, when trained with less tightly bound inhibitors. The present method is also applicable to prediction of the binding free energy of a ligand to a receptor. The application of this approach to the study of transition state inhibitors and ligands would permit evaluation of chemical libraries of potential inhibitory, agonistic, or antagonistic agents. The method is amenable to incorporation in a computer-readable medium accessible by general-purpose computers.

Claims

1. A method for determining a free energy of binding of a potential inhibitor to an enzyme, comprising the steps of:

obtaining a structure and a free energy of binding to said enzyme for each of two or more enzyme substrates or inhibitors;

orienting said structures of said two or more enzyme substrates or inhibitors for maximum geometric coincidence with each other;

determining an electrostatic potential at each of more than one point on a van der Waals surface of each of said enzyme substrates or inhibitors;

thereafter, mapping each of said electrostatic potentials of each of said enzyme substrates or inhibitors onto a geometric surface of one of said two or more enzyme substrates or inhibitors, each of said two or more enzyme substrates or inhibitors being thereby described by an identical surface geometry but a different electrostatic potential surface, and each of said electrostatic potentials being described by positional information relating said electrostatic potentials to said geometric surface;

thereafter, inputting said electrostatic potentials, said positional information, and said known free energy of binding of one of said two or more enzyme substrates or inhibitors into a neural network;

thereafter, training said neural network until said neural network predicts said free energy of binding of said one of said two or more enzyme substrates or inhibitors;

repeating said steps of inputting and training for each of the remaining said two or more enzyme substrates or inhibitors to produce a trained network;

thereafter, determining a potential inhibitor electrostatic potential at each of more than one point on a van der Waals surface of said potential inhibitor, said potential inhibitor having a known structure and an unknown free energy of binding to said enzyme;

orienting said structure of said potential inhibitor for maximum geometric coincidence with said structures of said two or more enzyme substrates or inhibitors;

thereafter, mapping each of said electrostatic potentials of said potential inhibitor onto a geometric surface of one of said two or more enzyme substrates or inhibitors, said potential inhibitor having a surface geometry identical to that of said two or more enzyme substrates or inhibitors, but a different electrostatic potential surface, and each of said electrostatic potentials of said potential inhibitor being described by positional information relating said electrostatic potentials to said geometric surface;

thereafter, inputting said electrostatic potentials and said positional information of said electrostatic potentials of said potential inhibitor into said trained network; and

using said trained network to calculate a free energy of binding of said potential inhibitor to said enzyme.

2. The method of claim 1, wherein the neural network contains an input layer, a hidden layer, and an output layer, and is a feed forward network with back propagation of error that learns with momentum.

3. The method of claim 2, wherein the network uses between 4 and 12 hidden layer neurons, a learning rate between 0.1 and 0.5, a momentum term between 0.8 and 0.9, and between 5×

10³ and 1×

10⁶ learning iterations.

4. The method of claim 3, wherein the number of hidden layer neurons, number of iterations, learning rate, and momentum can be adjusted so that the network outputs a binding energy with 98% accuracy.

5. A computer readable medium, comprising:

computer-readable information;

said information capable of interacting with a computer to produce an output;

said output being a calculated free energy of binding of a potential inhibitor to an enzyme;

said output being calculated by;

orienting structures of two or more enzyme substrates or inhibitors for maximum geometric coincidence with each other;

each of said two or more enzyme substrates or inhibitors having a predetermined structure and a predetermined free energy of binding to said enzyme;

determining an electrostatic potential at each of more than one point on a van der Waals surface of each of said enzyme substrates or inhibitors;

thereafter, mapping each of said electrostatic potentials of each of said enzyme substrates or inhibitors onto a geometric surface of one of said two or more enzyme substrates or inhibitors, each of said two or more enzyme substrates or inhibitors being thereby described by an identical surface geometry but a different electrostatic potential surface, and each of said electrostatic potentials being described by positional information relating said electrostatic potentials to said geometric surface;

thereafter, inputting said electrostatic potentials, said positional information, and said known free energy of binding of one of said two or more enzyme substrates or inhibitors into a neural network;

thereafter, training said neural network until said neural network predicts said free energy of binding of said one of said two or more enzyme substrates or inhibitors;

repeating said steps of inputting and training for each of the remaining said two or more enzyme substrates or inhibitors to produce a trained network;

thereafter, determining a potential inhibitor electrostatic potential at each of more than one point on a van der Waals surface of said potential inhibitor, said potential inhibitor having a known structure and an unknown free energy of binding to said enzyme;

orienting said structure of said potential inhibitor for maximum geometric coincidence with said structures of said two or more enzyme substrates or inhibitors;

thereafter, mapping each of said electrostatic potentials of said potential inhibitor onto a geometric surface of one of said two or more enzyme substrates or inhibitors, said potential inhibitor having a surface geometry identical to that of said two or more enzyme substrates or inhibitors, but a different electrostatic potential surface, and each of said electrostatic potentials of said potential inhibitor being described by positional information relating said electrostatic potentials to said geometric surface;

thereafter, inputting said electrostatic potentials and said positional information of said electrostatic potentials of said potential inhibitor into said trained network; and

using said trained network to calculate a free energy of binding of said potential substrate to said enzyme.

6. The computer readable medium of claim 5, wherein the neural network contains an input layer, a hidden layer, and an output layer, and is a feed forward network with back propagation of error that learns with momentum.

7. The computer readable medium of claim 6, wherein the network uses between 4 and 12 hidden layer neurons, a learning rate between 0.1 and 0.5, a momentum term between 0.8 and 0.9, and between 5×

10³ and 1×

10⁶ learning iterations.

8. The computer readable medium of claim 7, wherein the number of hidden layer neurons, number of iterations, learning rate, and momentum can be adjusted so that the network outputs a binding energy with 98% accuracy.

9. A method for determining a free energy of binding of a potential substrate to an enzyme, comprising the steps of:

obtaining a structure and a free energy of binding to said enzyme for each of two or more enzyme substrates or inhibitors;

orienting said structures of said two or more enzyme substrates or inhibitors for maximum geometric coincidence with each other;

determining an electrostatic potential at each of more than one point on a van der Waals surface of each of said enzyme substrates or inhibitors;

thereafter, mapping each of said electrostatic potentials of each of said enzyme substrates or inhibitors onto a geometric surface of one of said two or more enzyme substrates or inhibitors, each of said two or more enzyme substrates or inhibitors being thereby described by an identical surface geometry but a different electrostatic potential surface, and each of said electrostatic potentials being described by positional information relating said electrostatic potentials to said geometric surface;

thereafter, inputting said electrostatic potentials, said positional information, and said known free energy of binding of one of said two or more enzyme substrates or inhibitors into a neural network;

thereafter, training said neural network until said neural network predicts said free energy of binding of said one of said two or more enzyme substrates or inhibitors;

repeating said steps of inputting and training for each of the remaining said two or more enzyme substrates or inhibitors to produce a trained network;

thereafter, determining a potential substrate electrostatic potential at each of more than one point on a van der Waals surface of said potential substrate, said potential substrate having a known structure and an unknown free energy of binding to said enzyme;

orienting said structure of said potential substrate for maximum geometric coincidence with said structures of said two or more enzyme substrates or inhibitors;

thereafter, mapping each of said electrostatic potentials of said potential substrate onto a geometric surface of one of said two or more enzyme substrates or inhibitors, said potential substrate having a surface geometry identical to that of said two or more enzyme substrates or inhibitors, but a different electrostatic potential surface, and each of said electrostatic potentials of said potential substrate being described by positional information relating said electrostatic potentials to said geometric surface;

thereafter, inputting said electrostatic potentials and said positional information of said electrostatic potentials of said potential substrate into said trained network; and

using said trained network to calculate a free energy of binding of said potential substrate to said enzyme.

10. The method of claim 9, wherein the neural network contains an input layer, a hidden layer, and an output layer, and is a feed forward network with back propagation of error that learns with momentum.

11. The method of claim 10, wherein the network uses between 4 and 12 hidden layer neurons, a learning rate between 0.1 and 0.5, a momentum term between 0.8 and 0.9, and between 5×

10³ and 1×

10⁶ learning iterations.

12. The method of claim 11, wherein the number of hidden layer neurons, number of iterations, learning rate, and momentum can be adjusted so that the network outputs a binding energy with 98% accuracy.

13. A computer readable medium, comprising:

computer-readable information;

said information capable of interacting with a computer to produce an output;

said output being a calculated free energy of binding of a potential substrate to an enzyme;

said output being calculated by;

orienting structures of two or more enzyme substrates or inhibitors for maximum geometric coincidence with each other;

each of said two or more enzyme substrates or inhibitors having a predetermined structure and a predetermined free energy of binding to said enzyme;

determining an electrostatic potential at each of more than one point on a van der Waals surface of each of said enzyme substrates or inhibitors;

thereafter, mapping each of said electrostatic potentials of each of said enzyme substrates or inhibitors onto a geometric surface of one of said two or more enzyme substrates or inhibitors, each of said two or more enzyme substrates or inhibitors being thereby described by an identical surface geometry but a different electrostatic potential surface, and each of said electrostatic potentials being described by positional information relating said electrostatic potentials to said geometric surface;

thereafter, inputting said electrostatic potentials, said positional information, and said known free energy of binding of one of said two or more enzyme substrates or inhibitors into a neural network;

thereafter, training said neural network until said neural network predicts said free energy of binding of said one of said two or more enzyme substrates or inhibitors;

repeating said steps of inputting and training for each of the remaining said two or more enzyme substrates or inhibitors to produce a trained network;

thereafter, determining a potential substrate electrostatic potential at each of more than one point on a van der Waals surface of said potential substrate, said potential substrate having a known structure and an unknown free energy of binding to said enzyme;

orienting said structure of said potential substrate for maximum geometric coincidence with said structures of said two or more enzyme substrates or inhibitors;

thereafter, mapping each of said electrostatic potentials of said potential substrate onto a geometric surface of one of said two or more enzyme substrates or inhibitors, said potential substrate having a surface geometry identical to that of said two or more enzyme substrates or inhibitors, but a different electrostatic potential surface, and each of said electrostatic potentials of said potential substrate being described by positional information relating said electrostatic potentials to said geometric surface;

thereafter, inputting said electrostatic potentials and said positional information of said electrostatic potentials of said potential substrate into said trained network; and

using said trained network to calculate a free energy of binding of said potential substrate to said enzyme.

14. The computer readable medium of claim 13, wherein the neural network contains an input layer, a hidden layer, and an output layer, and is a feed forward network with back propagation of error that learns with momentum.

15. The computer readable medium of claim 14, wherein the network uses between 4 and 12 hidden layer neurons, a learning rate between 0.1 and 0.5, a momentum term between 0.8 and 0.9, and between 5×

10³ and 1×

10⁶ learning iterations.

16. The computer readable medium of claim 15, wherein the number of hidden layer neurons, number of iterations, learning rate, and momentum can be adjusted so that the network outputs a binding energy with 98% accuracy.