Method and apparatus for identification and optimization of bioactive compounds using a neural network

• US 6,587,845 B1
• Filed: 02/15/2000
• Issued: 07/01/2003
• Est. Priority Date: 02/15/2000
• Status: Expired due to Fees

• Current Assignee: Benjamin B. Braunheim
• Original Assignee: Benjamin B. Braunheim
• Inventors: Braunheim, Benjamin B.
• Primary Examiner(s): Voeltz, Emanuel Todd
• Assistant Examiner(s): Booker, Kelvin
• Application Number:
  US09/504,407
• Time in Patent Office: 1,232 Days
• Field of Search: 706/20, 706/21
• US Class Current: 706/21
• CPC Class Codes: G16C 20/30 Prediction of properties of chemical compounds, compositions or mixtures
• CPC Class Codes: G16C 20/70 Machine learning, data mining or chemometrics

Abstract

A computational method for the discovery and design of therapeutically valuable bioactive compounds is presented. The method employed has successfully analyzed enzymatic inhibitors for their chemical properties through the use of a neural network and associated algorithms. This method is an improvement over the current methods of drug discovery which often employs a random search through a large library of synthesized chemical compounds or biological samples for bioactivity related to a specific therapeutic use. This time-consuming process is the most expensive portion of current drug discovery methods. The development of computational methods for the prediction of specific molecular activity will facilitate the design of novel chemotherapeutics or other chemically useful compounds. The novel neural network provided in the current invention is “trained” with the bioactivity of known compounds and then used to predict the bioactivity of unknown compounds.

Claims

1. A computerized neural network system for predicting the chemical activity of at least one molecule of interest comprising:

a) an input layer consisting of at least one neuron where input data is sent as a vector value;

b) a weight matrix where every entry in the form of an input vector is multiplied by a set weight and then sent to at least one hidden layer neuron;

c) a hidden layer consisting of at least one neuron such that when said input vector is multiplied by a set weight said hidden layer contains said weight matrix, said weight matrix having the dimensions n by m where n is the length of an input vector and m is the number of hidden layer neurons available;

d) an output layer consisting of at least one neuron where weight matrix data is sent before it is input into a transfer function;

e) a transfer function that is non-linear in form and is capable of taking any value generated by said output layer and returning a number between −

1 and 1 or another predetermined range;

f) a training process for said neural network such that said neural network can accurately approximate a free energy of binding of at least one known training molecule with an output from said output layer; and

g) a test process in which a trained neural network is used to predict a free energy of binding for said at least one molecule of interest;

 

wherein the physicochemical descriptor of said at least one molecule of interest is the quantum mechanical electrostatic potential of said at least one molecule of interest at the van der Waels surface of said at least one molecule of interest and, wherein said test process Includes the use of at least one adjuster molecule such that after said training process said neural network is used to predict a free energy of binding for said at least one adjuster molecule, said at least one adjuster molecule having a known free energy of binding and having been excluded from the set of molecules comprising the set of said of at least one known training molecule.

2. The neural network of claim 1, wherein said neural network is able to accurately predict the free energy of binding of said at least one adjuster molecule within 10%.

3. A computerized double neural network system for predicting the chemical activity of at least one molecule of interest comprising:

a) an outer neural network further comprising;

i) an outer network input layer consisting of at least one neuron where input data is sent as a vector value;

ii) an outer network weight matrix where every entry in the form of an input vector is multiplied by a set weight and then sent to at least one hidden layer neuron;

iii) an outer network hidden layer consisting of at least one neuron such that when said input vector Is multiplied by a set weight said hidden layer contains said weight matrix, said weight matrix having the dimensions n by m where n is the length of an input vector and m is the number of hidden layer neurons available;

iv) an outer network output layer consisting of at least one neuron where weight matrix data is sent before it is input into a transfer function;

v) an outer network transfer function that is non-linear in form and is capable of taking any value generated by said output layer and returning a number between −

1 and 1 or another predetermined range;

vi) a training process for said outer neural network such that said neural network can accurately approximate a free energy of binding of at least one known training molecule with an output from said output layer;

b) an inner neural network capable of receiving data from said outer neural network further comprising;

i) an inner network weight matrix where every entry in the form of an input vector is multiplied by a set weight and then sent to at least one hidden layer neuron;

ii) an inner network hidden layer consisting of at least one neuron such that when said input vector is multiplied by a set weight said hidden layer contains said weight matrix, said weight matrix having the dimensions n by m where n is the length of an input vector and m is the number of hidden layer neurons available;

iii) an inner network output layer consisting of at least one neuron where weight matrix data is sent before it is input into a transfer function said inner network output layer having an output value;

iv) an inner network transfer function that is non-linear in form and is capable of taking said output value generated by said output layer and returning a number between −

1 and 1 or another predetermined range;

v) a training process for said inner neural network such that said neural network can accurately approximate a free energy of binding of at least one known training molecule with an output from said output layer

vi) a test process in which a trained neural network is used to predict a free energy of binding for said at least one molecule of interest;

 

wherein said inner neural network is integrated to function with the data generated from said outer neural network such that the rules for said free energy of binding learned by said outer neural network are utilized by said inner neural network to model a quantum object such that said double neural network is used to predict the chemical characteristics of said quantum object, said quantum object describing a molecule with improved chemical properties of binding relative to said at least one molecule of interest;

wherein said outer network output layer is the input layer of said inner neural network;

wherein said outer network hidden layer includes an error term, said error term being used to calculate the correction terms for said outer network input layer such that the weights and biases of said double neural network are optimized; and

wherein the physicochemical descriptor of said at least one molecule of interest is the quantum mechanical electrostatic potential of said at least one molecule of interest at the van der Waals surface of said at least one molecule of interest.

4. The double neural network of claim 3, wherein said test process includes the use of at least one adjuster molecule such that after said outer network training process said neural network is used to predict a free energy of binding for said at least one adjuster molecule, said at least one adjuster molecule having a known free energy of binding and having been excluded from the set of molecules comprising the set of said of at least one known training molecule.

5. The double neural network of claim 4, wherein said double neural network is able to accurately predict the free energy of binding of said at least one adjuster molecule within 10%.

6. The double neural network of claim 3, wherein only the weights and biases of said outer network weight matrix are allowed to vary during the training of said double neural network.

7. The double neural network of claim 3, wherein a bias is added to said outer network hidden layer and said outer network output layer neurons such that all values are scaled before they are input into said outer network transfer function.

8. The double neural network of claim 3, wherein said outer network hidden layer is composed of 5 hidden layer neurons.

9. The double neural network of claim 3, wherein said inner network hidden layer is composed of 5 hidden layer neurons.

10. The double neural network of claim 3, wherein said double neural network is run through at least 100,000 iterations.

11. The double neural network of claim 3, wherein the learning rate of said outer neural network is 0.1.

12. The double neural network of claim 3, wherein the learning rate of said inner neural network is 0.1.

13. The double neural network of claim 3, wherein the momentum term of said outer neural network is 0.9.

14. The double neural network of claim 3, wherein the momentum term of said inner neural network is 0.9.

15. The double neural network of claim 3, wherein the quantum chemical data sent to said outer network input layer is a vector value derived from calculating the electrostatic potential of a molecule at the van der Waals surface.

16. The double neural network of claim 15, wherein said computer is coupled to a display device and there exists a means for presenting the chemical properties of said at least one molecule of interest on said display device.

17. The double neural network of claim 3, wherein the process for carrying out the elements of said double neural network for predicting the chemical activity of said at least one molecule of interest are contained in a computer, said computer being capable of receiving data and performing said training process and said testing process.

18. The double neural network of claim 17, wherein the chemical characteristics of said quantum object are in the form of a three dimensional representation, said three dimensional representation allowing the identification of the molecular features of said quantum object that said double neural network determined could altered to improve the chemical characteristics of said at least one molecule of interest.

19. The double neural network of claim 3, wherein said at least one molecule of interest is selected from the group consisting of:

a) a pharmaceutical;

b) an enzyme;

c) a catalyst;

d) a polypeptide;

e) an amino acid derivative;

f) a carbohydrate;

g) a nucleotide;

h) a macromolecular compound;

i) an organic moiety of an alkyl, cycloalkyl, aryl, aralkyl or alkaryl group or a substituted or heterocyclic derivative thereof;

j) an industrial compound; and

k) a polymer.

20. The double neural network of claim 19, wherein said at least one molecule of interest is an enzyme.

21. The method of claim 20, wherein said at least one molecule of interest is selected from the group consisting of:

a) a pharmaceutical;

b) an enzyme;

c) a catalyst;

d) a polypeptide;

e) an amino acid derivative;

f) a carbohydrate;

g) a nucleotide;

h) a macromolecular compound;

i) an organic moiety of an alkyl, cycloalkyl, aryl, aralkyl or alkaryl group or a substituted or heterocyclic derivative thereof; and

j) an industrial compound; and

k) a polymer.

22. The double neural network of claim 3, wherein said output value is decreased by at least 1Δ

G/RT.

23. The double neural network of claim 3, wherein said output value is decreased by 3Δ

G/RT.

24. A computer implemented method for predicting the chemical activity of at least one molecule of interest by using a neural network comprising:

a) inputting data into an input layer consisting of at least one neuron where input data is sent as a vector value;

b) developing a weight matrix wherein every entry in the form of an input vector is multiplied by a set weight and then sent to at least one hidden layer neuron;

c) providing a hidden layer consisting of at least one neuron such that when said input vector is multiplied by a set weight said hidden layer contains said weight matrix, said weight matrix having the dimensions n by m where n is the length of an input vector and m is the number of hidden layer neurons available;

d) constructing an output layer consisting of at least one neuron where weight matrix data is sent before it is input into a transfer function;

e) utilizing a transfer function that is non-linear in form and is capable of taking any value generated by said output layer and returning a number between −

1 and 1;

f) employing a training process for said neural network such that said neural network can accurately approximate a free energy of binding of at least one known training molecule with an output from said output layer; and

 

employing a test process in which a trained neural network is used to predict a free energy of binding for said at least one molecule of interest wherein the physicochemical descriptor of said at least one molecule of interest is the quantum mechanical electrostatic potential of said at least one molecule of interest at the van der Waals surface of said at least one molecule of interest, wherein said test process includes the use of at least one adjuster molecule such that after said training process said neural network is used to predict a free energy of binding for said at least one adjuster molecule, said at least one adjuster molecule having a known free energy of binding and having been excluded from the set of molecules comprising the set of said of at least one known training molecule.

25. The method of claim 24, wherein said neural network is able to accurately predict the free energy of binding of said at least one adjuster molecule within 10%.

26. A computer implemented method for predicting the chemical activity of at least one molecule of interest by using a double neural network comprising:

a) utilizing an outer neural network further comprising;

i) an outer network input layer consisting of at least one neuron where input data is sent as a vector value;

ii) an outer network weight matrix where every entry in the form of an input vector is multiplied by a set weight and then sent to at least one hidden layer neuron;

iii) an outer network hidden layer consisting of at least one neuron such that when said input vector is multiplied by a set weight said hidden layer contains said weight matrix, said weight matrix having the dimensions n by m where n is the length of an input vector and m is the number of hidden layer neurons available;

iv) an outer network output layer consisting of at least one neuron where weight matrix data is sent before it is input into a transfer function;

v) an outer network transfer function that is non-linear in form and is capable of

b) taking any value generated by said output layer and returning a number between −

1 and 1;

i) an outer network training process for said neural network such that said neural network can accurately approximate a free energy of binding of at least one known training molecule with an output from said output layer;

c) providing an inner neural network capable of receiving data from said outer neural network further comprising;

i) an inner network weight matrix where every entry in the form of an input vector is multiplied by a set weight and then sent to at least one hidden layer neuron;

ii) an inner network hidden layer consisting of at least one neuron such that when said input vector is multiplied by a set weight said hidden layer contains said weight matrix, said weight matrix having the dimensions n by m where n is the length of an input vector and m is the number of hidden layer neurons available;

iii) an inner network output layer consisting of at least one neuron where weight matrix data is sent before it is input into a transfer function said inner network output layer having an output value;

iv) an inner network transfer function that is non-linear in form and is capable of taking any value generated by said output layer and returning a number between −

1 and 1;

v) an inner network training process for said neural network such that said neural network can accurately approximate a free energy of binding of at least one known training molecule with an output from said output layer;

vi) a test process in which a trained neural network is used to predict a free energy of binding for said at least one molecule of interest;

d) integrating said inner neural network to function with the data generated from said outer neural network such that the rules for said free energy of binding learned by said outer neural network are utilized by said inner neural network to model a quantum object such that said double neural network is used to predict the chemical characteristics of said quantum object, said quantum object describing a molecule with improved chemical properties of binding relative to said at least one molecule of interest;

e) constructing said outer network input layer such that said output layer of said outer neural network is the input layer of said inner neural network; and

f) providing said outer network hidden layer with an error term, said error term being used to calculate the correction terms for said outer network input layer such that the weights and biases of said double neural network are optimized;

 

wherein the physicochemical descriptor of said at least one molecule of interest is the quantum mechanical electrostatic potential of said at least one molecule of interest at the van der Waals surface of said at least one molecule of interest.

27. The double neural network of claim 26, wherein said test process includes the use of at least one adjuster molecule such that after said outer network training process said neural network is used to predict a free energy of binding for said at least one adjuster molecule, said at least one adjuster molecule having a known free energy of binding and having been excluded from the set of molecules comprising the set of said of at least one known training molecule.

28. The double neural network of claim 27, wherein said double neural network is able to accurately predict the free energy of binding of said at least one adjuster molecule within 10%.

29. The double neural network of claim 26, wherein only the weights and biases of said outer network weight matrix are allowed to vary during the training of said double neural network.

30. The double neural network of claim 26, wherein a bias is added to said outer network hidden layer and said outer network output layer neurons such that all values are scaled before they are input into said outer network transfer function.

31. The double neural network of claim 26, wherein said outer network hidden layer is composed of 5 hidden layer neurons.

32. The double neural network of claim 26, wherein said inner network hidden layer is composed of 5 hidden layer neurons.

33. The double neural network of claim 26, wherein said double neural network is run through at least 100,000 iterations.

34. The double neural network of claim 26, wherein the learning rate of said outer neural network is 0.1.

35. The double neural network of claim 26, wherein the learning rate of said inner neural network is 0.1.

36. The double neural network of claim 26, wherein the momentum term of said outer neural network is 0.9.

37. The double neural network of claim 26, wherein the momentum term of said inner neural network is 0.9.

38. The double neural network of claim 37, wherein said computer is coupled to a display device and there exists a means for presenting the chemical properties of said at least one molecule of interest on said display device.

39. The double neural network of claim 26, wherein the quantum chemical data sent to said outer network input layer is a vector value derived from calculating the electrostatic potential of a molecule at the van der Waals surface.

40. The double neural network of claim 26, wherein the process for carrying out the elements of said double neural network for predicting the chemical activity of said at least one molecule of interest are contained in a computer, said computer being capable of receiving data and performing said training process and said testing process.

41. The double neural network of claim 26, wherein said at least one molecule of interest is selected from the group consisting of:

a) a pharmaceutical;

b) an enzyme;

c) a catalyst;

d) a polypeptide;

e) an amino acid derivative;

f) a carbohydrate;

g) a nucleotide;

h) a macromolecular compound;

i) an organic moiety of an alkyl, cycloalkyl, aryl, aralkyl or alkaryl group or a substituted or heterocyclic derivative thereof;

j) an industrial compound; and

k) a polymer.

42. The double neural network of claim 26, wherein said output value is decreased by at least 1Δ

G/RT.

43. The double neural network of claim 26, wherein said output value is decreased by 3Δ

G/RT.

44. A computerized neural network system comprising a neural network having a first component trained to recognize binding energy for a first set of molecular descriptors based on geometric and/or electrostatic information and for a given binding energy returning a second set of the molecular descriptors through a second component of the network.

45. A computerized double neural network system comprising a trained neural network for predicting binding potency for a chemotherapeutic agent with a target molecule, the network having an input layer, and the network being coupled to an output layer of an outer neural network comprising one or more layers so that the output of the output layer of the outer neural network is the input to the input layer of the inner neural network.

46. The system of claim 45 wherein the chemotherapeutic agent is an inhibitor and the molecule is an enzyme.

47. A computer implemented method comprising providing a neural network having a first component trained to recognize binding energy for a first set of molecular descriptors based on geometric and/or electrostatic information and for a given binding energy returning a second set of the molecular descriptors through a second component of the network.

48. A computer implemented method comprising providing a trained neural network for predicting binding potency for a chemotherapeutic agent with a target molecule, the network having an input layer, coupling the network to an output layer of an outer neural network comprising one or more layers so that the output of the output layer of the outer neural network is the input to the input layer of the inner neural network.

50. The method of claim 48 wherein the chemotherapeutic agent is an inhibitor and the molecule is an enzyme.

49. A computer implemented method comprising providing a trained neural network for predicting binding potency for a chemotherapeutic agent with a target molecule, the network having an input layer coupled to an output layer of an outer neural network comprising one or more layers so that the output of the output layer of the outer neural network is the input to the input layer of the inner neural network, and inputting molecular descriptors based on geometric and/or electrostatic information into the input layer from the coupled outer layer.

51. A computer implemented method of customizing the binding features of a molecule of interest comprising:

providing a neural network comprising a first component trained to recognize binding energy for first set of molecular descriptors based on electrostatic and/or geometrical information, and for a given binding energy returning a second set of the molecular descriptors through a second component of the network;

selecting a molecule of interest and modifying it so that the resulting molecule has a set of molecular descriptors that more closely matches descriptors in the second set of descriptors returned by the second component of the network.

52. A computer implemented method of determining a set of molecular descriptors:

providing a neural network comprising an inner network trained to predict binding energy of a molecule of interest with a target molecule using a set of molecular descriptors based on geometric and/or electrostatic information for the molecule of interest the inner network having an input layer coupled to the output layer of an outer neural network for inputting molecular descriptors,

in the inner neural network, setting the binding energy for an unknown molecule of interest to a desired level; and

determining a set of molecular descriptors for an unknown molecule of interest by computing through the network a set of molecular descriptors that if output from the output layer of the outer neural network would yield a binding energy within a desired range of a predetermined binding energy set for the inner neural network.

53. The method of claim 52 wherein the target molecule comprises a protein having a binding site and the molecular descriptors are for an unknown target molecule that is a potential binding agent of the binding site and wherein the binding energy is set at least slightly above the binding energy of a known binding agent for the binding site.

54. The method of claim 53 wherein the protein is an enzyme and the binding agent is an inhibitor.

55. The method of claim 52 wherein there is a total of 5 inner and outer network layers.

56. The method of claim 52 wherein the molecular descriptors comprise electrostatic potential and geometric information.

57. The method of claim 52 wherein the binding energy level is set to a desired degree higher than the binding energy for a known molecule of interest and wherein the method further comprises determining the chemical structure of a molecule using the molecular descriptors for the unknown molecule of interest.

58. The method of claim 57 wherein the determined structure is derived from optimizing binding features in a known molecule of interest.

59. The method of claim 57 wherein the determined structure is a modification of a known molecule of interest having a known binding energy with the target molecule.