Modeling biological effects of molecules using molecular property models

• US 7,856,321 B2
• Filed: 12/14/2005
• Issued: 12/21/2010
• Est. Priority Date: 12/16/2004
• Status: Active Grant

• Current Assignee: VALO Health Incorporated
• Original Assignee: Numerate, Inc.
• Inventors: Lanza, Guido, Boardman, Paul, Duffy, Nigel P.
• Primary Examiner(s): Dejong; Eric S
• Application Number:
  US11/304,209
• Publication Number:
  US 20060161407A1
• Time in Patent Office: 1,833 Days
• Field of Search: None
• US Class Current: 702/19
• 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 20/00 ICT specially adapted for functional genomics or proteomics, e.g. genotype-phenotype associations
• CPC Class Codes: G16B 40/00 ICT specially adapted for biostatistics; ICT specially adapted for bioinformatics-related machine learning or data mining, e.g. knowledge discovery or pattern finding

Abstract

Method, apparatus, and article of manufacture for modeling the biological effects of a molecule are disclosed. A machine learning application may be configured to process a set of training examples regarding a biological property of interest. Once trained, the machine learning application may be configured to generate a prediction regarding a property of interest for a test molecule. Embodiments of the invention provide a meta-model configured to generate a biological effect prediction for a molecule based on the predictions generated for the test molecule by a plurality of molecular property models.

Claims

1. A method for determining whether one or more test molecules has a biological effect selected from the group consisting of an antibacterial effect, an antiviral effect and an anticancer effect, comprising:

training on a suitably programmed computer more than three different molecular property models using at least one of a first set of training data, wherein each trained molecular property model is configured to generate a molecular property model prediction regarding a property of interest of the one or more test molecules, wherein said molecular property models are used in determining different properties of interest distinct from the biological effect;

using machine learning techniques to train on a suitably programmed computer system a meta-model to predict the biological effect using a second set of training data, wherein a representation of at least one molecule in the second set of training data includes predictions for the one or more test molecules generated by said molecular property models; and

establishing a prediction, using the computer implemented trained meta-model, that one or more test molecules has the biological effect based on the molecular property model predictions.

2. The method of claim 1, after the more than three molecular property models are trained, comprising:

selecting the test molecule;

generating a representation of the test molecule for use by the plurality of molecular property models;

providing the representation of the test molecule to the plurality of molecular property models to obtain the molecular property model predictions.

3. The method of claim 2, after the meta-model is trained, comprising:

providing the molecular property model predictions obtained from each of the molecular property models to the meta-model; and

generating a biological effect prediction for the test molecule.

4. The method of claim 1, wherein the first set of training data comprises a set of individual training examples, and wherein each example represents at least one molecule, and wherein each training example comprises a representation of the at least one molecule and a value for the property of interest modeled by at least one respective molecular property model.

5. The method of claim 4, wherein training the more than three molecular property models, comprises:

generating a form of each training example for use by a machine learning algorithm; and

performing the machine learning algorithm using the set of individual training examples to generate a trained one of the plurality of molecular property models.

11. The method of claim 1, wherein one or more of said molecular property models is a single target activity model.

12. The method of claim 11, wherein the single target activity model is configured to predict activity against a non-human target and the meta-model is trained to predict a human biological effect.

13. The method of claim 11, wherein the single target activity model is configured to predict activity against a human target and the meta-model is trained to predict a non-human biological effect.

14. The method of claim 1, wherein one or more of said molecular property models is known not to be associated with the biological effect.

15. The method of claim 1, further comprising:

performing a laboratory experiment for the biological effect on the one or more test molecules predicted to have the biological effect.

6. A method for determining whether one or more test molecules has a biological effect selected from the group consisting of an antibacterial effect, an antiviral effect and an anticancer effect, comprising:

training on a suitably programmed computer more than three different single target activity models using at least one of a first set of training data, wherein each trained single target activity model is configured to generate a molecular property model prediction regarding a property of interest of the one or more test molecules, wherein said single target activity models are used in determining different properties of interest distinct from the biological effect;

using machine learning techniques to train on a suitably programmed computer the meta-model to predict the biological effect using a second set of training data, wherein a representation of at least one molecule in the second set of training data includes predictions for the one or more test molecules generated by said single target activity models; and

establishing a prediction, using the computer implemented trained meta-model, that one or more of the test molecules has the biological effect based on the molecular property model predictions.

7. The method of claim 6, after the plurality of single target activity models are trained, comprising:

selecting the test molecule;

generating a representation of the test molecule for use by the more than three single target activity models;

providing the representation of the test molecule to the plurality of single target activity models to obtain the single target activity model predictions.

8. The method of claim 7, after the meta-model is trained, comprising:

providing the single target activity model predictions obtained from each of the single target activity models to the meta-model; and

generating a biological effect prediction for the test molecule.

9. The method of claim 6, wherein the first set of training data comprises a set of individual training examples, and wherein each example represents at least one molecule, and wherein each training example comprises a representation of the at least one molecule and a value for the property of interest modeled by each respective single target activity model.

10. The method of claim 9, wherein training a plurality of single target activity models, comprises:

generating a form of each training example for use by a machine learning algorithm; and

performing the machine learning algorithm using the set of individual training examples to generate a trained one of the more than three single target activity models.

16. The method of claim 6, further comprising:

performing a laboratory experiment for the biological effect on the one or more test molecules predicted to have the biological effect.