Molecular interaction predictors

• US 7,894,995 B2
• Filed: 10/03/2006
• Issued: 02/22/2011
• Est. Priority Date: 02/16/2006
• Status: Active Grant

• Current Assignee: Zhigu Holdings Limited
• Original Assignee: Microsoft Corporation
• Inventors: Gomez, Manuel Jesus Reyes, Jojic, Nebojsa
• Primary Examiner(s): Skowronek; Karlheinz R
• Application Number:
  US11/538,405
• Publication Number:
  US 20070192036A1
• Time in Patent Office: 1,603 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 15/20 Protein or domain folding
• 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

Adaptive threading models for predicting an interaction between two or more molecules such as proteins are provided. The adaptive threading models have one or more learnable parameters that can be learned from all or some of the available data. The available data can include data relating to known interactions between the two or more molecules, the composition of the molecules and the geometry of the molecular complex.

Claims

1. A method executed on a processing unit, the method comprising:

providing an optimized set of weighted contact potentials, the set of optimized weighted contact potentials optimized utilizing one or more machine learning algorithms;

choosing a set of distances according to a structural template relating to a contact between a protein and a peptide, the set of distances defining minimum distances of the contact between amino acids of the protein and an amino acids of a peptide;

determining a score that rates the contact between the protein and the peptide, the score determined by evaluating the protein sequence and peptide sequence data according to the set of distances and the optimized set of weighted contact potentials;

inferring a hidden variable representing information about geometry of a protein-peptide complex to facilitate evaluating the contact between the protein and the peptide, wherein the inferring occurs by defining distances for the set of distances from structural data corresponding to the protein sequence; and

utilizing the hidden variable to influence the protein sequence.

2. The method of claim 1, wherein choosing the set of distances according to the structural template relating to the protein comprises inferring an identity of the structural template by making a Bayesian inference.

3. The method of claim 2, wherein the protein is obtained by mutating a naturally occurring protein.

4. The method of claim 1, wherein the one or more machine learning algorithms comprise an iterative optimization.

5. The method of claim 4, wherein the iterative optimization is iterative least squares.

6. The method of claim 1, wherein evaluating protein sequence and peptide sequence data according to the set of distances comprises utilizing an optimized soft step distance function.

7. The method of claim 6, further comprising optimizing the soft step distance function by using a gradient descent.

8. The method of claim 1, wherein the protein is a MHC molecule and the peptide is an amino acid sequence from about 8 to about 11 amino acids in length.

9. A system comprising:

a processing unit;

a memory coupled to the processing unit;

components executed on the processing unit including;

a prediction component providing an optimized set of weighted contact potentials, the set of optimized weighted contact potentials optimized utilizing one or more machine learning algorithms;

the prediction component choosing a set of distances according to a structural template relating to a contact between a protein and a peptide, the set of distances defining minimum distances of the contact between amino acids of the protein and amino acids of the peptide;

the prediction component determining a score that rates the contact between the protein and the peptide, the score determined by evaluating protein sequence and peptide sequence data according to the set of distances and the optimized set of weighted contact potentials;

an inference component inferring a hidden variable representing information about geometry of a protein-peptide complex to facilitate evaluating the contact between the protein and the peptide, wherein the inferring occurs by defining distances for the set of distances from structural data corresponding to the protein sequence; and

the prediction component utilizing the hidden variable to influence the protein sequence.

10. The system of claim 9, wherein choosing the set of distances according to the structural template relating to the protein comprises inferring an identity of the structural template by making a Bayesian inference.

11. The system of claim 10, wherein the protein is obtained by mutating a naturally occurring protein.

12. The system of claim 9, wherein the one or more machine learning algorithms comprise an iterative optimization.

13. The system of claim 12, wherein the iterative optimization is iterative least squares.

14. The system of claim 9, wherein evaluating the protein sequence and peptide sequence data according to the set of distances comprises utilizing a soft step distance function.

15. The system of claim 14, further comprising optimizing the soft step distance function by using a gradient descent.

16. The system of claim 9, wherein the protein is a MHC molecule and the peptide is an amino acid sequence from about 8 to about 11 amino acids in length.