Structure based predictive modeling

Abstract

Disclosed are methods for building a sequence activity model with reference to structural data, which model can be used to guide directed evolution of proteins having beneficial properties. Some embodiments use genetic algorithms and structural data to filter out uninformative data. Some embodiments use a support vector machine to train the sequence activity model. The filtering and training methods can generate a sequence activity model having higher predictive power than conventional modeling methods. Systems and computer program products implementing the methods are also provided.

What is claimed is: 
     
         1 . A method of conducting directed evolution, the method comprising:
 (a) receiving a data set having information from physical measurements of molecules, wherein the data set comprises the following information for each of a plurality of variant biomolecules: (i) activity of the variant biomolecule on a ligand in a binding site of the variant biomolecule, (ii) a sequence of the variant biomolecule, and (iii) one or more geometric parameters characterizing the geometry of the ligand in the binding site;   (b) filtering the data set to produce a filtered data subset by removing information for one or more of the variant biomolecules, wherein the filtering comprises testing the predictive power of sequence activity models trained with a plurality of selected data subsets, each selected data subset having information for a particular set of variant biomolecules removed from the data set of (a); and   (c) training an improved sequence activity model using the filtered data subset.   
     
     
         2 . The method of  claim 1 , wherein filtering the data set comprises removing at least one of the geometric parameters from the data set. 
     
     
         3 . The method of  claim 1 , wherein the filtering the data set is performed with a genetic algorithm. 
     
     
         4 . The method of  claim 3 , wherein the genetic algorithm varies thresholds for removing information associated with the geometric parameters for one or more of the variant biomolecules. 
     
     
         5 . The method of  claim 1 , further comprising applying the improved sequence activity model to identify one or more new biomolecule variants predicted by the improved sequence activity model to have activity meeting one or more criteria, wherein each of the one or more new biomolecule variants has a sequence that differs from the sequences of the biomolecule variants providing information for the data set of (a). 
     
     
         6 . The method of  claim 5 , wherein applying the improved sequence activity model to identify one or more new biomolecule variants comprises performing a genetic algorithm in which potential new biomolecule variants are evaluated using the improved sequence activity model as a fitness function. 
     
     
         7 . The method of  claim 5 , further comprising assaying the new biomolecule variants for activity. 
     
     
         8 . The method of  claim 5 , further comprising
 producing a structural model for each of the new biomolecule variants; and   using the structural models to generate geometric parameters for binding sites of the new biomolecule variants, wherein the geometric parameters characterize the geometry of the ligand in the binding sites of the new biomolecule variants.   
     
     
         9 . The method of  claim 1 , further comprising measuring the activity of the variant biomolecules by an in vitro assay. 
     
     
         10 . The method of  claim 1 , further comprising receiving structural models of biomolecule variants and determining the one or more geometric parameters using the structural models. 
     
     
         11 . The method of  claim 10 , wherein the structural models are homology models. 
     
     
         12 . The method of  claim 10 , wherein the homology models are prepared using physical structural measurement details of biomolecules. 
     
     
         13 . The method of  claim 12 , wherein the physical structural measurement details of biomolecules comprise three-dimensional positions of atoms obtained by NMR or x-ray crystallography. 
     
     
         14 . The method of  claim 10 , further comprising using a docker to determine the one or more geometric parameters. 
     
     
         15 . The method of  claim 1 , wherein the information for each of a plurality of variant biomolecules further comprises (iv) an interaction energy characterizing the interaction of the ligand in the binding site. 
     
     
         16 . The method of  claim 15 , further comprising using a docker to determine the interaction energy 
     
     
         17 . The method of  claim 1 , wherein the improved sequence activity model is obtained by a support vector machine, a multiple linear regression, a principal component regression, a partial least square regression, or a neural network. 
     
     
         18 . The method of  claim 17 , wherein the sequence activity model is obtained by a support vector machine. 
     
     
         19 . The method of  claim 1 , wherein the plurality of variant biomolecules comprises a plurality of enzymes. 
     
     
         20 . The method of  claim 19 , wherein the activity of the variant biomolecule on a ligand is the activity of an enzyme on a substrate. 
     
     
         21 . The method of  claim 20 , wherein the activity of an enzyme on a substrate comprises one or more features of a catalytic conversion of the substrate by the enzyme. 
     
     
         22 . The method of  claim 1 , further comprising using the improved sequence activity model to identifying one or more biomolecules having desired activity. 
     
     
         23 . The method of  claim 22 , further comprising synthesizing the biomolecules having desired activity. 
     
     
         24 . A computer program product comprising one or more computer-readable non-transitory storage media having stored thereon computer-executable instructions that, when executed by one or more processors of a computer system, cause the computer system to implement a method for conducting directed evolution, the method comprising:
 (a) receiving, by the computer system, a data set having information from physical measurements of molecules, wherein the data set comprises the following information for each of a plurality of variant biomolecules: (i) activity of the variant biomolecule on a ligand in a binding site of the variant biomolecule, (ii) a sequence of the variant biomolecule, and (iii) one or more geometric parameters characterizing the geometry of the ligand in the binding site;   (b) filtering, by the computer system, the data set to produce a filtered data subset by removing information for one or more of the variant biomolecules, wherein the filtering comprises testing the predictive power of sequence activity models trained with a plurality of selected data subsets, each selected data subset having information for a particular set of variant biomolecules removed from the data set of (a); and   (c) training, by the computer system, an improved sequence activity model using the filtered data subset.   
     
     
         25 . A computer system, comprising:
 one or more processors;   system memory; and   one or more computer-readable storage media having stored thereon computer-executable instructions that, when executed by the one or more processors, cause the computer system to implement a method for conducting directed evolution, the method comprising:
 (a) receiving a data set having information from physical measurements of molecules, wherein the data set comprises the following information for each of a plurality of variant biomolecules: (i) activity of the variant biomolecule on a ligand in a binding site of the variant biomolecule, (ii) a sequence of the variant biomolecule, and (iii) one or more geometric parameters characterizing the geometry of the ligand in the binding site; 
 (b) filtering the data set to produce a filtered data subset by removing information for one or more of the variant biomolecules, wherein the filtering comprises testing the predictive power of sequence activity models trained with a plurality of selected data subsets, each selected data subset having information for a particular set of variant biomolecules removed from the data set of (a); and 
 (c) training an improved sequence activity model using the filtered data subset.