CN104732115A - Protein conformation optimization method based on simple space abstract convexity lower bound estimation - Google Patents

Abstract

A protein conformation optimization method based on simple space abstract convexity lower bound estimation comprises the following steps that according to a coarsness energy model, a Rosetta Score 3 is adopted as an optimized objective function, and an energy calculation model is converted into a dihedral angle optimized space energy model; through feature vector extraction, a high-dimensional dihedral angle optimization problem is converted into an actually operable Descartes space optimization problem; based on Karmarker projective transformation, a Descartes space energy model is converted into a nonlinear optimization problem constrained by unit simplex, and an abstract convexity lower bound supporting face is constructed in this way, and is updated; fragment assembly and a Monte Carlo algorithm are combined to obtain a series of metastable state conformation; finally, high-resolution protein conformation is obtained through a Refinement service provided by a Rosetta sever. The method is high in sampling efficiency, low in complexity and high in prediction precision.

Description

A kind of protein conformation optimization method based on concise model abstract convex Lower Bound Estimation

Technical field

The present invention relates to bioinformatics, computer application field, in particular a kind of protein conformation optimization method based on concise model abstract convex Lower Bound Estimation.

Background technology

Bioinformatics is a study hotspot of life science and computer science crossing domain.At present, according to Anfinsen hypothesis, directly from amino acid sequence, based on Potential Model, adopt global optimization method, the state of minimum energy of search molecular system, thus high flux, predict at an easy rate and the native conformation of peptide chain become one of most important research topic of bioinformatics.Low or the polypeptide (small protein of <10 residue) for sequence similarity, ab initio prediction method is unique selection.Ab initio prediction method must consider following two factors: (1) protein structure energy function; (2) conformational space searching method.First factor belongs to molecular mechanics problem in essence, mainly in order to calculate energy value corresponding to each protein structure.There are some comparatively effective structural energy functions at present, as: simple mesh model HP and actual force field model M M3, AMBER, CHARMM, GROMOS, DISCOVER, ECEPP/3 etc.; Second factor belongs to Global Optimal Problem in essence, by selecting a kind of suitable optimization method, carries out fast search to conformational space, obtains the conformation corresponding with a certain global minima energy.Wherein, protein conformation space optimization belongs to the NP-Hard problem that a class is difficult to resolve very much.2005, D.Baker pointed out in Science, and conformational space optimization method is a bottleneck factor of restriction protein ab initio prediction method precision of prediction.

Therefore, there is sampling efficiency, complexity and precision of prediction aspect Shortcomings in existing conformational space optimization method, needs to improve.

Summary of the invention

In order to the deficiency that the sampling efficiency overcoming existing protein conformation optimization method is lower, complexity is higher, precision of prediction is lower, the present invention proposes the protein conformation optimization method based on concise model abstract convex Lower Bound Estimation that a kind of sampling efficiency is higher, complexity is lower, precision of prediction is higher.

The technical solution adopted for the present invention to solve the technical problems is:

Based on a protein conformation optimization method for concise model abstract convex Lower Bound Estimation, described conformational space optimization method comprises the following steps:

1) according to coarseness energy model, adopt Knowledge based engineering Rosetta Score3 energy model as objective function,

Shown in (1), and initialization population:

$f_1=f_1({\stackrel{\&OverBar;}{x}}^1,{\stackrel{\&OverBar;}{x}}^2,...,{\stackrel{\&OverBar;}{x}}^{\stackrel{\&OverBar;}{N}})---\left(1\right)$

Wherein represent N, C, O and C _fthe sum of atom, represent the coordinate of i-th atom $({\stackrel{\&OverBar;}{x}}_1^i,{\stackrel{\&OverBar;}{x}}_2^i,{\stackrel{\&OverBar;}{x}}_3^i),i=\mathrm{1,2},...,\stackrel{\&OverBar;}{N};$

2) to 1) in objective function carry out model conversion:

2.1) coordinate transformation method is adopted, by computation model f ₁be converted into dihedral angle and optimize dimensional energy model f ₂:

Wherein for backbone dihedral angles vector, N _rESrepresent residue number, φ _i, ω _irepresent i-th residue Atom C-N-C respectively _α-N, N-C _α-C-N, C _α-C-N-C _αdihedral angle;

2.2) hypervelocity shape recognition process is adopted, extract 4 unique points of protein structure, respectively: molecule barycenter CTD, from the atom CST that CTD is nearest, from CTD atom FCT farthest, from FCT atom FTF farthest, by calculating the mean distance of all atoms and four unique points in protein molecule coarseness skeleton pattern, distance variance, and range deviation index, 12 dimensional feature vectors of constitutive protein matter structure $\stackrel{\&RightArrow;}{M}=({\mathrm{\&mu;}}_1^{\mathrm{ctd}},{\mathrm{\&mu;}}_2^{\mathrm{ctd}},{\mathrm{\&mu;}}_3^{\mathrm{ctd}},{\mathrm{\&mu;}}_1^{\mathrm{cst}},{\mathrm{\&mu;}}_2^{\mathrm{cst}},{\mathrm{\&mu;}}_3^{\mathrm{cst}},{\mathrm{\&mu;}}_1^{\mathrm{fct}},{\mathrm{\&mu;}}_2^{\mathrm{fct}},{\mathrm{\&mu;}}_3^{\mathrm{fct}},{\mathrm{\&mu;}}_1^{\mathrm{ftf}},{\mathrm{\&mu;}}_2^{\mathrm{ftf}},{\mathrm{\&mu;}}_3^{\mathrm{ftf}},),$ Consider the factor of precision and complexity, select as protein structure characteristic coordinates, based on model (1), obtain following feature space energy model f ₃:

$f_3\left({\stackrel{\&OverBar;}{M}}_U\right)=f_3({\mathrm{\&mu;}}_1^{\mathrm{ctd}},{\mathrm{\&mu;}}_1^{\mathrm{cst}},{\mathrm{\&mu;}}_1^{\mathrm{fct}},{\mathrm{\&mu;}}_1^{\mathrm{ftf}})---\left(3\right)$

Wherein represent the mean distance of all atoms and unique point CTD, CST, FCT, FTF in protein coarseness skeleton pattern;

2.3) based on Karmarker photography conversion, model (3) is converted to the nonlinear optimal problem f under unit simplex S constraint ₄:

$f_4\left(x^{\&prime;}\right)\&equiv;f_4(x_1^{\&prime;},x_2^{\&prime;},x_3^{\&prime;},x_4^{\&prime;},x_5^{\&prime;}),s.t.x_i^{\&prime;}\&GreaterEqual;0,{\mathrm{\&Sigma;}}_{i=1}^5{x}^{\&prime;}=1,i=\mathrm{1,2,3,4,5}---\left(4\right)$

2.4) for f ₄, adopt strictly increasing to penetrate convex function transform method, increase a normal number at objective function item, the strictly increasing be converted under unit simplex constraint penetrates convex function f ₅(x ');

2.5) for K sampled point, for i-th sampled point x ^{' i}, calculate its abstract convex subdifferential, build f ₅(x ') is at sampled point x ^{' i}the support minorant h (x at place ^{' i}):

$h\left(x^{\&prime;i}\right)=f_5\left(x^{\&prime;i}\right)\min \{\frac{x_1^{\&prime;}}{x_1^{\&prime;i}},...,\frac{x_{K+1}^{\&prime;}}{x_{K+1}^{\&prime;i}}\}---\left(5\right)$

2.6) max-min piecewise linearity energy model f is set up ₆(x '):

f ₆(x′)＝maxh(x ^′i),i＝1,2,…,K (6)

2.7) consider that K+1 ties up support vector matrix L:

Wherein $l^i=(\frac{f\left(x^{\&prime;i}\right)}{x_1^{\&prime;i}},\frac{f\left(x^{\&prime;i}\right)}{x_2^{\&prime;i}},...,\frac{f\left(x^{\&prime;i}\right)}{x_{K+1}^{\&prime;i}})$ For support vector;

2.8) N-ary tree is set up to preserve Lower Bound Estimation information;

3) build concise model and underestimate model:

3.1) support vector is set up to each conformation in initial population;

3.2) find out in N-ary tree and do not satisfy condition leaf node, with build support vector replace;

3.3) judge whether the node after replacing meets meet, then retain this node, do not meet, then delete;

4) search procedure is performed:

4.1) end condition is set;

4.2) in population, select two different individualities at random;

4.3) produce new individuality by cross and variation: in parent individuality selected by Stochastic choice, the fragment of equal length exchanges, then fragment assembling is done to it, generate new individuality;

4.4) judge which region of search newly-generated individuality drops on;

4.5) its Lower Bound Estimation value E ' is calculated _c;

4.6) E ' is calculated _cthe value of delta 1 of the energy value of the individuality less with energy value in selected parent individuality, if δ is 1>0, then jumps out this and calculates, if δ is 1<0, calculate its true energy E _c;

4.7) E is calculated _cthe value of delta 2 of the energy value of the individuality less with energy value in selected parent individuality, if δ is 2<0, then replaces with it individuality that in population, selected parent energy value is higher;

5) judge whether to meet end condition, as no, turn 4.2); In this way, then terminate.

Further, described end condition is that iterations reaches settings.

Technical conceive of the present invention is: based on the framework of genetic algorithm (GA), with Rosetta Score3 for optimization object function, based on the expression model of coarseness, is that dihedral angle optimizes dimensional energy model by energy balane model conversion; By characteristic vector pickup, higher-dimension dihedral angle optimization problem is converted to actual exercisable cartesian space optimization problem; Based on Karmarker projective transformation, further cartesian space energy model is converted to the nonlinear optimal problem under unit simplex constraint, structure abstract convex lower bound supporting surface like this, and upgrade, carry out guidance search by the lower bound information constantly tightened up, and reach the object reducing heat-supplied number of times; Binding fragment assembling and Monte Carlo algorithm obtain a series of metastable state conformation; Finally, the Refinement service provided by Rosetta server obtains high-resolution protein conformation.

Beneficial effect of the present invention is: sampling efficiency is higher, complexity is lower, precision of prediction is higher.

Accompanying drawing explanation

Fig. 1 is that 2 dimensional region projects to unit simplex area schematic.

Fig. 2 is model conversion schematic diagram.

Fig. 3 optimizes the 2MKA protein three-dimensional structure comparison schematic diagram obtained.

Embodiment

Below in conjunction with accompanying drawing, the invention will be further described.

See figures.1.and.2, a kind of protein conformation optimization method based on concise model abstract convex Lower Bound Estimation (being called for short ACUE), comprises the following steps:

1) according to coarseness energy model, adopt Knowledge based engineering Rosetta Score3 energy model as objective function, shown in (1), and initialization population:

$f_1=f_1({\stackrel{\&OverBar;}{x}}^1,{\stackrel{\&OverBar;}{x}}^2,...,{\stackrel{\&OverBar;}{x}}^{\stackrel{\&OverBar;}{N}})---\left(1\right)$

Wherein represent the sum of N, C, O and CF atom, xi represents the coordinate of i-th atom $({\stackrel{\&OverBar;}{x}}_1^i,{\stackrel{\&OverBar;}{x}}_2^i,{\stackrel{\&OverBar;}{x}}_3^i),i=\mathrm{1,2},...,\stackrel{\&OverBar;}{N};$

2) to 1) in objective function carry out model conversion:

2.1) coordinate transformation method is adopted, by computation model f ₁be converted into dihedral angle and optimize dimensional energy model f ₂:

Wherein for backbone dihedral angles vector, N _rESrepresent residue number, φ _i, ω _irepresent i-th residue Atom C-N-C respectively _α-N, N-C _α-C-N, C _α-C-N-C _αdihedral angle;

2.2) hypervelocity shape recognition process is adopted, extract 4 unique points of protein structure, respectively: molecule barycenter CTD, from the atom CST that CTD is nearest, from CTD atom FCT farthest, from FCT atom FTF farthest, by calculating the mean distance of all atoms and four unique points in protein molecule coarseness skeleton pattern, distance variance, and range deviation index, 12 dimensional feature vectors of constitutive protein matter structure $\stackrel{\&RightArrow;}{M}=({\mathrm{\&mu;}}_1^{\mathrm{ctd}},{\mathrm{\&mu;}}_2^{\mathrm{ctd}},{\mathrm{\&mu;}}_3^{\mathrm{ctd}},{\mathrm{\&mu;}}_1^{\mathrm{cst}},{\mathrm{\&mu;}}_2^{\mathrm{cst}},{\mathrm{\&mu;}}_3^{\mathrm{cst}},{\mathrm{\&mu;}}_1^{\mathrm{fct}},{\mathrm{\&mu;}}_2^{\mathrm{fct}},{\mathrm{\&mu;}}_3^{\mathrm{fct}},{\mathrm{\&mu;}}_1^{\mathrm{ftf}},{\mathrm{\&mu;}}_2^{\mathrm{ftf}},{\mathrm{\&mu;}}_3^{\mathrm{ftf}},),$ Consider the factor of precision and complexity, select as protein structure characteristic coordinates, based on model (1), obtain following feature space energy model f ₃:

$f_3\left({\stackrel{\&OverBar;}{M}}_U\right)=f_3({\mathrm{\&mu;}}_1^{\mathrm{ctd}},{\mathrm{\&mu;}}_1^{\mathrm{cst}},{\mathrm{\&mu;}}_1^{\mathrm{fct}},{\mathrm{\&mu;}}_1^{\mathrm{ftf}})---\left(3\right)$

Wherein represent the mean distance of all atoms and unique point CTD, CST, FCT, FTF in protein coarseness skeleton pattern;

2.3) based on Karmarker photography conversion, model (3) is converted to the nonlinear optimal problem f under unit simplex S constraint ₄:

$f_4\left(x^{\&prime;}\right)\&equiv;f_4(x_1^{\&prime;},x_2^{\&prime;},x_3^{\&prime;},x_4^{\&prime;},x_5^{\&prime;}),s.t.x_i^{\&prime;}\&GreaterEqual;0,{\mathrm{\&Sigma;}}_{i=1}^5{x}^{\&prime;}=1,i=\mathrm{1,2,3,4,5}---\left(4\right)$

2.4) for f ₄, adopt strictly increasing to penetrate convex function transform method, increase a normal number at objective function item, the strictly increasing be converted under unit simplex constraint penetrates convex function f ₅(x ');

2.5) for K sampled point, for i-th sampled point x ^{' i}, calculate its abstract convex subdifferential, build f ₅(x ') is at sampled point x ^{' i}the support minorant h (x at place ^{' i}):

$h\left(x^{\&prime;i}\right)=f_5\left(x^{\&prime;i}\right)\min \{\frac{x_1^{\&prime;}}{x_1^{\&prime;i}},...,\frac{x_{K+1}^{\&prime;}}{x_{K+1}^{\&prime;i}}\}---\left(5\right)$

2.6) max-min piecewise linearity energy model f is set up ₆(x '):

f ₆(x′)＝maxh(x ^′i),i＝1,2,…,K (6)

2.7) consider that K+1 ties up support vector matrix L:

Wherein $l^i=(\frac{f\left(x^{\&prime;i}\right)}{x_1^{\&prime;i}},\frac{f\left(x^{\&prime;i}\right)}{x_2^{\&prime;i}},...,\frac{f\left(x^{\&prime;i}\right)}{x_{K+1}^{\&prime;i}})$ For support vector;

2.8) N-ary tree is set up to preserve Lower Bound Estimation information;

3) build concise model and underestimate model:

3.1) support vector is set up to each conformation in initial population;

3.2) find out in N-ary tree and do not satisfy condition leaf node, with build support vector replace;

3.3) judge whether the node after replacing meets meet, then retain this node, do not meet, then delete;

4) search procedure is performed:

4.1) end condition (as iterations reaches settings) is set;

4.2) in population, select two different individualities at random;

4.3) produce new individuality by cross and variation: in parent individuality selected by Stochastic choice, the fragment of equal length exchanges, then fragment assembling is done to it, generate new individuality;

4.4) judge which region of search newly-generated individuality drops on;

4.5) its Lower Bound Estimation value E ' is calculated _c;

4.6) E ' is calculated _cthe value of delta 1 of the energy value of the individuality less with energy value in selected parent individuality, if δ is 1>0, then jumps out this and calculates, if δ is 1<0, calculate its true energy E _c;

4.7) E is calculated _cthe value of delta 2 of the energy value of the individuality less with energy value in selected parent individuality, if δ is 2<0, then replaces with it individuality that in population, selected parent energy value is higher;

5) judge whether to reach end condition, as no, turn 4.2); In this way, then terminate.

The present embodiment is with the PDB ID protein that is 2MKA for embodiment, and a kind of protein conformation optimization method (ACUE) based on concise model abstract convex Lower Bound Estimation comprises the following steps:

1) according to coarseness energy model, adopt Knowledge based engineering Rosetta Score3 energy model as objective function, shown in (1), and initialization population: 100 initial population individualities are set in this example, random fragment assembling once generates initial configurations, as initial population respectively;

$f_1=f_1({\stackrel{\&OverBar;}{x}}^1,{\stackrel{\&OverBar;}{x}}^2,...,{\stackrel{\&OverBar;}{x}}^{\stackrel{\&OverBar;}{N}})---\left(1\right)$

Wherein represent the sum of N, C, O and CF atom, xi represents the coordinate of i-th atom $({\stackrel{\&OverBar;}{x}}_1^i,{\stackrel{\&OverBar;}{x}}_2^i,{\stackrel{\&OverBar;}{x}}_3^i),i=\mathrm{1,2},...,\stackrel{\&OverBar;}{N};$

2) to 1) in objective function carry out model conversion:

2.1) coordinate transformation method is adopted, by computation model f ₁be converted into dihedral angle and optimize dimensional energy model f ₂:

Wherein for backbone dihedral angles vector, N _rESrepresent residue number, φ _i, ω _irepresent i-th residue Atom C-N-C respectively _α-N, N-C _α-C-N, C _α-C-N-C _αdihedral angle;

2.2) hypervelocity shape recognition process is adopted, extract 4 unique points of protein structure, respectively: molecule barycenter CTD, from the atom CST that CTD is nearest, from CTD atom FCT farthest, from FCT atom FTF farthest, by calculating the mean distance of all atoms and four unique points in protein molecule coarseness skeleton pattern, distance variance, and range deviation index, 12 dimensional feature vectors of constitutive protein matter structure $\stackrel{\&RightArrow;}{M}=({\mathrm{\&mu;}}_1^{\mathrm{ctd}},{\mathrm{\&mu;}}_2^{\mathrm{ctd}},{\mathrm{\&mu;}}_3^{\mathrm{ctd}},{\mathrm{\&mu;}}_1^{\mathrm{cst}},{\mathrm{\&mu;}}_2^{\mathrm{cst}},{\mathrm{\&mu;}}_3^{\mathrm{cst}},{\mathrm{\&mu;}}_1^{\mathrm{fct}},{\mathrm{\&mu;}}_2^{\mathrm{fct}},{\mathrm{\&mu;}}_3^{\mathrm{fct}},{\mathrm{\&mu;}}_1^{\mathrm{ftf}},{\mathrm{\&mu;}}_2^{\mathrm{ftf}},{\mathrm{\&mu;}}_3^{\mathrm{ftf}},),$ Consider the factor of precision and complexity, select as protein structure characteristic coordinates, based on model (1), obtain following feature space energy model f ₃:

$f_3\left({\stackrel{\&OverBar;}{M}}_U\right)=f_3({\mathrm{\&mu;}}_1^{\mathrm{ctd}},{\mathrm{\&mu;}}_1^{\mathrm{cst}},{\mathrm{\&mu;}}_1^{\mathrm{fct}},{\mathrm{\&mu;}}_1^{\mathrm{ftf}})---\left(3\right)$

Wherein represent the mean distance of all atoms and unique point CTD, CST, FCT, FTF in protein coarseness skeleton pattern;

2.3) based on Karmarker photography conversion, model (3) is converted to the nonlinear optimal problem f under unit simplex S constraint ₄:

$f_4\left(x^{\&prime;}\right)\&equiv;f_4(x_1^{\&prime;},x_2^{\&prime;},x_3^{\&prime;},x_4^{\&prime;},x_5^{\&prime;}),s.t.x_i^{\&prime;}\&GreaterEqual;0,{\mathrm{\&Sigma;}}_{i=1}^5{x}^{\&prime;}=1,i=\mathrm{1,2,3,4,5}---\left(4\right)$

2.4) for f ₄, adopt strictly increasing to penetrate convex function transform method, increase an enough large normal number at objective function item, the constant increased in this example is 800, and the strictly increasing be converted under unit simplex constraint penetrates convex function f ₅(x ');

2.5) for K sampled point, in this example, K gets population scale 100, for i-th sampled point x ^{' i}, calculate its abstract convex subdifferential, build f ₅(x ') is at sampled point x ^{' i}the support minorant h (x at place ^{' i}):

$h\left(x^{\&prime;i}\right)=f_5\left(x^{\&prime;i}\right)\min \{\frac{x_1^{\&prime;}}{x_1^{\&prime;i}},...,\frac{x_{K+1}^{\&prime;}}{x_{K+1}^{\&prime;i}}\}---\left(5\right)$

2.6) max-min piecewise linearity energy model f is set up ₆(x '):

f ₆(x′)＝maxh(x ^′i),i＝1,2,…,K (6)

2.7) consider that K+1 ties up support vector matrix L:

Wherein $l^i=(\frac{f\left(x^{\&prime;i}\right)}{x_1^{\&prime;i}},\frac{f\left(x^{\&prime;i}\right)}{x_2^{\&prime;i}},...,\frac{f\left(x^{\&prime;i}\right)}{x_{K+1}^{\&prime;i}})$ For support vector;

2.8) N-ary tree is set up to preserve Lower Bound Estimation information;

3) build concise model and underestimate model:

3.1) support vector is set up to each conformation in initial population;

3.2) find out in N-ary tree and do not satisfy condition leaf node, with build support vector replace;

3.3) judge whether the node after replacing meets meet, then retain this node, do not meet, then delete;

4) search procedure is performed:

4.1) end condition (as iterations reaches settings, this example is set to iterations 20000 times) is set;

4.2) in population, select two different individualities at random;

4.3) produce new individuality by cross and variation: in parent individuality selected by Stochastic choice, the fragment of equal length exchanges, then fragment assembling is done to it, generate new individuality;

4.4) judge which region of search newly-generated individuality drops on;

4.5) its Lower Bound Estimation value E ' is calculated _c;

4.6) E ' is calculated _cthe value of delta 1 of the energy value of the individuality less with energy value in selected parent individuality, if δ is 1>0, then jumps out this and calculates, if δ is 1<0, calculate its true energy E _c;

4.7) E is calculated _cthe value of delta 2 of the energy value of the individuality less with energy value in selected parent individuality, if δ is 2<0, then replaces with it individuality that in population, selected parent energy value is higher;

5) judge whether to meet end condition, as no, turn 4.2); In this way, then terminate.

The protein being 2MKA with PDB ID is embodiment, uses above method to obtain the nearly native state conformation solution of this protein, as shown in Figure 3.

What more than set forth is the excellent results that an embodiment that the present invention provides shows, obvious the present invention is not only applicable to above-described embodiment, do not depart from essence spirit of the present invention and do not exceed content involved by flesh and blood of the present invention prerequisite under can do many variations to it and implemented.

Claims (2)

1. based on a protein conformation optimization method for concise model abstract convex Lower Bound Estimation, it is characterized in that: described conformational space optimization method comprises the following steps:

1) according to coarseness energy model, adopt Knowledge based engineering Rosetta Score3 energy model as objective function, shown in (1), and initialization population:

$f_1=f_1({\stackrel{\&OverBar;}{x}}^1,{\stackrel{\&OverBar;}{x}}^2,...,{\stackrel{\&OverBar;}{x}}^{\stackrel{\&OverBar;}{N}})---\left(1\right)$

Wherein represent N, C, O and C _fthe sum of atom, represent the coordinate of i-th atom $({\stackrel{\&OverBar;}{x}}_1^i,{\stackrel{\&OverBar;}{x}}_2^i,{\stackrel{\&OverBar;}{x}}_3^i),i=\mathrm{1,2},...,\stackrel{\&OverBar;}{N};$

2) to 1) in objective function carry out model conversion:

2.1) coordinate transformation method is adopted, by computation model f ₁be converted into dihedral angle and optimize dimensional energy model f ₂:

Wherein for backbone dihedral angles vector, N _rESrepresent residue number, φ _i, ω _irepresent i-th residue Atom C-N-C respectively _α-N, N-C _α-C-N, C _α-C-N-C _αdihedral angle;

2.2) hypervelocity shape recognition process is adopted, extract 4 unique points of protein structure, respectively: molecule barycenter CTD, from the atom CST that CTD is nearest, from CTD atom FCT farthest, from FCT atom FTF farthest, by calculating the mean distance of all atoms and four unique points in protein molecule coarseness skeleton pattern, distance variance, and range deviation index, 12 dimensional feature vectors of constitutive protein matter structure $\stackrel{\&RightArrow;}{M}=({\mathrm{\&mu;}}_1^{\mathrm{ctd}},{\mathrm{\&mu;}}_2^{\mathrm{ctd}},{\mathrm{\&mu;}}_3^{\mathrm{ctd}},{\mathrm{\&mu;}}_1^{\mathrm{cst}},{\mathrm{\&mu;}}_2^{\mathrm{cst}},{\mathrm{\&mu;}}_3^{\mathrm{cst}},{\mathrm{\&mu;}}_1^{\mathrm{fct}},{\mathrm{\&mu;}}_2^{\mathrm{fct}},{\mathrm{\&mu;}}_3^{\mathrm{fct}},{\mathrm{\&mu;}}_1^{\mathrm{ftf}},{\mathrm{\&mu;}}_2^{\mathrm{ftf}},{\mathrm{\&mu;}}_3^{\mathrm{ftf}}),$ Consider the factor of precision and complexity, select as protein structure characteristic coordinates, based on model (1), obtain following feature space energy model f ₃:

$f_3=\left({\stackrel{\&RightArrow;}{M}}_U\right)=f_3({\mathrm{\&mu;}}_1^{\mathrm{ctd}},{\mathrm{\&mu;}}_1^{\mathrm{cst}},{\mathrm{\&mu;}}_1^{\mathrm{fct}},{\mathrm{\&mu;}}_1^{\mathrm{ftf}})---\left(3\right)$

Wherein represent the mean distance of all atoms and unique point CTD, CST, FCT, FTF in protein coarseness skeleton pattern;

2.3) based on Karmarker photography conversion, model (3) is converted to the nonlinear optimal problem f under unit simplex S constraint ₄:

$f_4\left(x^{\&prime;}\right)\&equiv;f_4(x_1^{\&prime;},x_2^{\&prime;},x_3^{\&prime;},x_4^{\&prime;},x_5^{\&prime;}),s.t.x_i^{\&prime;}\&GreaterEqual;0,{\mathrm{\&Sigma;}}_{i=1}^5{x}^{\&prime;}=1,i=\mathrm{1,2,3,4,5}---\left(4\right)$

2.4) for f ₄, adopt strictly increasing to penetrate convex function transform method, increase a normal number at objective function item, the strictly increasing be converted under unit simplex constraint penetrates convex function f ₅(x ');

2.5) for K sampled point, for i-th sampled point x ^{' i}, calculate its abstract convex subdifferential, build f ₅(x ') is at sampled point x ^{' i}the support minorant h (x at place ^{' i}):

$h\left(x^{\&prime;i}\right)=f_5\left(x^{\&prime;i}\right)\min \{\frac{x_1^{\&prime;}}{x_1^{\&prime;i}},...,\frac{x_{K+1}^{\&prime;}}{x_{K+1}^{\&prime;i}}\}---\left(5\right)$

2.6) max-min piecewise linearity energy model f is set up ₆(x '):

f ₆(x′)＝maxh(x ^′i),i＝1,2,…,K (6)

2.7) consider that K+1 ties up support vector matrix L:

Wherein $l^i=(\frac{f\left(x^{\&prime;i}\right)}{x_1^{\&prime;i}},\frac{f\left(x^{\&prime;i}\right)}{x_2^{\&prime;i}},...,\frac{f\left(x^{\&prime;i}\right)}{x_{K+1}^{\&prime;i}})$ For support vector;

2.8) N-ary tree is set up to preserve Lower Bound Estimation information;

3) build concise model and underestimate model:

3.1) support vector is set up to each conformation in initial population;

3.2) find out in N-ary tree and do not satisfy condition leaf node, with build support vector replace;

3.3) judge whether the node after replacing meets meet, then retain this node, do not meet, then delete;

4) search procedure is performed:

4.1) end condition is set;

4.2) in population, select two different individualities at random;

4.3) produce new individuality by cross and variation: in parent individuality selected by Stochastic choice, the fragment of equal length exchanges, then fragment assembling is done to it, generate new individuality;

4.4) judge which region of search newly-generated individuality drops on;

4.5) its Lower Bound Estimation value E ' is calculated _c;

4.6) E ' is calculated _cthe value of delta 1 of the energy value of the individuality less with energy value in selected parent individuality, if δ is 1>0, then jumps out this and calculates, if δ is 1<0, calculate its true energy E _c;

4.7) E is calculated _cthe value of delta 2 of the energy value of the individuality less with energy value in selected parent individuality, if δ is 2<0, then replaces with it individuality that in population, selected parent energy value is higher;

5) judge whether to meet end condition, as no, turn 4.2); In this way, then terminate.

2., as claimed in claim 1 based on the protein conformation optimization method of concise model abstract convex Lower Bound Estimation, it is characterized in that: described end condition is that iterations reaches settings.