Improved Protein Structure Prediction using Advanced Scorin - PowerPoint Presentation

Slide1

Improved Protein Structure Prediction using Advanced Scoring Function and Effective Sampling

Avdesh Mishra, Md Tamjidul Hoque{amishra2, thoque}@uno.edu

Presented By: Avdesh MishraDepartment of Computer ScienceSlide2

Protein structure in its native state gains lowest free energy

The figure showsenergy landscape with energy barrierssome narrow path to native state

Energy

Native state

Unfolded state

Low

High

Decoy

Energy barriers

Background

Dill KA,

MacCallum

JL. The protein-folding problem, 50 years on. Science. 2012 Nov 23;338(6110):1042-6.Slide3

-10

-20

-100

-110

-140

Scoring Function

Scores

Scoring Function

Dill KA,

MacCallum

JL. The protein-folding problem, 50 years on. Science. 2012 Nov 23;338(6110):1042-6.Slide4

Scoring Function

Scoring Function

Knowledge-based Scoring Function

Physics-based Scoring FunctionSlide5

Physical-based Scoring Function

Approximates the energy based on empirical molecular mechanics force fields that contains terms associated with:Bond lengths and bond anglesTorsion anglesVan der waals interactions and

Electrostatic interactionsThese force fields calculations are computationally expensive

thus, physical-based scoring functions are less successful

.Slide6

Knowledge-based Scoring Function

Empirical scoring function derived from the statistics of the known structure.They have been successful because theyProvide decent approximation

Computationally less expensiveEasy to useSupposed to inherently incorporate the forces that have not been developed or used by scientists currentlySlide7

HP

ASA

uPsi

u

Phi

3DIGARS3.0 Knowledge-based Scoring Function

Sequence-specific accessible surface area

Ubiquitous phi orientation angles

Ubiquitous psi orientation angles

Hydrophobic and hydrophilic properties

3DIGARS3.0

The model is mathematically expressed as

E

3DIGARS3.0

= E

HP

+ (w

1

× E

ASA

) + (w

2

×

E

uPhi

) + (w

3

×

E

uPsi

)Slide8

3DIGARS – HP Properties

Hydrophobic and polar interaction plays vital role in tertiary structure formation.The spatial conformation of the protein can be divided into at least three different layers as shown in the figure.

Thus, we

collect the statistics of hydrophobic vs hydrophobic, hydrophilic vs hydrophilic and hydrophobic vs hydrophilic interactions among the atoms and incorporate in our function

Mishra A.

and

Hoque

M. T., Three-Dimensional

Ideal Gas Reference State Based Energy Function. 

Current Bioinformatics

, 

12, pp.171-180, 2017.Slide9

3DIGARS Results

Decoy

Sets

DFIRE

RWplus

dDFIRE

DFIRE2.0

3DIGARS

No. of Targets

Moulder

19

(-2.97)

19

(-2.84)

18

(-2.74)

19

(-2.71)

19

(-2.998)

20

Rosetta

20

(-1.82)

20

(-1.47)

12

(-0.83)

22

(-1.76)

31

(-2.023)

58

I-

Tasser

49

(-4.02)

56

(-5.77)

48

(-5.03)

53

(-4.548)

53

(-4.036)

56

4state-reduced

6

(-3.48)

6

(-3.51)

7

(-4.15)

6

(-3.16)

6

(-3.37)

7

Fisa_casp3

4

(-4.80)

4

(-5.17)

4

(-4.83)

4

(-5.08)

5

(-4.31)

5

Lmds

7

(-

0.88)

7

(-1.03)

6

(-2.44)

7

(-0.71)

7

(-1.96)

10

Best scores

are colored in green and

competitive scores

are colored

in yellow. Values within the parenthesis are average z-scores of the native structures.Slide10

3DIGARS2.0 – ASA Properties

Here, we incorporate the information present in amino acid sequence.Using REGAd3p and DSSP tools we obtain the predicted and real (ASA) of each amino acid

.We then compute the frequency distribution of the

difference between real and predicted

ASA.

Next, these frequency distributions are converted into energy scores.

Iqbal, S., Mishra, A. &

Hoque

, T. (2015). Improved Prediction of Accessible Surface Area Results in Efficient Energy Function Application.

Journal of Theoretical Biology

380

, 380-391

.Slide11

3DIGARS2.0 Results

Decoy Sets

DFIRE

RWplus

dDFIRE

GOAP

3DIGARS

3DIGARS2.0

No. of Targets

Moulder

19

(-2.97)

19

(-2.84)

18

(-2.74)

19

(-3.58)

19

(-2.998)

19

(-2.6728)

20

Rosetta

20

(-1.82)

20

(-1.47)

12

(-0.83)

45

(-3.70)

31

(-2.023)

49

(-2.9871)

58

I-

Tasser

49

(-4.02)

56

(-5.77)

48

(-5.03)

45

(-5.36)

53

(-4.036)

56

(-4.2964)

56

Best scores

are colored in green and

competitive scores

are colored in yellow. Values within the parenthesis are average z-scores of the native structures.Slide12

3DIGARS3.0 –

uPhi, uPsi Orientation AnglesThe main idea isto incorporate 3D shape of the protein structure

Phi and Psi angles define the backbone structure of the protein.We mine the orientation angles for all atoms to capture the 3D shape.

Calculation of Phi and Psi angle involves the backbone atoms

C(O)

n-1-Nn-C(α)

n

-C(O)

n

and

N

n

-C(α)n-C(O)n-Nn+1 respectively.

The uPhi and uPsi angles are computedIn the similar manner as dihedral anglesBut, we compute them for all the atoms

instead of only backbone atoms.

Mishra, A., Iqbal, S. &

Hoque, M. T. (2016). Discriminate protein decoys from native by using a scoring function based on ubiquitous Phi and Psi angles computed for all atom. Journal of theoretical biology

398, 112-121.Slide13

3DIGARS3.0 Results

Decoy Sets

DFIRE

RWplus

dDFIRE

GOAP

3DIGARS

3DIGARS2

3DIGARS3

Total No. Targets

Moulder

19

(-

2.97)

19

(-

2.84)

18

(-

2.74)

19

(-

3.58)

19

(-

2.998)

19

(-

2.6728)

20

(-3.851)

20

Rosetta

20 (-

1.82)

20 (-

1.47)

12 (-

0.83)

45

(-

3.70)

31

(-

2.023)

49

(-

2.9871)

46

(-2.683)

58

I-

Tasser

49 (-

4.02)

56

(-

5.77)

48 (-

5.03)

45

(-

5.36)

53

(-

4.036)

56

(-

4.2964)

56

(-5.573)

56

4state-reduced

6 (-3.48)

6 (-3.15)

7

(-4.15)

7

(-4.38)

6

(-3.371)4 (-2.642)7 (-3.456)7Fisa_casp34 (-4.80)4 (-5.17)4 (-4.83)5 (-5.27)5 (-4.319)5 (-4.682)4 (-4.076)5Hg_structa12 (-1.97)

12 (-1.74)

16

(-1.33)

22

(-2.73)

12 (-1.914)

12 (-1.589)

28

(-3.678)

29

Ig_structal

0 (0.92)

0 (1.11)

26

(-1.02)

47

(-1.62)

0 (0.645)

0 (0.268)60 (-2.526)61Ig_structal_hires0 (0.17)0 (0.32)16 (-2.05)18 (-2.35)0 (-0.002)1 (0.030)20 (-2.378)20

Best scores

are colored in green and

competitive scores

are colored in yellow. Values within the parenthesis are average z-scores of the native structures.Slide14

Protein Structure Prediction

Goal: predict the 3D structure of a protein from its amino acid sequenceSlide15

Protein Structure Prediction: Methods

PSP

Methods

Homology Modeling

Ab Initio Modeling

Can be applied regardless of the presence or absence of the pre-existing structure

Can only be applied if similar structure are presentSlide16

Ab Initio PSP Components

Ab InitioScoring Function

Conformational Search

Ab Initio modeling conducts a conformational search under the guidance of a designed scoring function

.

A successful Ab Initio method depends on two factors:

Effective search method

which can quickly identify the low-energy state through

conformational search

Accurate energy function

which can differentiate the native structure that corresponds to the most thermodynamically stable state, compared to the possible decoy structures.Slide17

3DIGARS-PSP Method

Dataset of 4332 Protein Structures

Obtain Secondary Structure (SS) and

Φ,

Ψ Angles using DSSP

Generate Frequency Distribution of

Φ

,

Ψ

Angles and SS Types

Backbone Models

Initialize GA Population using Single Point Angular Mutation

Save Best Model in Memory

Select 5% Elite Models

Fill Rest Randomly

Perform Angular Mutation

@ 60

%

Perform

Memory Assisted Crossover

@ 70

%

Obtain Full Model using SCWRL4 and Calculate Fitness using 3DIGARS3.0

Gen < 2000

End

Save Models

Best ModelsSlide18

3DIGARS-PSP Results

Figure 1 |

Cysteine Protease Inhibitor (PDB ID: 1nyc); towards top – superposition of 3DIGARS-PSP model on native (initial seeds from Rosetta); towards bottom – superposition of top Rosetta model (based on TMScore

) on native.

Figure 2 |

E. Coli protein

(PDB ID:

1pohA); towards top – superposition of 3DIGARS-PSP model on native

(initial seeds from

Rosetta); towards bottom – superposition of top Rosetta model (based on

TMScore

) on native.

Natives

are shown in cyan and pink and Models are shown in red and

yellowNative Vs 3DIGARS-PSP

Native Vs RosettaSlide19

3DIGARS-PSP Results

Mishra, A., & Hoque, M. T. (2017). Next Generation Evolutionary Sampling and Energy Function Guided ab initio Protein Structure Prediction. Biophysical Journal, 112(3), 55a.Slide20

3DIGARS-PSP ResultsSlide21

o3DIGARS (Optimized 3DIGARS)

Thus, we further improve the accuracy of the 3DIGARS3.0 scoring function by following stepsCollected three different datasetsDataset used to train 3DIGARS scoring functionDataset used to train the 3DIGARS2.0 scoring functionDataset used to train and test the SpineX

server.Computed predicted ASA and Phi-Psi angles using Spiders2 program.Computed

real ASA and Phi-Psi angles using DSSP program.

Computed energy distribution from the difference between real and predicted

ASA and Phi-Psi values respectively, for each datasets using two different reference states.

Heffernan R,

Paliwal

K, Lyons J,

Dehzangi

A, Sharma A, Wang J,

Sattar

A, Yang Y, Zhou Y. Improving prediction of secondary structure, local backbone angles, and solvent accessible surface area of proteins by iterative deep learning. Scientific reports. 2015;5.Kabsch, W. and C. Sander, Dictionary of Protein Secondary Structure: Pattern Recognition of Hydrogen-Bonded and Geometrical Features. Biopolymers, 1983. 22: p.

2577-2637Slide22

o3DIGARS Feature Set

o3DIGARS Score3DIGARS Score

ASA Score Computed using REGAd3p

uPhi

and

uPsi

Scores

Total Features = 2

Phi Angles Scores Computed from 3 Different Data and 2 Difference Reference State

Total Features = 6

ASA Score Computed from 3 Different Data and 2 Different Reference State

Total Features = 6

ASA Score Computed from 3 Different Data, 2 Different Reference State and Segregated Based on SS

Total Features = 6

Psi Angles Scores Computed from 3 Different Data and 2 Difference Reference State

Total Features = 6Slide23

o3DIGARS

Collected 24 new features.We obtained sequence-specific torsion and ASA energy components for independent training and test datasets.The training dataset comprises of the structures submitted in CASP10 and CASP11 for PSP.Similarly, the

test dataset comprises of the structures submitted in CASP8 and CASP9 for PSP.These features were multiplied with weights and linearly combined together.

Then the optimization was carried out to optimize the objective functionOBJ_Fxn

= PCC_Avg

+ (-1.0 * Tmscore_Low_Energy_Modle_Avg)Slide24

o3DIGARS Optimization

Using Genetic Algorithm we optimized the combination of o3DIGARS energy components.o3DIGARS = E3DIGARS + w1 * EASA_REGAd3p + w2 * EuPhi + w3 * E

uPsi + w4 * EASA_3DIGARS_Data_Ref1 + w5 * E

ASA_SSD_Data_Ref1 + w6

* EASA_Spiders2_Data_Ref1

+ w7 * EASA_3DIGARS_Data_Ref2

+

w

8

* E

ASA_SSD_Data_Ref2

+ w9 * EASA_SpineX_Data_Ref2 + w10 * EASA_SS_3DIGARS_Data_Ref1

+ w11 * EASA_SS_SSD_Data_Ref1 + w12

* EASA_SS_Spiders2_Data_Ref1 + w13 * EASA_SS_3DIGARS_Data_Ref2

+ w14 * EASA_SS_SSD_Data_Ref2 + w15 *

EASA_SS_SpineX_Data_Ref2 + w16 * EPhi_SS_3DIGARS_Data_Ref1 + w

17 * EPhi_SS_SSD_Data_Ref1 + w18 * EPhi_SS_Spiders2_Data_Ref1

+ w19 * EPhi_SS_3DIGARS_Data_Ref2 + w20 *

EPhi_SS_SSD_Data_Ref2 + w21 * EPhi_SS_SpineX_Data_Ref2 +w22

* EPsi_SS_3DIGARS_Data_Ref1 + w23

*

E

Psi_SS_SSD_Data_Ref1

+

w

24

*

E

Psi_SS_Spiders2_Data_Ref1

+

w

25

*

E

Psi_SS_3DIGARS_Data_Ref2

+ w26

* EPsi_SS_SSD_Data_Ref2 + w27 * EPsi_SS_SpineX_Data_Ref2

Excluded those features with zero weights and ran the optimization again.Final o3DIGARS model:o3DIGARS = E3DIGARS + w

1 * EASA_REGAd3p + w2 * EuPsi + w3

* EASA_3DIGARS_Data_Ref1 + w4 * EASA_Spiders2_Data_Ref1 + w5 * EASA_SSD_Data_Ref2

+ w6 * EASA_SpineX_Data_Ref2 +

w7 * EASA_SS_3DIGARS_Data_Ref1 + w

8 * EASA_SS_SSD_Data_Ref1 + w9 * EASA_SS_Spiders2_Data_Ref1 + w10

* EPsi_SS_3DIGARS_Data_Ref1 + w11

* E

Psi_SS_SSD_Data_Ref1

+

w

12

* E

Psi_SS_Spiders2_Data_Ref1

Slide25

o3DIGARS ResultsPlot shows the comparison of Rosetta and I-

Tasser’s TMscore (first model) with the TMscore of the model selected by o3DIGARS. The larger number of points below the y = x line shows that o3DIGARS is able to select the model with low TMscore more accurately.Slide26

o3DIGARS ResultsPlot shows the comparison of Rosetta and I-

Tasser’s Tmscore (first model) with the TMscore of the model selected by o3DIGARS. The larger number of points below the y = x line shows that o3DIGARS is able to select the model with low TMscore more accurately.Slide27

o3DIGARS ResultsPlot shows the comparison of Rosetta and I-

Tasser’s Tmscore (average of 5 models) with the TMscore of the model selected by o3DIGARS. The larger number of points below the y = x line shows that o3DIGARS is able to select the model with low TMscore more accurately.Slide28

o3DIGARS ResultsPlot shows the comparison of

Rosetta and I-Tasser’s Tmscore (average of 5 models) with the TMscore of the model selected by o3DIGARS. The larger number of points below the y = x line shows that o3DIGARS is able to select the model with low TMscore more accurately.Slide29

o3DIGARS Results

3DIGARS3.0

o3DIGARS

% Improvement

Avg. PCC

-0.56333

-0.65341

15.99%

Avg.

Tmscore

of Low Energy Model

0.645093

0.678777

5.22%

Table compares the performance

of 3DIGARS3.0 and

o3DIGARS methods based on average PCC and

TMscore

of Low Energy Models on

CASP8 independent test dataset.Slide30

o3DIGARS Results

3DIGARS3.0

o3DIGARS

% Improvement

Avg. PCC

-0.63487

-0.65553

3.25%

Avg.

Tmscore

of Low Energy Model

0.60976

0.635949

4.29%

Table compares the performance

of 3DIGARS3.0 and

o3DIGARS methods based on average PCC and

TMscore

of Low Energy Models on

CASP9 independent test dataset.Slide31

Software AvailabilitySoftware, dataset and related materials for all three versions of our energy functions (3DIGARS, 3DIGARS2.0 and 3DIGARS3.0) and the

Ab Initio (V1) are available freely onlinehttp://cs.uno.edu/~tamjid/Software/ab_initio/v1/PSP.ziphttp://cs.uno.edu/~tamjid/Software/3DIGARS/3DIGARS.ziphttp://cs.uno.edu/~tamjid/Software/3DIGARS/3DIGARS2.0.ziphttp://cs.uno.edu/~tamjid/Software/3DIGARS/3DIGARS3.0.zipSlide32

AcknowledgeI’d also like to gratefully acknowledge

the Louisiana Board of Regents through the Board of Regents Support Fund, LEQSF (2013-16)-RD-A-19.I’d like to thank my supervisor Dr. Md Tamjidul Hoque

and everyone who has helped me in this project.Slide33