Improving De novo Protein Structure Prediction using Contact Maps Information

Transcription

1 CIBCB 2017 IEEE Conference on Computational Intelligence in Bioinformatics and Computational Biology Improving De novo Protein Structure Prediction using Contact Maps Information Karina Baptista dos Santos PhD student in Computational Modeling National Laboratory for Scientific Computing (LNCC) Petrópolis, RJ - Brazil

2 Introduction Protein Structure Prediction (PSP) have a large potentital for biotechnological applications MQTVLKKRKKSGYGYIPD IADIRDFSYTPEKSVIAA LPPKVDLTPSFQVYDQGR IGSCTANALAAAIQFERI HDKQSPEFIPSRLFI Protein Function o Can provide essential clues to understanding diseases o Allows the creation of new proteins and the development of new drugs 2

3 Introduction CASP (Critical Assessment of Techniques for Protein Structure Prediction) o 12th CASP ( ) o Assess the state of the art of protein structure prediction methods Evaluation via process of blind prediction o Progress in PSP methodologies Use of protein contact maps in the search of native conformation o Highlight where the future effort may be focused Proceedings of the National Academy of Sciences pages , 30 NOV

4 Introduction Protein Contact Maps CASP10, CASP11 and CASP12 o A minimalist representation of the 3D protein structure. o Two residues in a protein are in contact if the Euclidean distance between their Cβ atoms (Cα for Glycine) is at most some threshold value, usually 8Å. Cβ i Cβ j 8.0Å Proteins: Structure, Function, and Bioinformatics pages 84-97, 16 OCT 2013 DOI: /prot

5 Introduction Protein Contact Maps prediction Residue covariation analysis o Sufficiently large number of homologous proteins (MSA) Which residue pairs are most likely to be in contact? Which residue pairs are important to be maintained by evolution? MSA inference Contact in 3D covariation These contacts are not template! Contact Map Precision threshold [0, 1] Probability of such residues being in contact Plos One, v.6, n.12 page e28766, 7 DEC 2011 DOI: /prot

6 Introduction Protein Contact Maps prediction o MetaPSICOV 1 (the CASP11 winner) Stage 1 Classic contact prediction features 3 three coevolution-based methods PSICOV, mfdca-free Contact and CCMpred Infer correlations signals in MSA Stage 2 Fill in the gaps in contact map and remove outliers (Filter) Hydrogen bonds patterns between residues in contact 1 Kosciolek, Tomasz, and David T. Jones. "Accurate contact predictions using covariation techniques and machine learning." Proteins: Structure, Function, 6 and Bioinformatics (2015).

7 Objectives Develop a strategy to use the information of protein contact maps in the GAPF PSP program Contact Map Term a new term of the GAPF Fitness Function Assess the predictive information capacity of the contact maps provided by MetaPSICOV Stage 1 7

8 Methods GAPF 2, 3 - Genetic Algorithm for Protein Folding o Phenotypic crowding-based steady-state genetic algorithm o Candidate solutions are encoded by the backbone dihedral angles o Coarse-Grained representation for the side chain atoms 2 GAPF - Custódio, F. L.. Algoritmos Genéticos para Predição Ab Initio de Estrutura de Proteínas. Laboratório Nacional de Computação Científica Custódio, Fábio Lima, Helio J. C. Barbosa, and Laurent Emmanuel Dardenne. A Multiple Minima Genetic Algorithm for Protein Structure Prediction. Applied Soft Computing Journal 15:

9 Methods GAPF - Genetic Algorithm for Protein Folding o Protein Fragment Libraries Profrager 4 Fragment Structure Protein Model o Secondary Structure Prediction PSIPRED 5 4 Trevizani, R., Custódio, F.L., dos Santos, K.B., and Dardenne, L.E. Critical Features of Fragment Libraries for Protein Structure Prediction. PLOS ONE, vol.12, no.1, pp.1-22, Jones, D.t., Protein secondary structure prediction based on position specific scoring matrices. Journal of molecular biology, vol.292, no2, pp ,1999.

10 Methods GAPF - Fitness Function 7 o Based on the energy from the interaction between the atoms of the protein Dihedral potential (from Gromos96 force field); Atomic repulsive term; Hydrophobic compaction term; 4 Hydrogen bond terms Hydrogen bonds 1a Rocha, Gregório Kappaun. Desenvolvimento de Metodologias Para Predição de Estruturas de Proteínas Independente de Moldes. Laboratório Nacional de Computação Científica

11 Methods Contact Map Term Distance constraints λ = Contribution value add to the predicted model energy γ = Precision threshold associated with the probability of such residues being in contact σ(i, j) = Euclidean distance between Cβ atoms of residues i and j o The more residues are respecting the distance constrainsts, the greater the contribution to the fitness function 11

12 Methods Test Set Experimentally determined 3D structures 3FIL 56 residues T0773-D1 67 residues (PDB 2N2U) T0820-D1 90 residues T0769-D1 97 residues (PDB 2MQ8) T0766-D1 108 residues (PDB 4Q53) 12

13 Methods Contact Maps o Native Map All residues pairs found in native structure (PDB file) Precision threshold (PT) = 1.0 R1 R2 d min d max PT

14 Methods Contact Maps o Filtered Map Only the residues pairs predicted by MetaPSICOV Stage1 that are present in the Native Map. Precision threshold (PT) = MetaPSICOV Predicted value R1 R2 d min d max PT

15 Methods Contact Maps o Filtered Map Only the residues pairs predicted by MetaPSICOV Stage1 that are present in the Native Map. Precision threshold (PT) = MetaPSICOV Predicted value R1 R2 d min d max PT Native X Filtered Same Residues / Different Precision Threshold 15

16 Filtered Map - Precision threshold profile

17 Native x False contacts in the MetaPSICOV predicted map Contact Pairs Filter native contacts from predicted map Great challenge filA T0773-D1 T0820-D1 T0769-D1 T0766-D1 All MetaPSICOV Stage1 contacts Native contacts o No method described is able to do this selection efficiently Precision threshold above some value (e.g. 0.5, 0.4 ) Number of contacts associated with the sequence length ( e.g., L/5, L/2 ) 17

18 Results Impact of the use of Contact Maps o 4000 models predicted by GAPF Best of All Models Predicted Top5 Best Energy Models o RMSD 4.0Å Good predictions 18

19 Results Impact of the use of Contact Maps o 4000 models predicted by GAPF Best of All Models Predicted T0820-D1 Top5 Best Energy Models 6Å >10Å o RMSD 4.0Å Good predictions 19

20 Results Satisfied contact does not mean correct contact! Best of All Top 5 Best Energy Native Map 60% 65% Filtered Map 48% 44% Cβ i Cβ j 8.0Å Native Structure Cβ- Cβ = 7.71 Å Model Cβ- Cβ = 4.23 Å 20

21 Results RMSD (Å) RMSD (Å) RMSD (Å) RMSD x Energy (Native Map) The greatest contribution to GAPF 3FIL T0773-D1 Energy (Kcal/mol) Energy (Kcal/mol) T0820-D1 Energy (Kcal/mol) 21

22 Results RMSD x Energy (Standard Procotol) o Difficult to select the best structure model methodology to select decoys 3FIL T0820-D1 22

23 Results T0820-D1 - Summary of improvements GAFP_St GAFP_Cm Native map GAFP_Cm Filtered map Best of All Models RMSD = 9.60 Å GDT = % RMSD = 3.19 Å GDT = % RMSD = 3.73 Å GDT = % Top 5 Best Energy Models RMSD = Å GDT = % RMSD = 3.61 Å GDT = % RMSD = 3.65 Å GDT = % Experimental Structure Model predicted by GAPF program 23

24 Conclusions Contact Map in the form of Distance Constrains Useful Strategy for PSP Naive Potential Properly combined with GAPF fitness function Predictors give probability for all possible residue-residue contacts o Lack of full confidence in the prediction of the contacts Development of strategies to filter and enhance contact maps o Which contacts are most valuable for PSP 24

25 Acknowledgment Thank you! 25