Transcript Round 1
Protein Folding: Interrelation between Secondary and Tertiary Structure Determination Karl F. Freed James Franck Institute and Department of Chemistry University of Chicago KITPC, Beijing China, July 29, 2009. Human, Dog, Rat, Worm Genome Projects: Obtain genes which code for protein sequences www.genomesonline.org/images/gold_s1.gif www.ncbi.nlm.nih.gov/Genbank/genbankgrowth.j pg Proteins: The primary functional biomolecules insulin: hormone glucose levels hemoglobin: oxygen transport cytochrome c: heme group electron transfer ribonuclease: lysozyme: cleaves RNA cleaves carbohydrates antibody:recognize/target foreign bodies Function follows from structure myoglobin: oxygen storage glutamine synthetase: synthesize glutamine Our goal is to determine the folded structures Protein Folding Problem SEQUENCE 3D STRUCTURE Folding 0.00001 – 10+ sec “Unfolded State” Many conformations Amino acids - polar, neutral, charged, or hydrophobic “Native State” Responsible for function Why haven’t we solved The Protein Folding Problem after 40+ years? Sequence “Mother Folding” Unfolded State, “statistical coil” Sequence determines structure Two aspects: 1. Predict pathways 2. Predict structure Give me an aa sequence; I’ll produce a pathway and the final structure. Why so difficult? 1. We’re not smart enough 2. It’s a very complex system 3. … Give me an aa sequence; I’ll produce a pathway and the final structure. Why so difficult? Complex problem: •Too many atoms (not enough computing power) •Force fields inaccurate (pairwise interactions inadequate) •Complex interplay between secondary and tertiary structure formation (local vs. long-range structure) •High degree of folding cooperativity •Averaging doesn’t work (no mean field models) •H2O solvent is difficult to treat •Don’t know all the rules? •Not enough information? •Reductionist models (e.g. H/P) often too simple What are the fundamental principles needed to predict pathways and structures? Two aspects of “The Protein Folding Problem”: - Mechanistic studies: “How does it get from the U-state to the N-state?” - Predict structure from sequence: Common Strategy: Use sequence similarity (homology) to proteins with known structures. …ALEVADKVAIYTRE… …ALEVAEKVAIWTRE… ? Use parts or full structure to predict new structure (homology). Successful when have homology, but side-steps The Question: “What are the Principles?” Why do they fold into specific structures? Hydrophobic (greasy) atoms in the center away from water Form hydrogen bonds between all the nitrogens and oxygens Pack all atoms with no spaces H2O H2O H2O H2O H2O H2O a helix bstrand van der Waals interactions What level of representation is needed? Ca Cb Beads-on-a-string Monomer structure & miscibility of polyolefins K. F. Freed and J. Dudowicz, Adv. Polym. Sci. 183, 63-126 (2005). Major themes and challenges in protein folding 1) Polymer bends in certain ways 2) Satisfy main-chain hydrogen bonds and form secondary structure. y f 3) Bury hydrophobic residues and pack the atoms Must satisfy 1, 2 & 3 simultaneously Vast Conformational Search: Levinthal Paradox How does a protein find the time to fold? Polypeptide backbone is flexible, adopting specific conformations Poly-Proline II basin Beta basin preferences y Helical, turn basin Ramachandran Map f Each f,y pair: 3 conformations ~3n ~100.5n conformations/amino acid e.g., 1050 states for a 100 a.a. protein Random, unbiased search would take 1035 s even with fsec sampling And we have to search too! What info is needed to fold proteins? all-atom dihedral angles Step 1: Remove solvent – ( hydrophobic, screening, etc terms) Step 2: Replace side chains by Cb atoms. All-atom protein + solvent Simulation would take decades y Cb f All-atom protein Only main chain heavy atoms and Cb atoms. Backbone f,y angles are the only coordinates. 2N total for N aa’s (need ~10x3N for x,y,z of heavy atoms) Reduced representation: Side chains are only Cb y f Retains the 3 themes Big Challenge: Retain sequence information lost with removal of side chains 1) Proteins bend in certain ways b-basin PPII-basin Sampling in dihedral space: f-y angles and Ramachandran BASINS y f b PP2 e Extended f & y show very strong Helical preferences for certain regions of the f-y map, called Ramachandran basins. (due to steric & electrostatic interactions) Where to get this information? aL a-basin Phi From computer simulations? Psi But, force fields can vary widely Zaman et al. JMB 2003 Data mine protein data base (PDB) of crystal structures: Extract the distributions for each type of amino acid PDB THR ALA y f y f Ramachandran Map of ALA with neighbors=ALA, ASP ALAALA ALAASP Map depends on neighbor type and conformation Strong correlations in sequence Move-set uses highly selected trimers from PDB trimer library PDB Filter by residue type & conformation trimer move Specifying side chain type and backbone geometry implicitly includes all-atom side chain information Knowledge-based energy function: Assign interaction energies according to the observed distances in the PDB PDB 6000 Probability 5000 4000 3000 2000 1000 0 ProbPDB(rij) Probability of finding 2 atoms some distance apart e.g. Dist(Caala – Caval). Energy = - log (Prob) F2 0 2 0 2 4 6 10 8 10 0 -2 -4 4 6 ala val distance (Ca -Cb ) EnergyPDB(rij) = -ln( ProbPDB(rij) ) 8 Secondary structure prediction methods… Machine Learning Statistics Single residue and nearest neighbors Chou and Fasman (1974) QVD Secondary structure probability conditional on identity of neighbors Neural Network Jones (1999) Dor & Zhou (2007) Support Vector Machines Kim (2003) Hu et al. (2004) Current Accuracy of Best Methods is 75 – 80 %: Good for a, but poorer for b and a/b Multiple sequence alignments + SASA, … McGruffin and Jones (2003): PSIPRED Dor and Zhou (2007) What prevents accuracy of secondary structure prediction from reaching ~90% ? Secondary structure often depend on long range interactions, i.e. tertiary structure This is supported by the following studies: •The same fragment from different parts of Minor and Kim (1996) protein G forms varying secondary structures •Secondary structure prediction accuracy decreases with increasing contact order •The same sequence fragment can be found in multiple native secondary structure types Attempt to integrate secondary and tertiary structure Add 3-D models to Machine Learning Folds target & homologs (~ 5% gain) Kihara (2005) Pan et al. (1999) Jacobini et al. (2000) Zhou et al. (2000) Ikeda and Higo (2006) Neural Network + tertiary Meiler and Baker (2003) What we do differently: Couple secondary and tertiary structure during the folding process Restrict possible secondary structure as the chain folds Tertiary structure Secondary structure “Eliminate all other factors, and the one which remains must be the truth.” A. C. Doyle, The Sign of the Four (1890) ItFix Algorithm Trimer library: Full PDB Keep track of where the trimers came from (Helix, extended, coil) Starting configuration (no SecStr restriction) Stage 1. Fold 100 times with trimers from entire PDB Round 1 Round 2 Stage 2. Refine each Stage 1 end structure Stage 1. Fold 100 times with restricted library Stage 2. Refine each Stage 1 end structure Round 1: Filter by residue identity only Eliminate an option (H,E,C) if PHelix, PCoil < 10%, PExt < 5% Round 2 Stage 1. Fold 100 times with restricted library Round 3 Stage 2. Refine each Stage 1 end structure Repeat until no further fixing is possible… Final Round Stage 1. Fold 100 times with restricted library Stage 2. Refine each Stage 1 end structure helix strand Not(Strand) Not(Helix) Coil: B, G, I, S, T, N Predicted secondary structure and 3-D model Rounds 3,4 … B) Iterations mimic steps in folding pathway 1 Major pathway Unfolded state b1 b2 helix b4 b5 310 b3 Round 1 0 Round Rnd 1 10 Rnd 0 b1-b2 hairpin +helix + b4 Rnd 2 Rnd 1 1 0 + b3 Round 2 Round 3 Rnd 3 1 0 + b3 Rnd 2 Round 4 + b4 1 0 +helix Round 6 10 +310 helix Rnd 9 Round 9 + b5 b2 helix b4 b5 0 b1 1 residue index 310 Pred’n b3 73 Native state Ca-RMSD 3.1 Å Figure 4 A 1af7 prediction 2.9 Å 1r69 prediction 4.2 Å 1ubq prediction 5.3 Å 1af7 best 2.5 Å 1r69 best 2.4 Å 1ubq best 3.1 Å B Energy C 1af7 Ca rmsd 1r69 Ca rmsd 1ubq Ca rmsd Round 5 1b72A 1di2 1r69 1 1af7 Round 0 Round 0 Round 0 Round 1 Round 1 Round 1 Round 1 Round 2 Round 2 Round 2 Round 2 Round 3 Round 3 Round 4 Round 3 Round 5 Round 4 Round 6 Round 4 Round 7 Round 6 Round 8 Round 6 10 10 1 0 10 0 Secondary Structure frequency 10 Round 0 residue index 1AF7 3.4 Å RMSD 1TIF 5.4 Å RMSD 1SAP 7.8 Å RMSD Novel Aspects •Predict 2° & 3° structure without using homology •Use principles of protein structure and folding: i) couple 2° & 3° structure formation ii) sequential stabilization •Iterative fixing to reduce the search. •Use a Cb representation •Potential function: orientational & 2° structure dependence in a Cb model •Q8-level 2° structure, (f,y) prediction, can outperform PSIPRED •Outputs pathway information Conclusions We couple the formation of tertiary structure and the identification of secondary structure to improve both types of structure prediction Goal: Predict secondary structures – Exceed the 80% barrier… without homology Future applications: Improved ab initio structure prediction Assist in identifying and ranking of threading targets Future improvements: Better energy functions = better folding and predictions Bigger proteins – how far can we push secondary structure accuracy even when we fail in determining folded structure Acknowledgements Prof. Tobin Sosnick Joe DeBartolo - ab initio folding, secondary structure prediction Dr. Andres Colubri – Folding simulations, software Dr. Abhishek Jha (MIT) – coil library, unfolded state, a, b propensities, structure refinement, electrostatics James Fitzgerald (Stanford) - Statistical potentials, torsional dynamics Prof. M. Zaman (UT Austin) – Simulations on peptides All-atom statistical potential Dr. Min-yi Shen, Prof. A. Sali, UCSF Funding: NIH, NSF, Burroughs Wellcome Fund