Application of Monte Carlo Simulation: Removing averaging artifacts in protein structure prediction Dukka Outline • Background • Protein Structure Prediction and CASP • TASSER algorithm • MCORE algorithm Background • Experimental or computational method often output results as an ensemble of protein structures. – NMR, Protein Structure Prediction, Protein Docking, RNA Structure Prediction • • • • A single representative structure is required to compare or do further analysis. Representative structure (consensus structure) = a centroid structure by averaging the Cartesian coordinates of the ensemble of superimposed structures. RMSD between the ‘averaged structure’ and any reference structure is always less than or equal to the average RMSD of the individual members. (Zagrovic et al.) However, the centroid structure has averaging artifacts rendering bond angles and bond lengths to be unphysical. Protein Structure Prediction and CASP • Critical Assessment of Structure prediction of Proteins (CASP) is a biannual contest where different groups try to predict structure of a protein whose structure is not released to the outside world. • One of the most popular and objective contest in the bioinformatics field. • CASP8 just over. • Major observations from CASP7: – Methods are more or less ripe enough – Consensus servers usually outperform individual servers – A lot of work needed to be done in the refinement step Refinement • Given a set of conformations obtain a conformation that is closest to the native structure. • Molecular force fields like AMBER, CHARMM can be utilized but as we know they are not perfect. • Furthermore, still lack of perfect definition of “closest”. Hence, CASP coming up with new ideas of other measures to measure the closeness to the native like HB score and so on. • Often, the ‘most closest prediction’ is not ranked top 1. Hence, ‘Refinement’ is getting a lot of attention. TASSER algorithm (Threading/ASSembly/Refinement) Centroid Structure Zhang & Skolnick, 2004 Problem Identification • TASSER is one of the best prediction server in both CASP7 and CASP8. • A large number of conformations is generate after the assembly step. However, we can submit only a couple of models. • Clustering is utilized and the centroid of the largest cluster (Combo model) is predicted as the output and has proven to be successful. • Artifacts in ‘Tasser (combo) output’ – Unrealistic bond lengths and bond angles due to averaging artifacts Scope – To fix these unrealistic bond lengths and bond angles C-alpha Space Energy Minimization! Combo and Closc Models Fraction of clashes COMBO model : The centroid structure of the most dense cluster. CLOSC model : The structure that is closest to the centroid of the most dense cluster. QuickTime™ and a TIFF (Uncompressed) decompressor are needed to see this picture. QuickTime™ and a TIFF (Uncompressed) decompressor are needed to see this picture. PULCHRA • PULCHRA - based on steepest descent minimization and a simple force field. • Sometimes, can not come out of the kinetic trap. • Heavily distorted chain, the minimization procedure does not converge or the optimized model still exhibits irregularities. Rotkiewicz and Skolnick, 2008 QuickTime™ and a TIFF (Uncompressed) decompressor are needed to see this picture. MCORE Generate an extended structure based on Combo model Start from a ‘close-by model’ Monte-Carlo Minimization Output the best structure Generation of Extended Structure Using the distance distribution from the PDB, mainly three types: x-Pro = 3.77, x{ALA|ARG|ASN|LEU|LYS|MET} = 3.81, and x{ASP|CYS|GLU|GLY|HIS|ILE|PHE|SER|THR|TRP|TYR|VAL} = 3.80 Monte-Carlo • Two major components of any Monte-Carlo Approach – Energy Function • Can be generic force field or any combination of terms – Move Sets • Critical to the performance of the algorithm, more of an art(?) – Convergency Criteria • Naïve way (Run for certain number of steps) • Introduce some criteria based on the generated conformations Monte-Carlo: Metropolis Criteria • Starting from a state A, make a change in the configuration to obtain a new (nearby) configuration B. • Compute EB • If (EB < EA), assume the new configuration, since it is a desirable thing. • If (EB > EA), calculate the probability p p e(E B E A ) / T • Draw r from uniform distribution [0,1], if r < p then accept the new configuration B else reject the new configuration B. Move Sets • Move Sets – Global move-set • Rest-all bead move – Local move-set • 1-bead move • 2-bead move • 3-bead move • 4-bead move • 5-bead move – End-bond move • 1,2,3-bead C-terminal end bond move • 1,2,3-bead N-terminal end bond move Move Sets • Calculate the unit vector along axis defined by i-1 and i+1 • Calculate the rotation matrix around this vector • Calculate the new position of i • Important thing is to preserve the bond length i.e. to preserve the distance between consecutive Calphas. i i-1 i+1 One bead move Two bead move Three-bead move Rest-bead move Four-bead move Five-bead move Axis of rotation End-bond Move Sets Axis of rotation QuickTime™ and a TIFF (Uncompressed) decompressor are needed to see this picture. QuickTime™ and a TIFF (Uncompressed) decompressor are needed to see this picture. Energy Function Excluded volume N 2 kexcl N (rkl ro _ excl k1 lk2 Penalize if the distance is less than 4.0A Bond angle )2 k N 2 N 2 2 ang ( i,i1,i2 o _ ang) kclos (d kk t d o _ clo ) k1 k1 Penalize if the angle is not between 70 and 150 kang 0.075 ro _ excl 4.0 Closeness to target kclos 5.4 Penalize if the difference in C-alpha position between the target and starting structure is not with-in certain cutoff kexcl 2.9 do _ clo 0.001 o _ ang 70 if i,i1,i2 70 and150if 150 and same otherwise N: Number of C-alpha atoms Assessment of Move Sets and Energy Function • Before doing the actual computation, have to test whether the move sets and energy function is properly working or not. • So, have to design some test cases. Positive test cases would be to drive extended structure to native structure. – Desired results: should be able to drive ‘very close’ to extended structure to native structure in relatively short number of steps Data Set • 1363 proteins less than 200 residues and the combo RMSD to the native is lesser than 6.5 Å. • 1363 Centroid structures (COMBO models) • 1363 CLOSC models • 1363 Close-by structures (CLOSC models + Pulchra Refinement) • 1363 Native structures. Driving Extended to Native Steps 10000 steps RMSD = 0.039 Average RMSD to NATIVE (Å) Average Energy 0.045 0.06 Steps 0.041 Driving Extended to Native 0.033Å Ext-refined Vs CA Convergency criteria i l | rmsd_diff((i –l))| < Tolerance value, where l = i+j , j=1,…,L Tried with different value of L and L=49 and Tolerance value = 0.005 seems reasonable. Propose two algorithms • MCORE: Start from a ‘close-by model and drive it towards the COMBO model. • CLOSC models as the close-by models. – When close-by model is readily available • MCORE-EXT: Start from an extended structure and drive it towards the COMBO model. – When close-by model is not readily available Average Energy MCORE: Driving Close-by models to COMBO Steps Fraction of Atoms Clashing in MCORE Fraction of Atoms Clashing in COMBO Why cannot go much closer to COMBO? RMSD of MCORE to COMBO (Å) RMSD of MCORE to COMBO (Å) RMSD of MCORE to NATIVE (Å) MCORE Vs Combo RMSD of COMBO to NATIVE 38 proteins had even lesser RMSD than the respective combo model RMSD to Native (Å) Fraction of Atoms in Clashes Comparison of Different Models 3.35 3.36 3.54 3.28 0.010 0.065 0.000 0.63 TM-score of four models 0.770 0.746 0.747 0.754 Results Avg. RMSD to Native (Å) Avg clash < 1.9 Avg clash < 3.6 Combo 3.28 0.03 0.630 Closc 3.54 0 0.614 MCORE 3.35 0 0.010 Pulchra (Closc) 3.54 0 0 MCORE(EXT, 2000 steps) 3.35 0 0.011 Pulchra (Combo) 3.36 0.005 0.065 Some Examples 0.78Å 0.354Å 1akhA refined Vs native 1akhA com Vs refined 12 clashes 0.674Å 0.68Å 1akhA com Vs native 1akhA pulchra Vs native 0.852Å 2.948Å 3bbn_ refined Vs combo 3bbn_ comboVs Native 2.918Å 3bbn_ pulchra Vs Native 3.099Å 3bbn_ refined Vs Native All-atom model reconstruction • Built the main chain atoms of the refined Cα trace. • Rebuilt side-chains using two methods – Pulchra (-c) – Scwrl 3.92(all) 2.868(cα) 3.95(all) Conclusion • Designed an algorithm to remove averaging artifacts and applied it to refine combo model. • Acknowledgments – Dr. Jeff Skolnick and all the members of the Skolnick Lab, especially Lila, Shashi, Hongyi, Seung Yup,……. – Dr. Dennis Livesay • Future Works – Refinement in All-atom space