Prediction of Protein Structure and Function on a Proteomic Scale Jeff Skolnick Director Center of Excellence in Bioinformatics
General Approach
Prediction of Protein Structure
Overview of CASP5 Results:
Comparative Modeling (CM) Results
T0153 CM COORDINATE SUPERPOSITION RMSD = 1.74 Å ( 129 / 134 aa ) NATIVE (discontinuous line) : PREDICTED (continuous line) : 1mq7 A rank #1
Fold Recognition (FR) results
T0135 FR(A) GLOBAL COORDINATE SUPERPOSITION RMSD = 4.80 Å ( 106 / 106 aa ) NATIVE (discontinuous line) : PREDICTED (continuous line) : rank #1
T0135 FR(A) GLOBAL COORDINATE SUPERPOSITION RMSD = 4.80 Å ( 106 / 106 aa ) NATIVE (discontinuous line) : PREDICTED (continuous line) : rank #1 Yellow line: region originally aligned to the template (1h6kX )
New Fold (NF) results
T0181 (NF) PREDICTED: rank #2
How representative is the set of solved PDB structures?
The PDB is a covering set of protein structures at low resolution Results from a new structure alignment program, SAL Kihara & Skolnick, J. Mol. Biol, 2003:333:393-802
Structural alignments to proteins of different secondary structure Different CATH ids 100 residue proteins
Use of best structural alignments Can we build good models starting from protein templates with average sequence id of 7%?
TASSER:Threading/ASSEmbly/Refinement
Very large scale structure prediction benchmark
Comprehensive benchmark set of PDB structures Length range: 41~200 Sequence identity cut-off: 35% In total: 1489
Summary of Results
Summary of Overall Folding Results SAL TASSER MODELLER Besta Alignb Top-5c Top-1d <RMSD>e 2.510 1.877 2.246 2.352 2.708 3.740 4.318 <COV>f 82% 100% NRMSD<6.0 NRMSD<5.5 NRMSD<5.0 NRMSD<4.5 NRMSD<4.0 NRMSD<3.5 NRMSD<3.0 NRMSD<2.5 NRMSD<2.0 NRMSD<1.5 NRMSD<1.0 1489 1485 1472 1440 1369 1255 1064 776 498 218 46 1488 1476 1422 1250 922 411 83 1487 1481 1468 1447 1396 1259 987 623 253 52 1475 1464 1450 1423 1359 1206 928 582 241 49 1462 1431 1395 1336 1141 1008 750 520 244 37 1326 1266 1195 1116 984 834 647 475 300 124 20 1202 1138 1060 962 841 697 551 397 85 15
Some Examples:
Summary At low resolution, the PDB is most likely complete for single domain proteins Can build acceptable full length models in the majority of cases Can refine the initial structures to move closer to native, even starting from the best structural alignment
Results from threading/refinement “Real Life” situation
TASSER:Threading/ASSEmbly/Refinement
“Easy” Cases: At least two threading templates identified with significant consensus region or One template with z-score that is highly significant
“Medium ” Cases: At least two threading templates identified without any significant consensus region or One template with z-score above threshold for correct fold assignment
Composite Threading Results We can identify the correct global fold in 92% of the entire representative set of small PDB structures Can generate good template alignments in 59% of the cases Good substructures 67% of the cases
Summary of Results
Examples of Alignment improvement Medium Easy Template Final model Template Final model Thin lines: Native; thick lines: Template/model Two factors mainly contribute to the improvement: geometric connectivity Better packing of local structure and side group because of the force field
Comparison to Ensemble of NMR Structures (Predicted Structure to Centroid/Farthest NMR Structure to Centroid) Thick Line is Predicted Structure
Benchmark set of larger proteins (201-300 residues) 487 Single-domain proteins 236 two-domain proteins 22 three-, four-domain proteins 745
Successful Predictions of Transmembrane Proteins
Application to ORFS <201 residues in E. coli 61% Easy (829/1360) 38% Medium (521/1360) 10 Hard TASSER 68% (920/1360) Good models
Summary Acceptable model in about 2/3 of the cases (969/1489) Application to E coli Yields similar results ~2/3 of proteins should have good model -Almost all (90%) have a good template
Development of Active Site Descriptors
Representation of an Automated Functional Template [ AFT ] Types of functional sites from SwissProt: METAL BINDING ACT_SITE SITE cm SCj cm SCi Cai+1 Caj-1 Caj Cai Caj+1 Cai-1 Set of distances between: cm SCk Cak Ca atoms and center of mass of the side chains corresponding to 3 to 5 functional residues, Cak-1 Cak+1 Ca atoms corresponding to the adjacent residues.
Specificity parameters of AFTs 30 28 26 24 22 20 18 16 14 12 10 8 6 4 2 Positive hits Negative hits Restrictive cutoff: average value of DRMSDMaxPos and DRMSDMinNeg. Permissive cutoff: expected number of false positive matchs is less than 0.005 in a random structure. Number of hits in the subset of PDB High confidence DRMSD interval Low confidence DRMSD interval DRMSDMaxPos DRMSDMinNeg 0.0 0.5 1.0 1.5 2.0 2.5 DRMSD [ Å ]
Fraction of decoys correctly annotated vs Fraction of decoys correctly annotated vs. ranking of the best true positive hit Global Ca crmsd from the native structure Local Ca drmsd from the native structure 73% 56% 48% 35% The recognition by an AFT matching the first three components of the true EC number is considered a true positive hit.
Threading of Entire Genomes
Summary of Fold Assignments Organism Total Protein ORFs ORF Coverage (%) Amino Acid Coverage % FASTA (%) PSI-BLAST – PDB (%) PSI-BLAST – PDBseq (%) GTOP (%) Pedant (%) Gerstein (%) M. genitalium 484 387 (80.0) 48.1 231 (47.7) 205 (42.4) 259 (53.5) 273 (56.4) 214 (44.2) E. coli 4289 3356 (78.2) 50.2 1660 (38.7) 1516 (35.3) 1906 (44.4) 2032 (58.5) 1954 (45.6) 1191 (27.8) B. subtilis 4106 2988 (72.8) 47.2 1465 (35.7) 1314 (32.0) 1732 (42.4) 1947 (60.2) 1963 (47.7) 1121 (27.3) A. aeolicus 1522 1297 (85.2) 48.0 646 (42.4) 592 (38.9) 771 (50.7) 827 (53.1) 800 (52.6) 527 (34.6) S. cerevisiae 6343 4610 (72.7) 30.0 1962 (30.9) 1804 (28.4) 2422 (38.2) 2694 (42.5) 2766 (42.9) 1699 (27.3)
Comments on fold distribution Protein folds can be assigned to 72-85% of genes in each genome. 30-50% of the total amino acids in a genome are covered by the assigned folds. Generally, distribution of folds are similar in the 5 organisms. Folds of a/b type are abundant. Folds of multi-functions are abundant in a genome. Kinase fold shows up in top 5 only in S.cerevisiae.
MULTIPROSPECTOR: Prediction of Protein-Protein Interactions L. Lu, H. Lu, J. Skolnick. Proteins, 2002, 49, 350-364.
Overall Idea of Multimer Threading Monomer threading A X: GELPIAPIGRIIKNA GAERVSDDARIALAK VLEEMGEEIASEAVK LAKHAGRKTIKAEDI KLARKMFK Y: GEVPIAPLGRIIKNA VLEEMGEEIASEAIR LAKHAGRKTIKAEDV KLAKKMFK B A B X: GELPIAPIGRIIKNA GAERVSDDARIALAK VLEEMGEEIASEAVK LAKHAGRKTIKAEDI KLARKMFK Y: GEVPIAPLGRIIKNA VLEEMGEEIASEAIR LAKHAGRKTIKAEDV KLAKKMFK X Y Assign fold on the basis of Z score and Interface Energy Multimer Threading Multimer Structure Library A B
Preliminary test on Known Dimers and Monomers Homodimers: 58 Heterodimers: 20 Monomers: 96 96 5 91 20 20 54 4 58 Proteins predicted to be dimers Proteins predicted to be monomers
Procedure for genomic scale prediction of protein-protein interactions by MULTIPROSPECTOR
Comparison of colocalization index for different methods
Distribution of predicted interactions in functional categories
Conclusions
Completeness of the PDB Conclusions Completeness of the PDB PDB is a covering set of single domain proteins at low to moderate resolution Protein Structure prediction problem can be solved with more powerful threading algorithms!!
TASSER For single domain proteins: In almost all cases, for all ranges of initial RMSD, even when starting from the “best” structural alignment, the final results are better than the initial template- the models move closer to native Based on a comprehensive folding benchmark, we expect low resolution structures for ~ 2/3 of proteins with low sequence identity to PDB structures Weak dependence on secondary structure type
Structure to Function Low resolution structures can be used to identify active sites. Genome scale threading – greater than 70% of ORFs can be assigned to known folds Extension to protein-protein interactions Comparable accuracy to agreement between two experimental methods
Acknowledgements http://bioinformatics.buffalo.edu/ Center of Excellence in Bioinformatics Yang Zhang Adrian Arakaki Purdue University Daisuke Kihara Yale University Long Lu University of Illinois Hui Lu $$$$$ NIH, NSF & The Oishei Foundation http://bioinformatics.buffalo.edu/