understanding intelligence

Predicting Protein Structures and Structural Features on a Genomic

Scale

Pierre BaldiSchool of Information and Computer Sciences

Institute for Genomics and BioinformaticsUniversity of California, Irvine

UNDERSTANDING INTELLIGENCE

• Human intelligence (inverse problem)• AI (direct problem)• Choice of specific problems is key• Protein structure prediction is a good

problem

PROTEINS

R1 R3

| |

Cα N Cβ Cα

/ \ / \ / \ / \N Cβ Cα N Cβ

| R2

Utility of Structural

Information

(Baker and Sali, 2001)

CAVEAT

A L L -A L P H A A L L -B E TA

M E M B R A N E (2 5 % ) G L O B U L A R (7 5 % )

P R O TE IN S

REMARKS• Structure/Folding• Backbone/Full Atom• Homology Modeling• Fold Recognition (Threading) • Ab Initio (Physical Potentials/Molecular

Dynamics, Statistical Mechanics/Lattice Models)

• Statistical/Machine Learning (Training Sets, SS prediction)

• Mixtures: ab-initio with statistical potentials, machine learning with profiles, etc.

PROTEIN STRUCTURE PREDICTION (ab initio)

DECOMPOSITION INTO 3 PROBLEMS

1. FROM PRIMARY SEQUENCE TO SECONDARY

STRUCTURE AND OTHER STRUCTURAL FEATURES 2. FROM PRIMARY SEQUENCE AND STRUCTURAL

FEATURES TO TOPOLOGICAL REPRESENTATION 3. FROM TOPOLOGICAL REPRESENTATION TO 3D

COORDINATES

Helices

1GRJ (Grea Transcript Cleavage Factor From Escherichia Coli)

Antiparallel β-sheets

1MSC (Bacteriophage Ms2 Unassembled Coat Protein Dimer)

Parallel β-sheets

1FUE (Flavodoxin)

Contact map

Secondary structure prediction

GRAPHICAL MODELS: BAYESIAN NETWORKS

• X1, … ,Xn random variables associated with the vertices of a DAG = Directed Acyclic Graph

• The local conditional distributions P(Xi|Xj: j parent of i) are the parameters of the model. They can be represented by look-up tables (costly) or other more compact parameterizations (Sigmoidal Belief Networks, XOR, etc).

• The global distribution is the product of the local characteristics:

P(X1,…,Xn) = Πi P(Xi|Xj : j parent of i)

DATA PREPARATION

 Starting point: PDB data base.       Remove sequences not determined by X ray diffraction.       Remove sequences where DSSP crashes.       Remove proteins with physical chain breaks (neighboring AA

having distances exceeding 4 Angstroms)       Remove sequences with resolution worst than 2.5 Angstroms.       Remove chains with less than 30 AA.       Remove redundancy (Hobohm’s algorithm, Smith-Waterman,

PAM 120, etc.) Build multiple alignments (BLAST, PSI-BLAST, etc.)

SECONDARY STRUCTURE PROGRAMS

DSSP (Kabsch and Sander, 1983): works by assigning potential backbone hydrogen bonds (based on the 3D coordinates of the backbone atoms) and subsequently by identifying repetitive bonding patterns.

  STRIDE (Frishman and Argos, 1995): in addition to hydrogen bonds, it uses also dihedral angles.

  DEFINE (Richards and Kundrot, 1988): uses difference distance matrices for evaluating the match of interatomic distances in the protein to those from idealized SS.

SECONDARY STRUCTURE ASSIGNMENTS

DSSP classes: • H = alpha helix• E = sheet• G = 3-10 helix• S = kind of turn• T = beta turn• B = beta bridge• I = pi-helix (very rare)• C = the restCASP (harder) assignment: • α = H and G• β = E and B• γ = the restAlternative assignment: • α = H• β = B• γ = the rest

ENSEMBLES

Profiles n s f o W Q3 residue

No 7 2 8 11 1611 68.7% No 9 2 8 11 1899 68.8% No 7 3 8 11 1919 68.6% No 8 3 9 11 2181 68.8% No 20 0 17 11 2821 67.7% Output 9 2 8 11 1899 72.6% Output 8 3 9 11 2181 72.7% Input 9 2 8 11 1899 73.37% Input 8 3 9 11 73.4% Input 12 3 9 10 2757 73.6% Input 7 3 8 11 1919 73.4% Input 8 3 9 10 2045 73.4% Input 12 3 9 11 2949 73.2%

SSpro 1.0 and SSpro 2.0 on the 3 test sets, Q3

SSpro 1.0 SSpro 2.0R126 H 0.8079 0.8238

E 0.6323 0.6619C 0.8056 0.8126

Q3 0.7662 0.7813EVA H 0.8076 0.8248

E 0.625 0.6556C 0.7805 0.7903

Q3 0.76 0.7767CASP4 H 0.8386 0.8608

E 0.6187 0.6851C 0.8099 0.822

Q3 0.778 0.8065

FUNDAMENTAL LIMITATIONS

100% CORRECT RECOGNITION IS PROBABLY IMPOSSIBLE FOR SEVERAL REASONS

• SOME PROTEINS DO NOT FOLD SPONTANEOUSLY OR MAY NEED CHAPERONES

• QUATERNARY STRUCTURE [BETA-STRAND PARTNERS MAY BE ON A DIFFERENT CHAIN]

• STRUCTURE MAY DEPEND ON OTHER VARIABLES [ENVIRONMENT, PH]

• DYNAMICAL ASPECTS • FUZZINESS OF DEFINITIONS AND ERRORS IN

DATABASES

BB-RNNs

2D RNNs

2D INPUTS

• AA at positions i and j• Profiles at positions i and j• Correlated profiles at positions i and j• + Secondary Structure, Accessibility, etc.

PERFORMANCE (%)

6Å 8Å 10Å 12Å

non-contacts

99.9 99.8 99.2 98.9

contacts 71.2 65.3 52.2 46.6

all 98.5 97.1 93.2 88.5

Protein ReconstructionUsing predicted secondary structure and predicted contact map

PDB ID : 1HCR, chain ASequence: GRPRAINKHEQEQISRLLEKGHPRQQLAIIFGIGVSTLYRYFPASSIKKRMNTrue SS : CCCCCCCCHHHHHHHHHHHCCCCHHHHHHHCECCHHHHHHHCCCCCCCCCCCPred SS : CCCCCCCHHHHHHHHHHHHCCCCHHHHEEHECHHHHHHHHCCCHHHHHHHCC

PDB ID: 1HCR

Chain A (52 residues)

Model # 147

RMSD 3.47Å

Protein ReconstructionUsing predicted secondary structure and predicted contact map

PDB ID : 1BC8, chain CSequence: MDSAITLWQFLLQLLQKPQNKHMICWTSNDGQFKLLQAEEVARLWGIRKNKPNMNYDKLSRALRYYYVKNIIKKVNGQKFVYKFVSYPEILNMTrue SS : CCCCCCHHHHHHHHCCCHHHCCCCEECCCCCEEECCCHHHHHHHHHHHHCCCCCCHHHHHHHHHHHHHHCCEEECCCCCCEEEECCCCHHHCCPred SS : CCCHHHHHHHHHHHHHCCCCCCEEEEECCCEEEEECCHHHHHHHHHHHCCCCCCCHHHHHHHHHHHHHCCCEEECCCCEEEEEEECCHHHHCC

PDB ID: 1BC8

Chain C (93 residues)

Model # 1714

RMSD 4.21Å

CASP6 Self AssessmentEvaluation based on GDT_TS of first submitted model

GDT_TS: Global Distance Test Total Score

GDT_TS = (GDT_P1 + GDT_P2 + GDT_P4 + GDT_P8 ) / 4

Pn : percentage of residues under distance cutoff n

Hard Target Summary

N: number of targets predictedAv.R.: average ranksumZ: sum of Z scores on all targets in setsumZpos: sum of Z scores for predictions with positive Z score

group N Av.R. sumZ sumZposBAKER-ROBETTA 25 9.12 27.81 27.94baldi-group-server 24 10.04 20.47 22.33Rokky 25 11.60 17.56 18.46Pmodeller5 20 12.55 14.35 16.11ZHOUSPARKS2 25 15.00 12.67 15.91ACE 25 13.08 11.63 14.89Pcomb2 24 16.04 10.82 13.58RAPTOR 24 16.33 9.81 13.21zhousp3 25 15.72 9.30 12.90PROTINFO-AB 19 16.74 8.62 12.59

•Top 10 groups displayed, of 65 registered servers

•Assessment on 25 new fold and fold recognition analogous target domains

Hard Target Summary•Top 10 groups displayed, of 65 registered servers

•Assessment on 19 new fold and fold recognition analogous target domains less than 120 residues

N: number of targets predictedAv.R.: average ranksumZ: sum of Z scores on all targets in setsumZpos: sum of Z scores for predictions with positive Z score

group N Av.R. sumZ sumZposbaldi-group-server 19 6.74 20.61 20.61BAKER-ROBETTA 19 9.11 20.44 20.57Rokky 19 12.11 12.30 13.20PROTINFO-AB 16 12.63 11.53 12.59ZHOUSPARKS2 19 15.32 8.91 11.87Pcomb2 18 15.39 9.54 11.48Pmodeller5 15 14.47 9.25 11.00PROTINFO 18 16.22 8.66 10.56ACE 19 14.21 6.98 10.24RAPTOR 18 17.00 7.22 9.64

• Target Information– Length: 70 amino acids– Resolution: 1.52 Å– PDB code: 1WHZ– Description: Hypothetical Protein From Thermus Thermophilus Hb8– Domains: single domain

Target T0281Detailed Target Analysis

• Assessment– GDT_TS server rank of our 1st model: 2– GDT_TS: 51.07– RMSD to native: 6.15

Target T0281Contact Map Comparison

True Map vs. Predicted Map True Map vs. Recovered Map

*note: true map is lower left

Target T0281Structure Comparison

true structure predicted structure

Target T0281Structure Comparison Superposition

True structure: thick trace

Predicted structure: thin trace

• Target Information– Length: 51 amino acids– Resolution: 2.00 Å– PDB code: 1WD5– Description: Putative phosphoribosyl transferase, T. thermophilus– Domains: 2nd domain, residues 53-103 of 208 AA sequence

Target T0280_2Detailed Target Analysis

• Assessment– GDT_TS server rank of our 1st model: 1 (also 1st among human groups)– GDT_TS: 54.41– RMSD to native: 5.81

Target T0280_2Contact Map Comparison

*note: true map is lower left

True Map vs. Predicted Map True Map vs. Recovered Map

Target T0281Structure Comparison

true structure predicted structure

Target T0281Structure Comparison Superposition

True structure: thick trace

Predicted structure: thin trace

THE SCRATCH SUITEwww.igb.uci.edu

– DOMpro: domains– DISpro: disordered regions– SSpro: secondary structure – SSpro8: secondary structure– ACCpro: accessibility– CONpro: contact number– DI-pro: disulphide bridges– BETA-pro: beta partners– CMAP-pro: contact map– CCMAP-pro: coarse contact map– CON23D-pro: contact map to 3D– 3D-pro: 3D structure (homology + fold recognition + ab-initio)

SISQQTVWNQMATVRTPLNFDSSKQSFCQFSVDLLGGGISVDKTGDWITLVQNSPISNLLCCCECCCCCCEEEECCCCCCCCCCCCEEEEEEECCCCEEEECCCCCCEEEEECCHHHHHHCCCEEEEECEEEEECCCCCCCTCCCCEEEEEEEETCSEEEECTTTTEEEEEECCHHHHHH-----------+--------------+++++++++-+---------++++----++++++---------+-+++-----------++++++++++-+-++++---+++++++++++++++-----------+--++---------+++++++++----+++-----++++++++++++++---------+-++++++--------+++++++++---++++-----++++++++++++++eeeeee---e--e-e-eee-ee-eee---------e-e--eeeeee--------------

RVAAWKKGCLMVKVVMSGNAAVKRSDWASLVQVFLTNSNSTEHFDACRWTKSEPHSWELIHHHHHHCCCEEEEEEEEEECCEEECCCCCEEEEEEEECCCCCCCCCEEEEEECCCCCCCCHHHHHHTTCEEEEEEEEEEEEEEECCCCCEEEEEEEECCCTTCCCEEEEEEECCTCCEEE+++++--+++++++++++-+----------++++++---------+-+++----------+++----++++++++++++----------+++++++++------++++++++++-+-+--++++---+++++++++++++--+------+++++++++------++++++++++---+-+++++-+-++++++++++++----------+++++++++------++++++++++---+-+-----ee---e-------e-e-ee-e-e-e-----e--eeee--e-------e-e-ee-e

..

Solvent accessibility threshold: 25%PSI-BLAST hits : 24

..

Query served in 151 seconds

Advantage of Machine Learning• Pitfalls of traditional ab-initio

approaches• Machine learning systems take time to

train (weeks).• Once trained however they can predict

structures almost faster than proteins can fold.

• Predict or search protein structures on a genomic or bioengineering scale .

DAG-RNNs APPROACH

• Two steps:– 1. Build relevant DAG to connect inputs, outputs, and hidden

variables– 2. Use a deterministic (neural network) parameterization together

with appropriate stationarity assumptions/weight sharing—overall models remains probabilistic

• Process structured data of variable size, topology, and dimensions efficiently

• Sequences, trees, d-lattices, graphs, etc• Convergence theorems• Other applications

Convergence Theorems

• Posterior Marginals:σBNdBN in distributionσBNdBN in probability (uniformly)

• Belief Propagation:σBNdBN in distributionσBNdBN in probability (uniformly)

Structural Databases

• PPDB = Poxvirus Proteomic Database

• ICBS = Inter Chain Beta Sheet Database

Strategies for drug design

O

HN

O

NHR O

RR

NH

O

O

HN

O

NH RO

R R

NH

OHN

HN

O

NHR

HN

R O

O

HN

O

NHR O

RR

NH

O

O

HN

O

NH RO

R R

NH

protein 2

O

protein 1

HN

protein 1

HN

O

NHR

protein 2

HN

R O

O

HN

O

NHR O

RR

NH

O HN

O

NHR

O

HN

O

NHR O

RR

NH

O HN

O

NHR

HN

O OHN

HN

HN

O OHN

HN

-sheet mimic -sheet mimic

protein 1 protein 1

• Block, modulate, mediate β-sheet interactions

• [Covalent modification of a chain to prevent β-sheet formation]

Three-Stage Prediction of Protein Beta-Sheets Using Neural Networks, Alignments, and Graph

Algorithms

Jianlin Cheng and Pierre Baldi

School of Info. and Computer Sci.University of California Irvine

Beta-Sheet Architecture

Importance of Predicting Beta-Sheet Structure

• AB-Initio Structure Prediction• Fold Recognition• Model Refinement• Protein Design• Protein Stability

Previous Work• Methods

– Statistical potential approach for strand alignment. (Hubbard, 1994; Zhu and Braun, 1999)

– Statistical potentials to improve beta-sheet secondary structure prediction.(Asogawa,1997)

– Information theory approach for strand alignment. (Steward and Thornton, 2000)

– Neural networks for beta-residue contacts. (Baldi, et.al, 2000)

• ShortcomingsFocus on one single aspect; not utilize structural contexts and evolutionary information; not exploit constraints enough; not publicly available.

Three-Stage Prediction of Beta-Sheets

• Stage 1 Predict beta-residue pairings using 2D-

Recursive Neural Networks (2D-RNN).• Stage 2 Align beta-strands using alignment algorithms.• Stage 3 Predict beta-strand pairs and beta-sheet

architecture using graph algorithms.

Dataset and StatisticsNum

Chains 916

Betaresidues

48,996

ResiduePairs

31,638

BetaStrand

10,745

StrandPairs

8,172

BetaSheet

2,533

Stage 1: Prediction of Beta-Residue Pairings Using 2D-RNN

Input Matrix I (m×m)

2D-RNNO = f(I)

Target / Output Matrix (m×m)

(i,j)

i-2 i-1 i i+1 i+2 j-2 j-1 j j+1 j+2 |i-j|

20 profiles 3 SS 2 SA

Tij: 0/1Oij: Pairing Prob.

(i,j)

Iij

Total: 251 inputs

An Example Target

Protein 1VJGBeta-Residue Pairing Map (Target Matrix)

An Example Output

Stage 2: Beta-Strand Alignment

• Use output probability matrix as scoring matrix

• Dynamic programming

• Disallow gaps and use simplified searching algorithms

1 m

n 1

1 m

1 n

Anti-parallel

Parallel

Total number of alignments = 2(m+n-1)

Strand Alignment and Pairing Matrix

• The alignment score (Pseudo Binding Energy) is the sum of the probabilities of paired residues.

• The best alignment is the alignment with maximum score.

• Strand Pairing Matrix.

Strand Pairing Matrix of 1VJG

Stage 3: Prediction of Beta-Strand Pairings and Beta-Sheet Architecture

Strand Pairing Constraints

Minimum Spanning Tree Like Algorithm

Strand Pairing Graph (SPG)

Goal: Find a set of connected subgraphs that maximize the sum of pseudo-energy and satisfy the constraints.

Algorithm: Minimum Spanning Tree Like Algorithm.

Example of MST Like Algorithm

0

1.3 0

.94 .37 0

.02 .02 .04 0

.02 .02 .03 1.9 0

.10 .05 .74 .04 .04 0

.02 .02 .03 .02 .02 .20 0

1234567

1 2 3 4 5 6 7

4 5

Strand Pairing Matrix of 1VJG

Assembly of beta-strandsStep 1: Pair strand 4 and 5

Example of MST Like Algorithm

0

1.3 0

.94 .37 0

.02 .02 .04 0

.02 .02 .03 1.9 0

.10 .05 .74 .04 .04 0

.02 .02 .03 .02 .02 .20 0

1234567

1 2 3 4 5 6 7

4 5

2 1

Strand Pairing Matrix of 1VJGA

N

Assembly of beta-strandsStep 2: Pair strand 1 and 2

Example of MST Like Algorithm

0

1.3 0

.94 .37 0

.02 .02 .04 0

.02 .02 .03 1.9 0

.10 .05 .74 .04 .04 0

.02 .02 .03 .02 .02 .20 0

1234567

1 2 3 4 5 6 7

4 5

2 1 3Strand Pairing Matrix of 1VJGA

N

Assembly of beta-strandsStep 3: Pair strand 1 and 3

Example of MST Like Algorithm

0

1.3 0

.94 .37 0

.02 .02 .04 0

.02 .02 .03 1.9 0

.10 .05 .74 .04 .04 0

.02 .02 .03 .02 .02 .20 0

1234567

1 2 3 4 5 6 7

4 5

2 1 3 6Strand Pairing Matrix of 1VJGA

N

Assembly of beta-strandsStep 4: Pair strand 3 and 6

Example of MST Like Algorithm

0

1.3 0

.94 .37 0

.02 .02 .04 0

.02 .02 .03 1.9 0

.10 .05 .74 .04 .04 0

.02 .02 .03 .02 .02 .20 0

1234567

1 2 3 4 5 6 7

4 5

2 1 3 67Strand Pairing Matrix of 1VJGA

N

C

Assembly of beta-strandsStep 5: Pair strand 6 and 7

Beta-Residue Pairing Results

• Sensitivity = Specificity = 41%

• Base-line: 2.3%. Ratio of improvement = 17.8.

• ROC area: 0.86• At 5% FPR, TPR is 58% • CMAPpro: Spec. and Sens. is 27%. ROC

area:0.8. TPR=42% at 5% FPR.

Strand Pairing Results

• Naïve algorithm of pairing all adjacent strands– Specificity = 42%– Sensitivity = 50%

• MST like algorithm– Specificity = 53%– Sensitivity = 59%– >20% correctly predicted strand pairs are non-adjacent

strand pairs

Strand Alignment Results

Paring Direction

Align.All

Align.Anti-P

Align.Para.

Align.Bridge

Acc. 93% 72% 69% 71% 88%

On the correctly predicted pairs:

Pairing Direction

Align.All

Align.Anti-P

Align.Para.

Align.Bridge

Acc. 84% 66% 63% 66% 73%

On all native pairs:

•Pairing direction is 15% higher than of random algorithm.•Alignment accuracy is improved by >15%.

Application and Future Work• New methods for beta-residue pairings (e.g. Linear

Programming, SVM), and strand alignment and pairings. More inputs (Punta and Rost, 2005).

• Applications– AB-Initio Structure Sampling (beta-sheet)– Fold Recognition (conservation of beta-sheets)– Contact Map– Model Refinement (pairing direction/alignment)

• Web server and datasethttp://www.ics.uci.edu/~baldig/betasheet.html

A New Fold Example (CASP6)1S12 (T0201, 94 residues)

1 2 3 4 5

1 0 1.71 .05 .29 .33

2 0 .06 .41 .12

3 0 .22 .04

4 0 .53

5 0

CEEEEEECCEEEECCCCCCCCHHHHHHHHHHHHHHHHHHHHHHHEHHCCCCEEEEHHHHHHHHHHHHHHHHHHHHHHHHHCCCCEEEEEEECCC

Predicted: 1-2, 2-4, 3-4, 4-5

CEEEEECCCEEEEECCCCCHHHHHHHHHHHHHHHHHHHHCCCEEEEEECCEEEEEECCCCHHHHHHHHHHHHHHHHHHHHCCCCEEEEECCCCCC

True: 1–2, 2-4, 3-4, 1-5

True SS:

Predicted SS:

124

3

5

Rendered in Rasmol

Strand Pairing Matrix

ACKNOWLEDGMENTS• UCI:

– Gianluca Pollastri, Pierre-Francois Baisnee, Michal Rosen-Zvi– Arlo Randall, S. Joshua Swamidass, Jianlin Cheng, Yimeng Dou, Yann

Pecout, Mike Sweredoski, Alessandro Vullo, Lin Wu

– James Nowick, Luis Villareal

• DTU: Soren Brunak• Columbia: Burkhard Rost• U of Florence: Paolo Frasconi• U of Bologna: Rita Casadio, Piero Fariselli

www.igb.uci.edu/www.ics.uci.edu/~pfbaldi

1DFN| Defensin

Sequence with cysteine's position identified: MSNHTHHLKFKTLKRAWKASKYFIVGLSC[29]LYKFNLKSLVQTALSTLAMITLTSLVITAIIYISVGNAKAKPTSKPTIQQTQQPQNHTSPFFTEHNYKSTHTSIQSTTLSQLLNIDTTRGITYGHSTNETQNRKIKGQSTLPATRKPPINPSGSIPPENHQDHNNFQTLPYVPC[173]STC[176]EGNLAC[182]LSLC[18

6]HIETERAPSRAPTITLKKTPKPKTTKKPTKTTIHHRTSPETKLQPKNNTATPQQGILSSTEHHTNQSTTQILength: 257, Total number of cysteines: 5Four bonded cysteines form two disulfide bonds :173 -------186 ( red cysteine pair)176 -------182 (blue cysteine pair)

Prediction Results from DIpro (http://contact.ics.uci.edu/bridge.html)Predicted Bonded Cysteines:173,176,182,186Predicted disulfide bondsBond_Index Cys1_Position Cys2_Position1 173 1862 176 182

Prediction Accuracy for both bond state and bond pair are 100%.

A Perfectly Predicted Example

Sequence with cysteine's position identified: MTLGRRLAC[9]LFLAC[14]VLPALLLGGTALASEIVGGRRARPHAWPFMVSLQLRGGHFC[55]GATLIAPNFVMSAAHC[71]VANVNVRAVRVVLGAHNLSRREPTRQVFAVQRIFENGYDPVNLLNDIVILQLNGSATINANVQVAQLPAQGRRLGNGVQC[151]LAMGWGLLGRNRGIASVLQELNVTVVTSLC[181]RRSNVC[187]TLVRGRQAGVC[198]FGDSGSPLVC[208]NGLIHGIASFVRGGC[223]ASGLYPDAFAPVAQFVNWIDSIIQRSEDNPC[254]PHPRDPDPASRTHLength: 267, Total Cysteine Number: 11Eight bonded cysteines form four disulfide bonds: 55 ----- 71 (Red), 151 ----- 208 (Blue), 181 ----- 187 (Green), 198 ----- 223 (Purple)

A Hard Example with Many Non-Bonded Cysteines

Prediction Results from DIpro (http://contact.ics.uci.edu/bridge.html)Predicted Bonded Cysteines:9,14,55,71,181,187,223,254Predicted Disulfide Bonds:Bond_Index Cys1_Position Cys2_Position1 55 71 (correct)2 9 14 (wrong)3 223 254 (wrong)4 181 187 (correct)Bond State Recall: 5 / 8 = 0.625, Bond State Precision = 5 / 8 = 0.625Pair Recall = 2 / 4 = 0.5 ; Pair Precision = 2 / 4 = 0.5Bond number is predicted correctly.

Bond Num

Bond State Recall(%)

Bond State Precision(%)

Pair Recall(%)

Pair Precision(%)

1 91 46 74 392 93 77 61 513 90 74 54 454 77 87 52 595 71 86 33 426 65 84 27 347 63 85 36 558 66 89 27 419 60 83 23 35

10 55 86 30 4511 62 86 34 4712 67 97 17 2315 50 94 27 5016 82 99 11 1317 61 96 22 3318 50 82 6 919 47 90 11 20Overall bond state recall: 78%; overall bond state precision: 74%;

bond number prediction accuracy: 53%; average difference between true bond number and predicted bond number: 1.1 .

Prediction Accuracy on SP51 Dataset on All Cysteines

CURRENT WORK

• Feedback:Ex: SS Contacts SS Contacts

• Homology, homology, homology:SSpro 4.0 performs at 88%