SSASS: Single Spectrum ASSignment of Protein Backbone Chemical Shifts

1
SSASS Single Spectrum ASSignment of Protein
Backbone Chemical Shifts

• Shan Sundararaj
• November 19, 2004

2
NMR Analysis of Protein Structure
Data Collection (days)
Data Processing (hours to days)
Sequential Chemical Shift Assignment
NOESY assignment
Structure Building
3
NMR Structure Determination

• Assign chemical shifts using HSQC and 3D
  experiments
• Use assigned chemical shifts to assign NOESY data
• Use distance constraints from NOESY data to
  assemble structure

4
NMR Structure Determination
Assigned Spectra Assigned NOESY 3D
Structure
12
123
34
43
76
45
16
77
11
99
52
23
89
88
12
78
123
79
33
76
34
9
44
32
43
76
45
16
51
44
32
24
76
45
16
56
11
53
32
98
76
45
47
56
53
11
89
52
89
47
77
49
89
52
66
33
47
76
56
44
52
133
44
143
50 ms 100 ms 250 ms
HSQC HNCO HNCA HNHA
CBCACONH HNCACB CACONH HACONH etc.
5
Can We Do It Even Faster?
HNCACB

• Collect just 1 HNCACB spectrum
• Determine structure without NOEs

Single Spectrum Spectroscopy
6
Chemical Shift Assignment

• Currently a slow, tedious process
• Several multidimensional spectra
• HSQC-TOCSY, CBCA(CO)NH, HNCA, etc.
• Processing, peak-picking for each spectrum
• Use HSQC to select 15N and 1H shifts, then use
  other experiments to pick out likely spin systems
• Assemble spin systems based on connectivity in
  various experiments ? assign chemical shifts to
  residues

7
Another Way?

• Chemical shifts themselves are directly
  influenced by structure
• Several tools to infer chemical shifts from
  structure or homology
• Manual assignment can be automated

8
SHIFTY

• Searches RefDB for homologous proteins already
  assigned by NMR
• Higher sequence homology ? more accurate
  prediction of chemical shifts

9
SHIFTX

• Given a structure or model of a protein, predicts
  chemical shifts based on local structure of each
  residue

10
SimPred

• Predicts chemical shifts based on secondary
  structure prediction

11
Automated Assignment

• Can use predicted chemical shift assignments as a
  guide to make assignments from real spectra
• What is the best method to get enough real
  experimental data to make assignments?
• Use single HNCACB experiment to generate all the
  shifts and connectivity information (Single
  Spectrum ASSignment SSASS!)

12
HNCACB Experiment

• First 2 dimensions are amide proton/ nitrogen
• Third dimension contains 13C chemical shifts
• Ca and Cb resonances of a residue AND of the
  residue before it

13
HNCACB Connectivity

• Each amide nitrogen couples more strongly to its
  own Ca than to the Ca of previous residue (higher
  intensity)
• Correlation of Ca and Cb will have opposite signs
  (i.e. Ca has ve intensity, Cb has ve)
• Each spin system should have 4 C shifts
• 
  
• CAi CBi CAi-1
  CBi-1

14
HNCACB Correlations
http//www-bioc.rice.edu/mev/spectra2.html
15
HNCACB Correlations
Residue i
Residue i-1
N 128.27
N 124.52
16
Chemical Shift Assignment

• HN NH CA CB CA-1 CB-1
• 7.91 105.2 45.1 0.0 55.3 33.1
• 8.43 111.3 61.2 68.3 45.0 0.0
• 8.21 114.9 56.4 32.6 61.2 68.2
• 7.92 122.4 52.6 36.9 56.5 32.7
• 7.59 128.1 60.5 41.8 52.6 36.9
• 8.61 124.5 50.8 38.7 60.3 41.9
• 8.92 128.3 56.4 36.5 50.8 38.7
• 8.46 111.9 58.1 38.2 56.7 0.0

G T E K I N M R
17
SSASS Algorithm

• Two initial, parallel steps
• Predict what true chemical shifts might be
• Process peak list to build spin systems
• Make best (most complete) assignment of spin
  systems to the protein sequence

18
SSASS Algorithm
INPUT
PROCESS SPINS
PREDICT
MAKE BLOCKS
ASSIGN SCORE
ASSIGNMENT
19
SSASS Algorithm
Protein Sequence Peak File

• 1) PsiPred prediction
• 2) Homodeller model structure
• 3) Predict chemical shifts

1) Filter shifts (peak selection) 2) Form
refine spin systems 3) Form blocks of spins
1) Score Assign highest scoring blocks 2) Fill
in blanks with singleton systems
Protein with assigned backbone chemical shifts
20
Input

• Input to SSASS
• the sequence of the protein
• an automatically picked peak list from an HNCACB
  experiment (from VNMR, NMRView, NMRPipe, SPARKY,
  etc.)
• Manually picked lists can be used as well, of
  course

21
Predict
gt35 ID to BMRB entry?
SHIFTY
gt95 ID to PDB entry, or model provided?
SHIFTX
Sequence
gt35 ID to PDB entry?
Homodeller SHIFTX
lt35 ID to PDB entry?
PsiPred SimPred
22
Peak List Processing

• Initial peak list should ideally be at low
  threshold cutoff (slightly noisy data more
  likely to contain more real peaks)
• All peaks are sorted by 15N dimension, then 1H
  dimension
• Filtering script (shifts) then removes spurious
  peaks (phasing and spectral artifacts, duplicated
  peaks, solvent peaks, side chain traces)
• No need if peaks are manually picked

23
Spin System Formation

• Remaining peaks are looked at iteratively to make
  spin systems, starting with the most intense peak
  (spin) stopping when 110 of expected systems
  found
• NOTE spin systems may be missing peaks or
  have overlapped peaks or be overlapped systems

24
(No Transcript)
25
Spin System Refinement

• Those systems that cant unambiguously be broken
  down into CAi, CBi, CAi-1, and CBi-1, are then
  combinatorially made into a list of alternate
  systems (list)
• If the two CA peaks are of similar intensity, two
  systems are made
• If more than 4 peaks are found, all possible
  systems of 4 are made

26
Spin System ? Amino Acid

• Assign the likelihood that each system
  corresponds to a type of amino acid
• Use distributions of known amino acid chemical
  shift information (calculated for coil, helix,
  and strand)
• we assume normal distributions
• e.g. CA 46.0 ? Glycine

27
Connectivity

• Use the info in CAi-1, and CBi-1 to make blocks
  of possible connected spin systems
• Algorithm uses breadth first search of
  connectivity, with pruning (i.e. all possible
  blocks of any length are considered, unless the
  level of connectivity drops below a certain
  threshold)
• 100s to 1000s to 100000s possible

28
Chemical Shift Assignment

• Each block is aligned to the sequence and a
  fitness score is calculated for it
• Score based on
• RMSD to predicted shift for each residue
• Length of block being evaluated
• Connectivity within block
• Intensity of spin systems
• Fit of shifts to sequence and 2o structure

29
Block Alignment
Predicted Shifts
Block
. . N P G T E I L A K L E L Q G . .
8.23 115.2 55.3 41.2 9.78 121.8 64.3 33.7 7.77
103.2 44.9 0.0 8.02 109.2 62.2 65.1 8.56 117.7
57.4 33.9 7.49 124.9 53.1 38.5 7.52 128.2 59.1
31.8 8.44 127.8 53.6 18.3 7.77 131.1 55.4
34.6 8.28 109.9 58.1 38.5 8.01 115.3 53.9
32.2 8.64 122.4 54.8 37.4
HN NH CA CB CA-1 CB-1 7.91 105.2 45.1
0.0 0.0 0.0 8.43 111.3 61.2 68.3 45.0 0.0 8.21
114.9 56.4 32.6 61.2 68.2 7.92 122.4 52.6 36.9
56.5 32.7 7.59 128.1 59.5 31.1 52.6 36.9 8.98
125.1 55.1 17.8 59.4 31.1 7.54 130.5 55.8 34.6
0.0 17.7 8.46 111.9 58.1 38.2 55.8 0.0
C C C C H H H H H H H H C C
30
Output Best Assignment!
31
Evaluation

• Used test set of 6 proteins of varying secondary
  structure
• MT0807
• MT0895
• pointed domain of ets-1
• CDC4
• UBC-13
• mms2

32
MT0807 (85 AA)
MT0895 (77 AA)
Ets-1 (110 AA)
Ubc-13 (151 AA)
Cdc4 (141 AA)
Mms2 (145 AA)
33
Spin System Analysis
34
Accuracy (No Homology)
35
Accuracy (Using SHIFTX/Y)
36
Effect of Homology Modeling
37
Effect of Homology Modeling
38
Current State

• Performance is best on smaller proteins (lt100
  residues)
• Performance suffers greatly when there is a lack
  of connectivity information in HNCACB
• Performance improves when results from homology
  are included

39
One Spectrum Not Enough?

• Can also handle input from HNCA and CBCA(CO)NH
  experiments ( HNCO)
• More proteins to test
• Troponin C (90 res)
• GA binding protein (91 res)
• Carnobacteriocin immunity protein (111 res)
• TEM-1 lactamase (263 res)

40
Future

• Improve rating of spin systems (relative
  intensity of peaks, order of peaks, overlapped
  peaks)
• Improve assignment process (build and assign
  blocks of most intense peaks first, repeat
  assignment at lower match thresholds, improve
  search algorithm)
• Optimize assignment scoring (best mix of homology
  prediction, secondary structure, spin
  connectivity, spin intensity) confidence score,
  explanation
• Allow more user control (manually set some
  assignments and re-assign protein)

41
Acknowledgements

• Dr. Hassan Monzavi
• Haiyan Zhang
• Trent Bjorndahl
• G. Amegbey
• Nelson Young
• Questions?