TEXTALTM : Artificial Intelligence Techniques for Protein Structure Determination

1
TEXTALTM Artificial Intelligence Techniques for
Protein Structure Determination
Kreshna Gopal, Reetal Pai, Thomas. R. Ioerger,
Tod.D.Romo, James. C. Sacchettini
Texas AM University
IAAI 2003 Acapulco, Mexico
2
Overview

• CAPRA
• LOOKUP
• Post processing Routines
• Results
• Discussion
• TEXTAL Availability
• Acknowledgements

3
CAPRA C? Pattern Recognition Algorithm
Map

• Scale input map to enable comparisons of
  patterns between
• different maps
• Trace gives connected skeleton of pseudo atoms
  through the
• backbone and the side chains
• The feature vectors are calculated at each of
  the tracer atoms

Scale
Tracer
Calculate features
4
CAPRA C? Pattern Recognition Algorithm

• The neural network has
• one input layer of 38 nodes
• hidden layer of 20 nodes
• one output node

Trained network parameters
Neural Network

• Neural network associates certain
  characteristics in the local density with an
• estimate of proximity to true C?
• The neural network predicts distances of pseudo
  atoms to true C?s

5
CAPRA C? Pattern Recognition Algorithm

• Select candidate C?s from all the pseudo atoms
  (trace)
• Use the distance predictions from the neural
  network

Selection of way-points

• Link all the C?s into linear chains
• Integrate intuitive criteria
• Identify the linearized sub-structure from the
  tracer graph

Build chains
C? chains
6
LOOKUP The Core Pattern Matching Routine

• Predicts co-ordinates of side chain atoms given
  location of C? atom
• Features used to determine regions with similar
  patterns of density
• Use of rotation invariant features
• Feature values calculated at different radii
• Database consists of feature vectors from regions
  within previously solved maps

7
LOOKUP The Core Pattern Matching Routine
TEXTAL uses weighted Euclidean feature difference
as measure of similarity
8
LOOKUP SLIDER

• SLIDER incrementally adjusts feature weights to
  increase matches and decrease mismatches
• For each region, a match and a mis-match are
  found and these 3-tuples are used to tune weights
• Weights can be optimized by finding a value
  between (0 ? 1) producing most positive
  crossovers
• Not guaranteed to find globally optimal weight
  vector, only a local optimum

9
Post Processing Routines

• To correct errors in stereochemistry
• Identify residues with backbone atoms in the
  wrong direction
• Identify correct backbone direction using a
  voting procedure
• Re-invoke LOOKUP
• Real Space Refinement
• Move atoms slightly to optimize their fit to
  density
• Preserve geometric constraints like bond
  distances and angles
• Sequence Alignment
• Corrects the identities of mislabeled amino acids
• Errors in LOOKUP output due to noise perturbing
  local density
• Errors corrected if the predicted fragment can be
  fit into actual sequence

10
TEXTALTM Input
08/01/03
Texas AM University
11
TEXTALTM Output
Mean residue density corr.
Mean length of output chains
Length of longest chain
No. of chains output
of structure built
All-atom rms error (Å)
Side chain structural similarity
C? rms error (Å)
Protein
A2u-globulin
2
88
68.5
85
0.85
0.84
0.99
48.9
Armadillo
9
217
46.7
89
0.98
0.82
N.A.
43.7
Cyanase
6
94
32.0
94
1.1
0.79
1.03
42.7
Gere
2
44
30.5
90
0.85
0.83
1.00
30.0
GM-CSF
4
46
25.0
82
0.91
0.84
0.94
28.9
Nsf-d2
6
79
39.5
92
0.96
0.83
1.13
33.5
Penicillopepsin
13
58
25.0
91
1.13
0.78
1.09
41.9
Psd-95
8
58
31.8
94
1.00
0.82
1.04
34.7
Rab3a
8
30
20.5
90
0.90
0.82
1.06
30.5
Rh-dehalogenase
8
66
36.5
97
0.92
0.83
0.99
54.6
CzrA
3
57
33.3
94
1.05
0.82
1.15
39.1
MVK
10
58
28.7
88
0.83
0.82
1.00
44.5
08/01/03
Texas AM University
12
CAPRA Output (CzrA)
08/01/03
Texas AM University
13
LOOKUP Output (CzrA)
08/01/03
Texas AM University
14
Discussion

• TEXTALTM has the potential to reduce one of the
  major bottlenecks in the way of high-throughput
  Structural Genomics
• Future Work
• Other feature comparison measures
• Clustering of database
• Reducing redundancy from database
• Stitching CAPRA chains

15
TEXTALTM Development

• Access to TEXTALTM through http//textal.tamu.e
  du12321
• PHENIX - Python-based Hierarchical ENvironment
  for Integrated Xtallography
• PHENIX Members
• Berkeley Lab
• University of Cambridge
• Los Alamos National Laboratory
• Texas AM University
• The alpha release of PHENIX is now available

16
Acknowledgements

• National Institute of Health
• TEXTALTM PIs
• Dr. Thomas R. Ioerger
• Dr. James C. Sacchettini
• TEXTALTM Development Team
• Kevin Childs
• Kreshna Gopal
• Reetal Pai
• Tod D Romo
• Vinod Reddy Melapudi