The TEXTAL System for Automated Model Building

1
The TEXTAL System for Automated Model Building

• Thomas R. Ioerger
• Texas AM University

2
Role of Automated Model Building

• input map gt output model (coords)
• goal automation
• what level is possible? need for human
  judgement/correction for difficult cases?
• incorporation in systems like PHENIX
• use on beam-lines
• detection of NCS molecular replacement
• iteration with phase improvement (Resolve)

3
The TEXTAL Approach

• Based on pattern recognition
• Consider a spherical region of 5Å radius...
• Have I ever seen a region of density similar to
  this in any previously-interpreted map?
• if so, use coordinates of atoms from matched
  region, translated and rotated
• metric density correlation, but must be
  rotation-invariant (optimize orientation)

4
Feature Extraction

• faster distance metric
• weighted Euclidean distance of feature vectors
• examples of (rotation-invariant) features
• standard deviation, other statistics in region
• distance to center of mass
• moments of inertia, ratios (for symmetry)
• search a database of regions from solved maps,
  with features extracted off-line

5
Outline of the Process
elec. dens. map
C-alpha chains (PDB file of predicted CA coords)
CAPRA
calculate features in 5A region around each
C-alpha search database for matches
1. sequence alignment 2. real-space refinement 3.
heuristics to fix backbone
LOOKUP
Post-Processing
initial model (complete coords)
structure factors (with est. phases)
atomic coords
6
CAPRA C-Alpha PatternRecognition Algorithm

• 1. Map scaling - adjust density so on average,
  gt1.0 captures to 20 of volume, lt-1.0 capture
  bottom 20
• 2. Tracing - skeletonization - pseudo-atoms on
  0.5A grid eliminate lowest density pts first
  dont break connectivity
• 3. Calculate features for 5A region around each
  pseudo-atom
• 4. Use neural network to predict distance to
  nearest C-alpha
• trained on features from random pts in 1A contour
  of known map
• 5. Select way-points predicted closest locally,
  gt2.5A apart
• 6. Link way-points together into C-alpha chains
• consider quality of neural net prediction
• prefer longer chains dont break off into
  side-chains
• take secondary structure into account
  straightness and helicity

7
Examples of CAPRA Steps
8
Example of CA-chains fit by CAPRA
9
Example of Models Built by Textal
10
Future Work

• correction by sequence alignment
• characterizing accuracy of Textal as function of
  resolution, phase quality
• at what point (of refinement) will it work?
• how well will it work? (rmsd, Derrph)
• iteration with phase improvement

11
Potential Points of Collaboration

• Tracer as a tool (and density scaling?)
• Using model-building for NCS detection, mask
  generation
• Interaction with solvent-flattening

12
Acknowledgements

• James C. Sacchettini
• Kreshna Gopal
• Reetal Pai
• Tod Romo
• funding from National Institutes of Health