1. Crystallization Image Analysis on the World Community Grid
Christian A. Cumbaa and Igor Jurisica
Jurisica Lab, Division of Signaling Biology
Ontario Cancer Institute,
Toronto, Ontario

2. Why automate classification of protein crystallization trial images?
clear
phase separation
precipitate
skin
crystal X
garbage
unsure
Hauptman-Woodward has 65,000,000 images.
They want 65,000,000 outcomes.

3. Why automate classification of protein crystallization trial images?
Assist or replace human screening
Speed the search phase in protein crystallization
Improve throughput, consistency, objectivity
Enables data mining and statistical optimization of the crystallization process
clear
precipitate
crystal

4. Image classification clear phase separation precipitate skin crystal
feature extraction
classification
clear
phase separation
precipitate
skin
crystal X
garbage
unsure
feature 1
feature 2 …
feature k
100000s of numbers
7 numbers
10s of numbers

5. Truth data 96 study NESG & SGPP 50% unanimously-scored images
96 proteins X 1536 images hand-scored by 3 experts
Presence/absence of 7 independent outcomes
NESG & SGPP
15000 images
Hand-scored by 1 expert, same scoring system
50% unanimously-scored images
10 most interesting compound categories
96-study
NESG (crystals)
SGPP (crystals)

6. Feature set 12375 features computed per image
A few basic statistics
50 microcrystal features
Euler number features, two variations
11 Blur levels
11 Blur levels X 4 thresholds
Image “energy”
11 blur levels
2925 Grey-Level Co-occurrence Matrix features
3 different grey-level quantizations
13 basic functions
25 sample distances
~100 directions
Computable from every point in the image
Distilled to max range, max mean, min mean
~9500 image-blob features
Radon & edge-detection

7. Our image analysis problem
Computing all 12,375 features takes >5 hours for a single image
We have 165,000 images in our training set
Features must be evaluated for quality
The best features (10s or low 100s) must be computed for the remaining 65,000,000 images
Massive computing resources required!

8. Image analysis on the World Community Grid
a global, distributed-computing platform for solving large scientific computing problems with human impact
377,627 volunteers contribute idle CPU time of 960,346 devices.
Our project: Help Conquer Cancer*
launched November 2007.
HCC has two goals:
To survey a wide tract of image-feature space and identify image analysis algorithms and parameters (features) that best determine crystallization outcome.
To perform the necessary image analysis on Hauptman Woodward’s archive of 65,000,000 crystallization trial images.
* fundraising slogan of the Ontario Cancer Institute and its parent organization.

9. Image analysis on the World Community Grid
HCC has two phases
Phase I: calculate 12,375 features per image on high-priority images, including 165,441 hand-scored images.
November 2007-May 2008
analysis on hand-scored images completed January 2008
Phase II: calculate the best features from Phase I on the backlog of HWI images
Grid members have contributed 8,919 CPU-years so far to HCC, an average of 55 CPU-years per day.

12. Phase I: feature assessment

13. Measuring feature quality
feature entropy
Treat as random variables:
Image class
Feature value
Measure the mutual information between them (unit: bits)
= entropy(class) + entropy(feature) – entropy(class,feature)
class entropy

14. Measuring feature quality
clear
precipitate (no crystal)
other

15. Information density: microcrystal counts parameter space
Clear
Precipitate
Crystal

16. Information density: GLCM maximum range parameter space
Clear
Precipitate
Crystal

17. Information density: Radon-Sobel soft sum parameter space
Clear
Precipitate
Crystal

18. Information density: Radon-Sobel blob metrics (means) parameter space
Clear
Precipitate
Crystal

19. Towards Phase II: image classification

20. Building classifiers
handpicked 74 features from peaks in the clear, precipitate and other mutual information plots
two classification schemes
three-way: clear, non-crystal precipitate, other
ten-way: clear, phase separation, phase + precipitate, skin, phase + crystal, precip, precip + skin, precip + crystal, crystal, garbage
naïve Bayes model
leave-one-out cross-validation

21. Measuring classifier accuracy: precision and recall
crystals
false negatives
recall
“I think these are crystals”
precision
true
positives
false positives

22. Three-class distribution
Clear
24.3%
Precipitate AND NOT crystal
52.7%
Other
23.0%
17095
5258
5109
15928
45112
1819
617
817
27615
clear
non-crystal precipitate
other
non-crystal precipitate
machine says
true
class
Confusion matrix

23. Recall & precision

24. 10-class distribution Clear 33.83% Phase separation 7.00%
Phase separation + precipitate
0.50%
Skin
0.79%
Phase separation + crystal
2.32%
Precipitate
34.25%
Precipitate + skin
4.95%
Precipitate + crystal
7.53%
Crystal
8.34%
Garbage
0.55%

25. Confusion matrix machine says true class clear phase separation
313 20 2 52 1 49 4 28
129
3129
1072 90
219
649
586 56
345
888 8
914
2852
611
1063
562
111 85
222 35 29
305
395
2008
692
328
243 33
205 12
385
512
4088
3440
16907
553
617
494
1972
441 10
551
292 88 75
511 37
268 74
105 5 13 6
372
126 3 31
107 81 97 51 32 24 91
503
139
298
668
281 40
2433
1446
1193 92
815
1135
227
25585
clear
phase separation
phase and precipitate
skin
phase and crystal
precipitate
precipitate and skin
precipitate and crystal
crystal
garbage
machine says
true
class

26. Recall & precision

27. Acknowledgements Hauptman-Woodward Medical Research Institute
George DeTitta, Joe Luft, Eddie Snell, Mike Malkowski, Angela Lauricella, Max Thayer, Raymond Nagel, Steve Potter, and the 96-study reviewers.
World Community Grid
Bill Bovermann, Viktors Berstis, Jonathan D. Armstrong, Tedi Hahn, Kevin Reed, Keith J. Uplinger, Nels Wadycki
IBM Deep Computing:
Jerry Heyman
Jurisica Lab:
Richard Lu
All crystallization images were generated at the High-Throughput Screening lab at The Hauptman-Woodward Institute.
Funding from
NIH U54 GM074899
Genome Canada
IBM
NSERC
(and earlier work from)
NIH P50 GM62413
CITO