Recent developments in the CLiDE tool for extraction of chemical
structure data from patents and other documents
Aniko T. Valko
Keymodule Ltd.
Peter Johnson Vilmos A. Valko
Summary
1) About CLiDE What is CLiDE for?
2) Performance against a benchmark set of images About the benchmark set Performance of CLiDE Enhancements made in CLiDE Comparison with selected systems
3) Performance against selected patents About patents Performance of CLiDE Comparison with selected systems
4) Conclusions and future work
Part 1:
About CLiDE
What is CLiDE for?
CLiDE is an Optical Chemical Structure Recognition (OCSR) software application, aimed at converting structure diagrams to computer-readable structures (i.e. connection tables)
PDF, DOC, DOCX, HTML
BMP, GIF, JPEG, PBM, PGM, PNG, PNM,
PPM, TIFF, XBM, XPM
Molfile, RGfile, SDfile CDX, CML, MRV
XML
Valence-violated atom
Non-interpreted atom
Clashing atoms
Small bond angle
Atoms at which CLiDE broke up the structure
Part 2:
Performance against a benchmark set of images
Benchmark set
● Images of isolated structures, one structure per image
● #images: 5735
● US Patent Office Complex Work Unit
US07321045-20080122-C00150 US07320974-20080122-C00070 US07323286-20080129-C00108 US07317070-20080108-C00008 US07316739-20080108-C00281 US07314700-20080101-C00001 US07320972-20080122-C00016 US07314876-20080101-C00035 US07314576-20080101-C00035 US07314511-20080101-C00002
● Available on the OSRA web site
● Verification set: Each image is associated with a Molfile meant to describe the correct connection table
Test runs on the benchmark set
● Test environment
CPU 3 Ghz Core 2 Duo
Memory 4GB
Linux distribution Ubuntu 10.0.4 (64-bit)
● Test run per image
1) CLiDE was run on an image
2) CLiDE analysed the image and generated a connection table
3) The connection table extracted by CLiDE was compared (using canonical SMILES) to the corresponding connection table from the verification set (so called ‘ground truth’)
● Performance measurements
1) Accuracy rate: the percentage of images that were correctly processed by CLiDE
2) Runtime: the total runtime measured over all the test runs
Performance against benchmark set
57,62%
59,00%
58,91%
59,30%
59,30%
81,81%
82,75%
84,60%
85,78%
86,55%
87,79%
87,96%
00:00
07:12
14:24
21:36
28:48
36:00
43:12
50:24
0%
10%
20%
30%
40%
50%
60%
70%
80%
90%
100%
3.2
.0
4.2
.0
4.4
.0
5.0
.0
5.2
.0
5.2
.1
5.4
.0
5.5
.0
5.5
.1
5.5
.2
5.5
.3
5.5
.4
Accuracy rate Runtime (min:sec)
Optimization and improvements in CLiDE’s document segmentation method (see later)
Auto correction of atom labels Better handling of aromatic rings Parsing chemical formulas Avoidance of loss of characters in atom labels Better handling of thick bonds Further improvements to chemical formula parsing 57.62% 87.98% 44 min 20 min
Enhancements in CLiDE
Corrections in atom labels 59.30% 81.81%
● Auto correction of OCR errors in atom labels
● Avoidance of misinterpretation of ‘Cl’ labels as Carbons
Enhancements in CLiDE
Chemical formula parsing 82.75% 84.60%
● Parsing the chemical formula into a sub connection table
Two-step process:
● Generating atom coordinates for the sub connection table
Super Atom Database: over 1000 super atoms, e.g. Me, Ph, Boc, TBDMS
Problem categories:
● Super atoms in chemical formulas ● Left- and right-aligned chemical formulas ● Branching in chemical formulas ● Chemical formulas with multiple attachments
● Chemical formulas with multiple attachments (―OCH2CH2O―)
● Super atoms in chemical formulas (―CO2Ph)
● Left- and right-aligned chemical formulas (―CH2NH2 vs NH2CH2 ―)
● Branching in chemical formulas (―OC(CH3)3)
Future work:
● Variables in chemical formulas (―CO2R, ―NHZ, ―SiR3)
Enhancements in CLiDE
Avoidance of loss of characters from atom labels
84.60% 85.78%
Enhancements in CLiDE
Better handling of thick bonds (stereo indicators)
85.78% 86.55%
Comparison with selected systems
74,48%
70,27%
68,68%
68,68%
43,20%
61,28%
0%
10%
20%
30%
40%
50%
60%
70%
80%
90%
100%
OSRA Imago
Accuracy rate
57.62%
87.96%
Runtime (hour:min)
05:51
04:50
04:50
04:54
00:15
01:52
0:00
1:12
2:24
3:36
4:48
6:00
7:12
OSRA Imago
00:44
00:20
Is the benchmark set correct?
Verification set #Molfiles to be corrected: 117
US07314693-20080101-C00370.TIF US07314693-20080101-C00370.MOL
● Anomalies: 10
● Stereo bonds: 22
US07316472-20080108-C00239.TIF US07316472-20080108-C00239.MOL
US07314872-20080101-C00024.TIF US07314872-20080101-C00024.MOL
● Incorrect sub connection tables for chemical formulae (e.g. NC, H3CO2S, OCF3): 63
● Errors in atom label: 14
● Other kinds of error: 17
Is the benchmark set correct?
Input images #images to be excluded: 16
● incorrect chemical formula: 1
US07314693-20080101-C00112.TIF
● disconnected atom: 1
● incorrect or ambiguous stereo bond: 6
US07314874-20080101-C00551.TIF USRE039991-20080101-C00187.TIF USRE039991-20080101-C00188.TIF
● arrow with unknown meaning: 8
US07320974-20080122-C00022.TIF
Performance after corrections
CLiDE 5.5.4 87.96% 90.11%
● #images: 5735
● #corrected Molfiles: 117
● #excluded images: 16
OSRA 1.4.0 68.68% 69.84%
Imago 2.0 beta 61.28% 61.91%
Part 3:
Performance against selected patents
About patents
Patents # non-Markush
structures
US6410540 218
WO2008099019 668
Challenges:
● Chemical structure diagrams have to be identified within the document page
● Interpretation of Markush structures
Markush structures were excluded from our tests
Challenge: Document segmentation
Page 65 of US6410540
Underlined text
5.5.4 4.4.0
Challenge: Document segmentation
Page 188 of WO2008019099 Table
5.5.4 4.4.0
Performance measurements
● Accuracy rate
● Runtime
● #Garbage structures: The number of structures that were assigned to non-chemical structure diagrams
Performance of CLiDE
88.90% 87,15%
80%
90%
100%
4.4.0 5.5.4
Accuracy rate
US6410540
00:22
00:06 00:00
00:14
00:28
4.4.0 5.5.4
Runtime (hour:min)
145
81
0
100
200
4.4.0 5.5.4
#Garbage structures
WO2008019099
57.63% 74,25%
0%
50%
100%
4.4.0 5.5.4
Accuracy rate
01:34
00:04 00:00
01:12
02:24
4.4.0 5.5.4
Runtime (hour:min)
1225
29
0
1000
2000
4.4.0 5.5.4
#Garbage structures
Comparison with selected systems Comparison with OSRA
87.15%
70,18%
60%
80%
100%
CLiDE OSRA
Accuracy rate
US6410540
00:06 00:09
00:00
00:07
00:14
CLiDE OSRA
Runtime (hour:min)
81
29
0
50
100
CLiDE OSRA
#Garbage structures
WO2008019099
74.25%
51,64%
40%
90%
CLiDE OSRA
Accuracy rate
00:04
00:58
00:00
01:12
CLiDE OSRA
Runtime (hour:min)
29
183
0
100
200
CLiDE OSRA
#Garbage structures
Part 4:
Conclusions and future work
Conclusions
● There has been considerable progress in OCSR, but nevertheless there still remain many problems to be solved
● The test sets showcased the diversity and the frequency of the problem types
● Regarding performance: • CLiDE has greatly improved during the last few years • CLiDE compares well with the other OCSR systems available to us for
testing
● In favourable cases, OCSR as exemplified by CLiDE now approaches OCR in accuracy (90%)
Future work
● Further improvements to structure recognition
● Filtering out garbage structures
● Identification and exclusion of non-chemical structure diagrams
● Further improvements to document segmentation
Short-term goals:
Long-term goals:
● Contextual document analysis, aimed at linking structures to text data
Flavours of CLiDE
CLiDE is released in three variants, designed for individual user needs
CLiDE Standard
CLiDE Professional
CLiDE Batch
Designed for the individual chemist who wishes to convert selected images into editable structures for use in reports etc.
GUI enterprise version to process whole documents with interactive editing
Unsupervised extraction for database creation etc.
Acknowledgment
Peter Johnson Keymodule Ltd. and University of Leeds
Anthony P. Cook University of Leeds
Vilmos A. Valko Keymodule Ltd.
Reseller agents • SimBioSys Inc. (North America) • NeoTrident Technology Ltd. (China) • Hulinks Inc. (Japan)
All users who gave us constructive feedback
Thank you for your attention