Automatic extraction of bioactivity data from patents

253rd ACS National Meeting, San Francisco CA, USA 4th April 2017

Automatic extraction of bioactivity data from patents

Daniel Lowe*, Stefan Senger† and Roger Sayle*

*NextMove Software Cambridge, UK †GlaxoSmithKline, Stevenage, UK


Example Use cases

• “A patent has recently come out on a topic of interest, can the key compounds be extracted with their activity data?”

• “Which compounds have been found to be active against this target?”


US Patent data freely available

patents.reedtech.com

(Or from the USPTO: bulkdata.uspto.gov)


= text-mined

What are these

compounds?


Understanding table semantics

SureChEMBL Google Patents After text-mining for chemical entities:

Green = substituent Purple = molecule

Source: US20170050925A9


SureChEMBL

Google Patents Patent PDF

PatFetch (NextMove Software) Source: US20010016661A1


Understanding table semantics

5 columns

6 columns

• Columns merged such that header and body have same number of columns


Getting the compound structures

• Chemical names

• Chemical sketches

• R-group tables

• Compound identifier associated with any of the above


Chemical names

• OPSIN (Open Parser for systematic IUPAC nomenclature)

• Dictionaries (ChEMBL/PubChem/NextMove)

• Chemical line formula parsing, especially useful for peptide names and R-group definitions


Chemical sketches

• Utilize the ChemDraw sketches provided by the USPTO

• Detection and handling of repeat brackets and positional variation

• Fixing obvious errors e.g. undervalent nitrogen near to H atom with no bond

• Labels reinterpreted


Formula Interpretation Input ChemDraw 15 This work

HATU

C4F9

H3PO4

CON(cHex)2 No result

III-2 No result

N

N+

O

NN

N

N

FP

-F

FF

FF

AT U

C C

FFF

F

F

FF F

F

FF

F F

FF

FF

F

O

N

PO

O

OOH

HH HO P

O

OH

OH

II

2-I


R-group tables


Resolving Identifiers

• Need to “name space” identifiers – “Compound 1”, “Reference compound 1”,

“Example 1” – But “Compound 1” = “cmpd 1” = “cpd. #1”

• Where a column is just called “#” is it a compound number, example number or just a table row number!

• Identifier may be defined multiple times e.g. as a sketch and chemical name


Resolving Identifiers (text-mining)


Resolving Identifiers (Sketches)


Resolving Identifiers (Tables)


Extracting compound-activity relationships


Excel table export


Extracting compound-activity relationships

What is the target?


Assay identification

• Naïve Bayes classifier trained from assay descriptions identified by BindingDB curators

• 10-fold cross validation: 98.9% recall, 94.7% precision

• Paragraph associated with next table or table mentioned in paragraph

• Target/organism detected • Care taken to avoid common irrelevant

organisms/proteins e.g. bovine serum albumin


Results


Results From US Patent applications (2001-Mar 2017)

Red = Bioactivity


Activities with associated structures per year

0

100,000

200,000

300,000

400,000

500,000

600,000

Activ

itty-

stru

ctur

e re

latio

nshi

ps e

xtra

cted

Publication Year


Comparison with BindingDB

• Activity data from ~1500 US patent grants (2013-2016) manually extracted over the course of 3 years

• ~150,000 activities • Comparison done on the subset that was made

available in ChEMBL 22_1 (98,898 activity values, 1012 patents)

• As some assay results are missed by the automatic extraction, and some are considered out of scope by BindingDB, difficult to distinguish differences in coverage from genuine disagreements


Comparison with BindingDB

• Values normalized into nM – 1000s of instances of measurements in nanometers!

• Mid point of ranges taken • Structures compared by StdInChI • Target name normalized to ChEMBL target ID

(organism specific), using either: – ChEMBL target synonyms – Normalize to HGNC symbol and check if HGNC symbol is a

ChEMBL target synonym


Comparison

Expected values found

Expected structures

found

Expected value +

structure found

Expected value +

structure + target

75% 65% 53% 18%


Unclear structure assignment

? ?


Stereochemistry and salts

OH

O

O

N

H

CH3H3CBr

H

H

Patent BindingDB This work


Long tail of difficult cases

What does this superscript term mean?

What are the units?


Targets of patent data compared to journal data

ChEMBL 22_1 (excluding BindingDB)

US Patent Applications Common Target Classes

0%

5%

10%

15%

20%

25%

30%

35%

40%

2002

2004

2006

2008

2010

2012

2014

2016

% p

er y

ear

Kinase

GPCR (Family A)

Protease

Nuclear receptor

Voltage-gated ionchannel

Electrochemicaltransporter

Oxidoreductase0%

5%

10%

15%

20%

25%

30%

35%

40%

2002

2004

2006

2008

2010

2012

2014

2016


Upcoming target classes

0.0%

0.5%

1.0%

1.5%

2.0%

2.5%

3.0%20

02

2003

2004

2005

2006

2007

2008

2009

2010

2011

2012

2013

2014

2015

2016

Perc

enta

ge o

f doc

umen

ts w

ith a

ctiv

ity v

alue

s aga

inst

ta

rget

cla

ss

Epigenetic writer (Patents)

Epigenetic reader (Patents)

Epigenetic writer (ChEMBL exBindingDB)

Epigenetic reader (ChEMBL exBindingDB)


Future work

• Support for more complex R-group tables • Improve recognition and resolution of protein

target names • Support for activities specified in text e.g.

Example 1 has an IC50 of 12 nM measured at rat EP4

• Resolution of symbols for activity ranges e.g. “A” indicates an IC50 value of less than 100 nM

• Improve assay metadata extraction cf. BioAssay Express


Disambiguation of Conflicting structure descriptions

Image from original filing

Redrawn by US patent office in

ChemDraw

Intended structure from chemical name


Conclusions

• Processing all US patents from 2001 to present can be done in less than a day on a desktop PC

• Technique applicable to chemical properties other than activity values

• Compound number <-> structure relationships useful for key compound identification

• For the majority of patents, extracting structure-activity relationships can be significantly expedited


Acknowledgements

• Noel O`Boyle • John Mayfield • Funding provided by:


Thank you for your time!

http://nextmovesoftware.com http://nextmovesoftware.com/blog

[email protected]

Automatic extraction of bioactivity data from patents

Software

Transcript of Automatic extraction of bioactivity data from patents