Post on 02-Jan-2016
Machete
Shannon Bradshaw, Marc Light, and Brian AlmquistDepartment of Management SciencesSchool of Library and Information ScienceDepartment of Linguistics Department of Computer ScienceThe University of Iowa
Reducing Info Management Problems
• Evolutionary biology – Many organisms– Many proteins– Many pathways
• Many information management problems
• A veritable goldmine for people like us
Knowledge Management (KM)
• Key idea: – Reduce duplicated effort in an organization or
community
• Simple example:– Bob has a question – An effective KM framework will point Bob to Alice
or Sharon who both know the answer and will share it– Ineffective KM would require Bob to invest a great
deal of time deciphering the answer for himself
• Want to reuse the experiences and previous efforts of a community to help an individual
Text Mining
• Extracting structured information from prose
• Example:– A table of protein-protein interactions
distilled from individual interactions described in sentences scattered across several documents
KM and Text Mining
• Large research communities in both spaces
• Want to interleave them in a single tool
• Targeted to bioscience literature
• We call this tool Machete
Example
• Context: – Experiments identified many genes that were
ankyrins or contained ankyrin repeats.
• Need: – Learn about ankyrins and ankyrin repeats
Knowledge Artifact
Using Artifacts: Personal level
Using Artifacts: Organizational level
Instead of doing the digging again
Lab members can reuse this
Using Artifacts: Community level
Finding documents
Reference Directed Indexing (RDI)
• Objective: To combine strong measures of both relevance and significance in a single metric
• Intuition: The opinions of authors who cite a document effectively distinguish both what a document is about and how important a contribution it makes
• Builds on the idea of using of anchor text to index Web documents
Example
• Paper by Andrade, Perez-Iratxeta, and Ponting on protein repeats
A single reference to Andrade
The ankyrin repeat motif mediates protein–protein interactions and is found in a diverse array of protein families, including transcription factors, cytoskeletal proteins, proteins which regulate development, and toxins (Andrade et al., 2001).
Leveraging multiple citations
• For any document cited more than once…
• We can compare the words of all authors
• Terms used by many referrers make good index terms for a document
• Phrases and statements in citation sentences bring to the surface important findings
Repeated use of words and phrases
Repeat proteins mediate numerous key protein–protein interactions in nature.[1. and 2.] Their repetitive architecture permits the adaptation of their size…
The ankyrin repeat motif mediates protein–protein interactions such as ankyrin and ß-propeller repeats [42]
Ankyrin repeats are thought to be important for protein–protein interaction events between integral membrane proteins and cytoskeletal proteins [Andrade et al., 2001].
The ankyrin repeat motif mediates protein–protein interactions
A voting technique
• RDI treats each citing document as a voter
• The presence of a query term in referential text is a vote of “yes”
• The absence of that term, a “no”
• The documents with the most votes for the query terms rank highest
Extraction possibilities
• In addition to retrieval, citation sentences may also provide a valuable source of data for information extraction
• However, for the time being we are focusing on the content of documents for extraction purposes
Finding information within documents
Text Mining
• Summarize gene function• Support for GO assignment• Speculative passages
PassageRetrieval
Retrieve Docs by First Finding Genes
• Associate words with genes• Collect word counts from user
query doc set• Return genes for which counts of
associated words went up• For each such gene, return docs
where associated words were found
Retrieve Docs by First Finding Genes(DNA AND repair)
c words
4 Lung6 CPD
c words
6 Lung2 CPD2 TTD
(STAT6 …
(XPD xeraderma…c words
3 mRNA4 IL-4
Look at the XPD gene and documents containing the Lung and CPD words
Text Mining
InfoExtraction
• prot interacts-with prot• prot located-in organella• gene associated phenotype
•Linguistics knowledgeAnkyrins bind to cell adhesion molecules of the CD44 family and the L1 CAM family…
This facilitates assembly of a repressor complex containing HDAC, Rb, and E2F that blocks transcription of the gene for IGF-1…
•Semantic knowledge –dictionaries and ontologies
•Counts –co-occurrence statistics–redundancy, e.g., that x interacts with y is mentioned 345 times
How It Works
Protein Interaction Extraction System We’re Building
• Inputs: – Pubmed query (“Ankyrins”) – List(s) of proteins
• Output:– Table of interacting protein pairs and links
Screen Shots
Clause-based Extraction
Collectively, these mutations also suppressed association of VDR with the coactivators GRIP1 and steroid receptor coactivator 1 in vitro but had little or no effect on ligand binding, heterodimerization with the retinoid X receptor, or association with a VDR-specific DNA recognition element
Method Section Mining
What fold concentrated Taq DNA polymerase buffer is optimal for the PCR reaction?
What plasmid DNA concentrations are needed for restriction digests?
In preparation for a Western blot, how long should GST lysate columns be incubated?
• We’re trying to build a system that can find answers to such questions
Dictionary Construction
• You people use so many words for the same thing: abbreviations, different uses of punctuation, totally different names– histone deacetylase 4, HDAC, HD, KIAA0288
• What is a poor computer to do?• Computers need synonym lists and other
information about words
The Info Is Out There(it just needs to be collated)
• Gene and protein entries in SwissProt, HUGO, GDB, OMIM, GenAtlas, LocusLink, InterPro have aliases
• They are all stored in different formats• They each contain some of the synonyms• They are only partially cross-referenced• Genes from non-model organisms are less likely to
be in some database somewhere (unless there is a homolog) (???)
• Grunt work is required
Nuts + bolts
PDF: Human sees
Machine sees
Challenges
• Each word/few words placed using x,y coordinates• Acrobat is just painting a picture. It has no sense of
the content of documents.• Difficult to:
– Follow flow of prose• Single or multi-column?• Some text spans multiple columns• Headers/footers
– Determine section breaks– Distinguish image/figure caption from body text– To parse bibliography entries
• Every document has a different layout format
Article has 3 columns, but text in PDF file may flow from left to right
Is this one block of text or part of two columns?
Is this part of the body or footer information?
Is this part of the article?
PDF Highlighting
• Multivalent Browser Annotations– Primarily useful for highlighting– Alternative annotations
• Highlighting with comments
– Stored separately from document• Local to user/machine
• How would this information be shared?
• Can they be “fused” with the document?
Multivalent Interface
Multivalent
• Highly extensible– We have some degree of freedom to modify– Interface is treated as part of the viewed
document
PDF: Inserting Hyperlinks
• Current system– Finds specified terms– Adds specified hyperlinks as an overlay over
each instance of a search term– Outputs modified PDF
<<links to Before/After files???>>
PDF: Inserting Hyperlinks
• Design Goals– Multi-platform support– Web-based interface
• Maintaining list of terms/URLs
• Submitting PDFs/URLs to URLs
– Extend to other forms of annotations
• Limitations– Certain PDFs cannot be converted to Text: (scanned
image, certain PostScript and DVI conversions)– Search is not robust: no hyphenations