PhenoMining Services

PhenoMining Services provide a set of Web-based text mining tools for the Consortium for Neuropsychiatric Phenomics (CNP). PhenoMining services consist of the following seven tools: Top Concepts, Relational Concept Explorer, Assertion Mining, Category-based Query Expansion, Cross Species Concept Explorer, Figure & Table Mining and Path Mining. The services are based on a corpus, generated automatically from Pubmed Central via relevant CNP key words, and a lexicon, developed by domain experts. PhenoMining provides properties and relationship among phenotypes such as the statistics of concept terms used and their corresponding phenotypes in the corpus, the relations between concepts at different phenolevels and their conditional co-occurrence.

To get started, click a link in the top navigation bar, or learn more about each service by clicking the links in the list of services below.

• Top Concepts -- A text cloud of lexical concepts

• Top Concepts displays a cloud-like view of the lexica concepts to provide a visualization of the relative popularity of concepts and phenolevels in the corpus. The tool computes the statistical distribution of all lexica concepts in the corpus and then ranks the terms according frequency of occurrence. The size of the text in the cloud indicates the relative frequency of the concept (larger font is more frequent) and the colors indicate the PhenoLevel of each term.

• Relational Concept Explorer -- Discover which concepts are most related to a lexical term

• The Relational Concept Explorer accepts a concept as input, and then generates an ordered list of closely related concepts. The tool discovers related concepts by calculating co-occurrence statistics at a user-specified semantic granularity to provide a measure for expressing relations between concepts. The Concept Explorer can sort the results ordinally by co-occurrence, categorically or alphabetically.

• Assertion Mining -- Mine full-text for sentences and paragraphs relating two or more terms

• Assertion Mining reveals full-text relationships between lexica concepts chosen by the user. An assertion is a formal statement describes a fact about two or more concepts. For example, he following statement describes the concepts ‘working memory’ and ‘dopamine’.

  … working memory impairment in schizophrenia has been associated with abnormal dopamine signaling in the prefrontal cortex, which is likely under complex genetic control. (Joshua et al. 2008)

  Assertions in a document are derived from the PhenoMining Services concept index database; the index contains data about whether two concepts occur in the same statement.

• Query Expansion -- Formulate a complex phenotype query with category based expansion

• Query Expansion assists the user with the creation of complex queries by suggesting related terms to help the user generate a more focused and more accurate query. Searching for scientific papers with query terms is a challenge due to:
  
  Mismatch of query terms and document terms; query expansion may be used to remedy this problem, and category-based query expansion [3] uses semantic information and yield better performance than statistical expansion. Note the following example shows that the expansion terms of dopamine for different Phenotypes (e.g. brain region, cognition and gene) level are quite different from pure statistical co-occurrence.

  

• Cross Species Concept Explorer -- Explore the relationship of lexical terms across species

• This tool aids researchers by aligning similar concepts in different species. For example, it is difficult to relate aspects of human cognition, such as language processing, to cognition in other species. The cross-species concept explorer leverages commonalities found in the biology supporting cognition, such as genes, cellular signaling and neurobiology, to reveal similarities across species, as shown in the following example.

  

• Figure & Table Mining -- Search for figures and tables

• Figure and Table Mining locates and displays relevant data in the corpus by mining text in figure and table legends for co-occurrence of serach terms. Researchers are often interested in data and fact discovery reported in the literature, and these are concisely captured in data tables and figures. Figure and table legends provide a concise description of results, and PhenoMining indexes lexical terms in these legends to discover relevant results based on the co-occurrence of terms in the legend text.

• Path Mining -- Connects concepts across the CNP hierarchy

• PhenoPaths are sequences of related concepts involving intermediate phenotypes in a Multi-level Phenome (a phenotype hierarchy).
  e.g. Schizophrenia -- working memory -- prefrontal cortex -- drd2
  

  Path Mining takes a phenome query, specified by multiple lexica concepts, and generates a graph (PhenoGraph) of PhenoPaths
  e.g. Syndrome -- cognition concepts -- neuro systems -- Gene level
  

  The path mining tool uses conditional co-occurrence of concept terms as a measure to form a PhenoPath which can be either at the sentence, paragraph, or article level.
  The conditional co-occurrence probability of Ci with C1, C2, … Cn
  
  

  Pr (Ci | C1, C2 ... Cn) = |Ci co-occurs with C1, C2, … Cn| / |C1, C2, … Cn| is a metric for Ci

  The following example shows a four level PhenoGraph for concepts in the phenolevels Syndrome, Cognition, Neuro systems, and Gene :

  

  A PhenoPath is a path in the PhenoGraph. PhenoMining yields a set of paragraphs or sentences from the corpus that match the Phenome. A sample of PhenoPath mining result is shown in the following screen shot. The phrase ‘dopamine(7, 0.0310)’ in the result means that 7 paragraphs in the corpus matched the phenome (Schizophrnia, working memory, pfc and dopamine receptor) with a conditional concurrence probability of 0.0310.

  

Document Content Explorer (beta) -- Mining the core details from documents, integrate with other informatic systems such as PhenoWiki
• Phenomining -- Lexica, corpus, and indexing

• PubMed query is widely used by researchers. However, it has many limitations. For example, PubMed search results do not rank retrieved articles, the results do not display relevant portions of returned articles, and the search does not allow rapid viewing of relevant figures and tables within the returned articles. Further, PubMed's search can not retrieve documents by Phenotype nor can it seek matches for a Multi-level Phenome -- a set of Lexica for each level in the phenotype hierarchies [1,2]. The PhenoMining services use techniques from text mining to remedy these shortcomings.

  Lexica and Corpus

  The text-mining techniques employed by the phenomining services require both a lexicon of concept terms and a corpus of text. The lexicon of concept terms was developed by domain experts within the CNP to capture key concepts most relevant to the data collected for the CNP's LA2K project. The lexicon represents a Multi-Level Phenome, comprising 61 concepts in six levels -- syndrome, symptom, cognitive concept, task, brain region, gene ( link to the hierarchical lexicon ). As used here, the word “concept” refers to a phenotype idea, represented by a list of terms chosen by domain experts to be considered synonyms by the mining software. For example, ‘basal ganglia’ and ‘striatum’ refer to the same anatomical idea for the purposes of the active lexicon. The list of the concepts and their occurence statistics can be accessed in XML format here .

  The corpus was derived from PubMed Central (PMC). The domain experts formulated the following query to retrieve a full-text articles from PubMed Central :

  ("Schizophrenia" OR "Bipolar Disorder" OR "Attention Deficit Disorder") OR ("Memory" OR "Response Inhibition" OR "executive function") OR ("Prefrontal Cortex" OR "Striatum" OR "caudate" OR "PFC" OR "Basal Ganglia")AND ("Neuregulin" OR "DISC1" OR "synapsin" OR "Dopamine receptor" OR "glutamate receptor")

  PubMed Central returned 7700 full-text documents from the query above, providing a rich and CNP-relevant corpus for PhenoMining. The contents of PMC constantly evolves as new research is published. Likewise, the terms relevant for the PhenoMining lexica evolve as our domain experts achieve new understanding and insight. PhenoMining will also evolve with automated periodic updates of the corpus and with domain expert revisions of the lexica.

  Indexing the Corpus

  Indexing the concepts from the lexica in the corpus requires locating each concept in relation to other lexica keywords. In order to provide maximum flexibility, the PhenoMining Services find relationships across semantic structures ranging from sentence to article.

  The full text papers from the PMC are in HTML format, and HTML does not natively capture all the granularities of semantic relationships pertaining to scientific research articles. To compensate for these shortcomings of HTML, the articles were translated to XML, and then indexed for the lexica concepts. Based on user input, the text mining tool searches the index database to derive the mining results.

  References

  1. 1. Wesley W. Chu, PhenoMining Services, 2010. (ppt)
  2. 2. Ying Wang, Wesley Chu, PhenoMining: Text Mining Services for CNP. (pdf)
  3. 3. Ying Wang, Chen Liu, Wesley W Chu, Fred Sabb, Stott Parker, and Andrew Howe, PhenoMining: Text Mining Tools for CNP, 2010. (ppt)
  4. 4. F W Sabb, C E Bearden, D C Glahn, D S Parker, N Freimer, R M Bilder. A collaborating knowledge base for cognitive phenomics, Molecular Psychiatry, 2008.
  5. 5. Bilder, Robert M., Sabb, Fred W., Parker, D. Stott, Kalar, Donald, Chu, Wesley W., Fox, Jared, Freimer, Nelson B. and Poldrack, Russell A. (2009) 'Cognitive ontologies for neuropsychiatric phenomics research', Cognitive Neuropsychiatry,14:4,419-450.
  6. 6. Zhenyu Liu and Wesley Chu. Knowledge-based query expansion to support scenario specific retrieval of medical free text. In Information Retrieval. Springer, 2007.

To learn more about the PhenoMining Services, please read the following materials:

1. 1. Wesley W. Chu, PhenoMining Services, 2010. (ppt)
2. 2. Ying Wang, Wesley Chu, PhenoMining: Text Mining Services for CNP. (pdf)
3. 3. Ying Wang, Chen Liu, Wesley W Chu, Fred Sabb, Stott Parker, and Andrew Howe, PhenoMining: Text Mining Tools for CNP, 2010. (ppt)

This application is part of the Consortium for Neuropsychiatric Phenomics and developed by Cobase Group of Computer Science Department , University of California, Los Angeles .