Artificial Intelligence Lab

Current focuses of the Arizona BioPathway research include automatic extraction of regulatory pathway relations from biomedical literature using NLP techniques, inference of genetic networks from genomic data using data mining approaches, and the integration of existing knowledge and text/data mining results of regulatory pathways using a variety of biomedical ontologies.

The text mining component of Arizona BioPathway is designed to extract genetic regulatory pathway relations from biomedical literature. We have experimented with two different approaches of natural language processing (NLP) to extract the pathway relations, shallow parsing and full parsing. The shallow parser uses templates based on closed-class words (e.g., prepositions) and model generic relations to capture relations between noun phrases, while the full parser uses a broad coverage syntactic-semantic hybrid grammar to identify grammatical verb relations. To increase the precision, both approaches use relevant biomedical lexicons such as Gene Ontology (GO), HUGO Gene Nomenclature, and the Specialist Lexicon of UMLS to filter the extracted relations. We are also studying various statistical learning techniques for biomedical entity recognition and relation extraction from biomedical text.

The data mining component is designed to extract gene regulatory relations from genomic and proteomic data including DNA microarray by machine learning techniques such as Bayesian networks. We are experimenting various techniques to learn regulatory networks from microarray data, either with existing prior knowledge or in combination with other types of biological experimental data, e.g., DNA methylation array or protein expression. The so-called joint learning approach is promising to learn the network more accurately, avoiding bias and incompleteness inherited by a particular type of data. Linkages extracted from heterogeneous genomic data sources provide different evidence about gene functional relations. In a recent study, we develop a Bayesian framework for integrating relations extracted from multiple sources, such as gene expression, biomedical literature, and genomic sequence information, into a genome-wide functional network. In addition, we conduct studies on cancer classification using gene array data. We are adopting and developing various feature selection techniques to identify marker genes and their interactions for cancer diagnosis and drug discovery.

The knowledge integration component leverages a variety of biomedical ontology and knowledge sources to form an integrated framework for pathway-related knowledge organization. We have developed a feature decomposition approach to the aggregation of extracted pathway relations and resolution of the redundancy, ambiguity and inconsistency among them, using existing lexicons and ontologies such Entrez Gene, RefSeq, Homologene, MeSH, UMLS and GO. Pathway relations extracted from text and learned from data, as well as known relations from existing knowledge sources will eventually be integrated into a consolidated knowledge base.

All these pathway relations can be combined to construct regulatory networks and be visualized by automatic graph drawing algorithms implemented in the Arizona BioPathway Visualizer (see the demo).

Testbed:

Text mining (PubMed, 2003)

P53 – Text Collection:
Content: All abstract with p53 or related genes in title or abstract
Abstracts: 20,360
Linguistic Parser Relations: 194,384
Co-occurrence Relations: 2,724,099

AP1-Text Collection:
Content: All abstract with ap1 or related genes in title or abstract
Abstracts: 23,339
Linguistic Parser Relations: 258,142
Co-occurrence Relations: 3,265,524

Yeast-Text Collection:
Content: All abstract with yeast in title or abstract
Abstracts: 66,197
Linguistic Parser Relations: 584,502
Co-occurrence Relations: 6,535,737

Arabidopsis -Text Collection:
Content: All abstracts with MeSH terms of ‘Arabidopsis’ or ‘Arabidopsis Proteins’
Abstracts: 10,548
Linguistic Parser Relations: 222
Co-occurrence Relations: 1,291

Data Mining

P53 – Microarray Data:
Content: Gene expression measurement of p53 mutant cell lines (provided by AZCC)
Gene expression measurements: 33
Genes (Homo sapiens ORFs): 5,306
Genes with greatest variations: 200

Yeast – Microarray Data:
Content: Microarray data of yeast cell cycle (Spellman et al. 1998)
Gene expression measurements: 77
Time series: 6
Genes (S. cerevisiae ORFs): 6,177
Genes whose expression varied over the different cell-cycle stages: 800

Arabidopsis – Micrarray data:
Content: two high-quality microarray series of Arabidopsis at http://www.weigelworld.org
Gene expression measurements: 237 for development and 298 for abiotic stress
Genes (Arabidopsis): 22,810

Arabidopsis – Genome sequence relations:
Content: gene relations extracted from genome sequence using four different methods in ProLink (http://dip.doe-mbi.ucla.edu/pronav)
Relations:
Phylogenetic profiling (PP): 132,637
Rosetta Stone (RS): 989,795
Gene neighbor (GN): 18,823
Gene cluster (GC): 11,586

MDS – Microarray data
Content: DNA methylation arrays from Arizona Cancer Center. It is derived from the epigenomic analysis of bone marrow specimens from healthy donors and individuals with myelodysplastic syndrome (MDS).
Measurements: 55 (10 normal and 45 tumor samples)
Genes: 678

Ovarian Cancer – Microarray data
Content: microarray-based measurements of DNA methylation from the Gynecologic Oncology tumor bank at the University of Iowa and made available through the Arizona Cancer Center.
Measurements: 114 (25 normal and 89 tumor samples)
Genes: 6,560

Techniques:

A shallow parser based on closed class English words extracting noun phrase relations
A full parser using syntax-semantic hybrid grammar extracting verb relations
Co-occurrence analysis based on Concept Space, which generates asymmetric relations between phrases ordered according to the strength of their relation
Conditional Random Field (CRF) methods for entity recognition
Kernel-based learning methods for relation extraction and classification
Feature decomposition for entity and relation aggregation
Bayesian Network frameworks for integrating gene functional relations from multiple data sources
Optimal search based feature subset selection methods for identifying marker genes for cancer classification