Gene Clustering by Latent Semantic Indexing of MEDLINE Abstracts Ramin Homayouni, Kevin Heinrich, Lai Wei, and Michael W. Berry University of Tennessee presented by J. Jiang

Outline  Brief Overview of Biomedical Literature Mining  The Gene Clustering Problem  Latent Semantic Indexing  Experiments  Conclusions and Discussions

Biomedical Literature Mining Brief Overview  Goal: to find useful information from the large amount of biomedical literature  Tasks include: Identifying relevant literature for a given gene/protein Connecting genes with diseases Grouping genes/proteins by functions Reconstructing and predicting gene networks (ISMB 05’ Tutorial Proposal, H. Shatkay)

Biomedical Literature Mining Brief Overview (cont.)  Approaches: IE & NLP: entities, relations, facts, etc. Many methods rely on co-occurrences of genes/proteins. IR: text categorization and summarization, etc. Hybrid: combining multiple techniques  Challenges include: No fixed nomenclature or sentence structure Indirect links Etc.

The Gene Clustering Problem  To group genes based on their functions  Previous work: Co-occurrence of gene symbols to extract gene relationships Implicit textual relationships Gene clustering using functional information in annotated indices or MEDLINE abstracts

Vector Space Model for Gene Clustering  Glenisson et al., 2003  Bag-of-words, vector space model  Cosine similarity  K-medoids algorithm This paper tries to improve the vector representation of documents using LSA.

Background: LSA  First studied by Deerwester et al., Indexing by Latent Semantic Analysis, J Info Sci, 1990  Motivation: inaccuracy of term matching due to polysemy and synonomy  Assumption: existence of latent semantic structure (“artificial concepts”)  Dimension reduction. Keep the most important dimensions. Similar to PCA.

Singular Value Decomposition  d documents, t terms (in general, t >> d)  d  t matrix X = [x ij ], where x ij denotes the frequency of term j in document i  X can be decomposed as: X = T 0 S 0 D 0, where columns of T 0 are the eigenvectors of XX, and columns of D 0 are the eigenvectors of X X. S 0 is diagonal. S 0 2 is the matrix of eigenvalues of XX (or X X).

SVD (cont.)  The diagonal elements of S 0 are constructed to be positive and ordered in decreasing magnitude.

SVD (cont.)  The eigenvector with the largest eigenvalue represents the dimension along which the variance of the data is maximized.  Keep the k largest elements in S 0, remove other elements, and remove corresponding columns (eigenvectors) in T 0 and D 0, X can be approximated by: X  X hat = TSD.

SVD (cont.)  X hat is the best least-square-fit to X with rank k.

Illustration The first eigenvector The second eigenvector (taken from “A Tutorial on PCA” by Lindsay Smith)

LSA with SVD  Terms are represented by rows of X hat and documents are represented by columns of X hat in the reduced space.  Doc-to-doc similarity: X hat X hat = DS 2 D = DS(DS).  Query is represented as pseudo-document: D q = X q TS -1, where X q is the query vector in the original space. D q is like a row of D.  Query-to-doc similarity: D q S (DS).

Experiments  50 genes in (1) development, (2) Alzheimer Disease, and (3) Cancer Biology are selected  Gene-document: concatenation of abstracts known to be related the gene  Gene-document represented as vectors:

Experiments (cont.)  Keyword query and accession number query  Reelin signaling pathway  GO classification terms and human disease  Direct genes and indirect genes  Hierarchical Clustering

Results

Results (cont.)

 Tried 5, 25, and 50 dimensions. 50 is shown to perform the best.  Tried reducing the numbers of abstracts of Reelin genes. Claimed that AP was not significantly reduced when 50% abstracts were removed.  Claimed that hierarchical clustering agrees with biological relationships.

Discussions  Pros Gene clustering by textual information. Applied LSA to biomedical literature. Indirect linkage can be found through latent concepts.  Cons Requires human annotation to construct gene-documents. Not applicable to new domain. Genes in the experiments are carefully chosen in 3 categories. How does the method perform in general?  Other gene clustering methods?

References  S. Deerwester et al. (1990). Indexing by Latent Semantic Analysis. Journal of the American Society for Information Science, 41-6,  M.A. Gerolami (2004). Latent Semantic Analysis A General Tutorial Introduction.  H. Shatkay (2005). ISMB 05’ Tutorial Proposal.  H. Shatkay & R. Feldman (2004). Mining the Biomedical Literature in the Genomic Era: An Overview. Journal of Computational Biology, 10-6,

The End  Questions?  Thank you!