Using text to enhance the interpretation of large multi-dimensional data sets

When numerical data about a large number of comparable entities is analyzed, information critical to understanding the patterns in the data is frequently available in the textual documentation about those entities. For example, high throughput gene expression assays in biology obtain expression for thousands of genes under dozens of conditions. Known properties about the genes that may partially or completely explain the expression patterns are described in the scientific literature describing the genes. I hypothesize that given a collection of entities, multidimensional numerical measurements on them, and free text documentation about them, an algorithm can search for and identify patterns within the numerical data that can be accounted for by shared properties, described in the documentation, of the entities that follow that pattern. Here I develop and present a method that uses text analysis to help find meaningful gene expression patterns that correlate with the underlying biology described in scientific literature.; First, I establish that computational text mining can discern key biological concepts from text. I demonstrate that supervised machine learning methods can be used to identify the biological function that is being addressed a document. Second, I present and evaluate the neighbor divergence per gene (NDPG) method that assigns a score to a given group of genes indicating the likelihood that the genes share a biological property or function. To do this, it uses only a reference index that connects genes to documents, and a corpus including those documents. Third, I present an approach, optimizing separating projections (OSP), to search for linear projections in gene expression data that separate functionally related groups of genes from the rest of the genes; the objective function in our search is the NDPG score of the positively projected genes. A successful search, therefore, should identify patterns in gene expression data that correlate with meaningful biology. I apply OSP to a published gene expression data set; it discovers many patterns that correspond to functionally related genes. Since the method requires only numerical measurements about entities with textual documentation, I conjecture that this method could be applied to other sorts of data sets.