Relation between the precision of the disambiguation and score values. Ten score values were chosen randomly. For each of these values, we considered the 50 terms that had scores immediately greater than the value, and evaluated if the term referred to a gene or gene product in the article where the prediction was made. Precision of the prediction was calculated as the number of correct predictions over the total number of predictions (50). The 500 names which were checked represent a random sample of 21,501 names in the catalog. The Figure shows that predictions made with higher scores have a greater probability to be correct. The label shown for each evaluation point indicates the percentage of terms found in the evaluation corpus with a score above the score of this specific point.

Steps involved in constructing the catalog of protein references. Terms are shown enclosed in rectangular boxes. Terms may occur in the context of sentences (when on a horizontal line, left), or in an article (right). Step 1: Articles are split into sentences, and sentences are split into tokens. Tokens roughly correspond to words (see text for details). Tokens with high frequency that are not eliminated by the exclusion lists (see Figure 1) are grouped into n-grams. On the figure, APE1/ref-1 is a n-gram that consists of two tokens: APE1 and ref-1, and can be recognized if the two terms co-occur frequently in sequence in a full length article. When the terms are recognized, each occurrence of a term in sentences of the article is identified. Step 2: Machine learning features are calculated from the context of the term in the article (see text for details) and the support vector machine (SVM) model classifies the context of the term. We obtain the score for each context of a term. In our experimental setup, smaller scores suggest that the context provides little evidence that the term refers to a protein, while larger scores (in absolute values) indicate more support. Step 3: We calculate the combined score (Sc) as the sum of the scores for each occurrence of a given term in a given article. The final catalog consists of a table with one row per term and article. Each row has three columns: PubMedID, term, and Sc.

Filtering process and exclusion lists. Oval boxes on the left of the figure present the exclusions lists used as input to the filtering process. An exclusion list is connected to the step of the filtering process that uses it. Each step filters out terms that match the exclusion list in sequence based on the rules described in the boxes of the second column. Some steps perform matches by considering the entire term (e.g., the first step on the top), other steps use only specific words in a term. These lists have been built and are being maintained manually. The rectangular boxes on the right show the number of terms that have been excluded at each step, when processing the most frequent terms from JBC2000 (numbers in parentheses indicate counts of unique terms).

Heuristics filtering of the protein catalog to produce the protein dictionary. The dictionary was constructed by considering the classification results of a particular term in different articles. Step 1: we filtered out terms that were predicted to be a protein in less than 75% of the articles where a prediction was made. For example, if term A appears in 4 articles and is classified as a protein name in 3 of them, term A is accepted in the dictionary. This process collected 61,312 terms. Step 2: we removed terms with two characters or less. Step 3: to remove ambiguity with protein names that are also common nouns, we filter the dictionary against the Webster's Revised Unabridged Dictionary (G & C. Merriam Co., 1913, edited by Noah Porter, provided by Patrick Cassidy of MICRA, Inc, and retrieved from ). We estimate that this edition contains about 80 common protein names (e.g., amylase). Step 4: we filter the dictionary against species names from the NCBI taxonomy database [30].