Consensus pattern alignment to find protein-protein interactions in text

“Don’t I know you from somewhere?” – comparing new to known texts plays a key role in the system we propose for searching protein–protein interactions (PPIs). Our system builds on an inexact pattern matching strategy, where patterns (linguistic frames) reflect the compositional structure of known occurrences of PPIs in text. To describe this structure, part-of-speech tags (verbs etc.) and entity classes (proteins), words, and word stems are used. Consider the sentences “Sky1p phosphorylates Npl3p” and “Akt phosphorylates beta-catenin”. Both have a structure in common that connects two proteins with a single verb. From comparable systems proposed before [1, 2], it became clear that collecting a suitable set of patterns is of major importance, and this step forms the main component of our system. From the IntAct database [5], we extract all pairs of proteins known to interact. We scan PubMed for textual evidences for each such interaction, and retain all single sentences that describe them. Using pairwise sentence alignment as a similarity scoring function, we perform a clustering on the resulting set of sentences. Within each cluster, multiple sentence alignment (MSA) identifies commonalities and variable positions across all sentences, expressed in a consensus pattern. Figure 1 shows an example MSA with four sentences that define one consensus pattern. We can now align such consensus patterns against arbitrary text to extract new PPIs. Our system yields a maximum recall of 69% –which was the best reported among all participating systems–, a maximum precision of 45% and maximum F1-measure of 41% on the BioCreative test set. Our method works completely independent from the training corpus, which we did not use at any stage. Thus, we intrinsically exclude any risk of overfitting, and believe that our approach should work equally well for related extraction problems, such as finding protein–disease associations.