Extraction of protein-protein interactions (PPIs) reported in scientific publications is a core topic of biomedical text mining. The ultimate goal is to devise a PPI extraction method that performs well on large amount of unseen text independently from the training corpus. One popular, machine-learning based approach to PPI extraction builds on the convolution kernels, i.e., similarity functions defined on the parse-based representation of sentences and interactions. Kernel functions differ in (1) the underlying sentence representation (bag-of-words, syntax tree parse, dependency graphs), (2) the substructures retrieved from the sentence representation to define interactions, and (3) calculation of the similarity function.
We present a novel kernel method called k-band shortest path spectrum kernel (kBSPS), an extension of the spectrum tree kernel (SpT) . It combines three ideas: First, interactions are represented as vertex-walks as in SpT but adapted to dependency graphs. The kBSPS kernel includes also edge labels into vertex-walks, thus also exploiting the dependency type of a relationship. Second, it uses a novel similarity function on vertex-walks permitting certain mismatches, thus allowing for linguistic variations. The tolerant matching distinguishes three types of nodes: dependency types (D), candidate entities (E), and other surface tokens (L). Mismatches / matches are scored differently depending on the type of nodes. Third, apart from the shortest path between the proteins of the candidate interaction, kBSPS also adds all nodes within distance k from this path to the vertex-walk representation.
We evaluated kBSPS kernel on the 5 standard PPI benchmark corpora (AIMed, BioInfer, HPRD50, IEPA, LLL) using document-level 10-fold cross-validation (CV) and cross-learning (CL; 4-vs-1) evaluation. CV evaluation is somewhat biased, because the training and the test data have very similar corpus characteristics and machine learners tend to learn that, therefore CL evaluation, where the training and test data sets are drawn from different distributions, provides a more unbiased picture. Our results are compared with three state-of-the-art kernel approaches to PPI extraction (see Table 1).
Table 1. Comparison of kBSPS in terms of AUC, F1-measure and classification time with other state-of-the-art kernels using the CV and CL evaluation scenarios
We have shown that kBSPS kernel is on par with state-of-the-art kernels at the more general CL evaluation. Furthermore, its performance is more stable (drops the least from CV) than other methods. Notably, kBSPS is also much faster than any other kernel, making it applicable to very large corpora.
Domonkos Tikk was supported by the Alexander-von-Humboldt Foundation.