A Rich Feature Vector for Protein-Protein Interaction Extraction from Multiple Corpora

Makoto Miwa, Rune Sætre, Yusuke Miyao, Jun'ichi TsujiiEMNLP 2009 · OpenReview

Abstract: Makoto Miwa, Rune Sætre, Yusuke Miyao, Jun’ichi Tsujii. Proceedings of the 2009 Conference on Empirical Methods in Natural Language Processing. 2009.

Introduction

Extracting protein-protein interactions (PPI) from scientific literature is essential for building comprehensive biomedical databases. However, automated Information Extraction (IE) systems face two primary challenges: data scarcity and corpus inconsistency. No single annotated dataset is large enough to train a perfect model, and existing datasets (such as AIMed, BioInfer, HPRD50, IEPA, and LLL) often use conflicting definitions for entities and interactions.

This research presents a high-performance PPI extraction system that overcomes these hurdles by combining a "rich feature vector" with a novel corpus-weighting strategy (SVM-CW), allowing models to learn effectively from multiple heterogeneous datasets simultaneously.

System Architecture

The proposed system processes raw text and pair information through a pipeline that includes parsing, feature extraction, and classification.

Overview of the PPI extraction system architecture.

By utilizing multiple parsers (like Enju and KSDEP), the system captures diverse linguistic perspectives. These are then converted into a unified feature representation for a linear Support Vector Machine (SVM) classifier. This approach allows the system to be significantly faster than previous kernel-based methods while maintaining high accuracy.

The Rich Feature Vector

A core contribution of this work is the engineering of a massive, high-dimensional feature vector. Unlike computationally expensive kernel-based methods, which calculate similarity between pairs of sentences directly, this linear approach represents complex linguistic structures as features.

The vector integrates three distinct layers of information:

  • Bag-of-Words (BOW): Captures the local lexical context surrounding the protein pair.
  • Shortest Path (v-walks and e-walks): Represents the grammatical dependencies connecting the two proteins in a sentence's parse tree.
  • Graph Features: Encodes broader structural information from the entire sentence graph.
Detailed construction and normalization of the rich feature vector.

A critical component of this representation is the normalization strategy. Features are first normalized within their respective groups (e.g., all BOW features) before the entire vector is normalized (L2-norm). This prevents any single feature type from dominating the model's decision-making process, ensuring a balanced contribution from lexical and structural data.

SVM-CW: Learning from Multiple Corpora

To address the inconsistency between datasets, the authors modified the SVM algorithm to include Corpus Weighting (SVM-CW). When training for a specific "target" corpus (e.g., AIMed), the system incorporates data from other "source" corpora (e.g., BioInfer) but assigns them lower weights. This "aggressive" transfer learning approach allows the model to learn general interaction patterns from the source data while prioritizing the specific nuances and definitions of the target corpus.

Key Results and Performance

The system achieved state-of-the-art results across all five standard PPI corpora. Notably, on the AIMed dataset, it reached an AUC of 89.3%, a significant improvement over previous benchmarks.

Beyond accuracy, the linear nature of the feature vector provides a substantial efficiency gain. The system is approximately 2x faster to train and test than previous kernel-based methods, making it highly suitable for large-scale literature mining across millions of PubMed abstracts. The study also found that certain corpora act as better "teachers" for others; for instance, IEPA significantly improved performance on the AIMed dataset.

Conclusion and Future Directions

The success of this approach demonstrates that well-engineered linear features can outperform complex graph kernels when combined with effective normalization and multi-corpus learning strategies. Future work will focus on improving coreference resolution (handling pronouns like "it" or "they") and further enriching the feature vector to capture deeper semantic meanings. Additionally, automating the tuning of corpus weights could make the system more robust for diverse domain adaptation tasks.

Summary updated

BibTeX
@inproceedings{miwa2009rich,
  author = {Makoto Miwa and Rune Sætre and Yusuke Miyao and Jun'ichi Tsujii},
  title = {{A Rich Feature Vector for Protein-Protein Interaction Extraction from Multiple Corpora}},
  booktitle = {Conference on Empirical Methods in Natural Language Processing},
  year = {2009},
  url = {https://openreview.net/forum?id=qE349r2XvKp},
}
Figure