Multi-level learning: improving the prediction of protein, domain and residue interactions by allowing information flow between levels

Kevin Y. Yip, Philip M. Kim, Drew McDermott, Mark GersteinBMC Bioinformatics 2009 · DOI · OpenReview

Introduction

Predicting protein-protein interactions (PPIs) is a fundamental task in systems biology, essential for understanding cellular processes and disease mechanisms. However, biological interactions are hierarchical, occurring at multiple scales: the whole protein level, the domain level, and the individual residue level. Traditionally, machine learning models have treated these scales as independent problems, failing to leverage the inherent dependencies between them.

This research introduces a multi-level machine learning framework that predicts interactions across all three scales simultaneously. By enabling bidirectional "information flow," where high-confidence predictions at one level serve as training data for others, the authors achieve significant improvements in prediction accuracy compared to treating each level as an independent problem.

Bridging the Granularity Gap

Interactions at different levels provide complementary biological insights. Protein-level data often comes from functional genomics (e.g., co-localization), domain-level data from evolutionary signatures, and residue-level data from physico-chemical properties like charge and solvent accessibility.

The primary challenge lies in the differing "granularities" of these features. It is difficult to directly apply a residue-level feature to a protein-level prediction. The proposed framework bridges these gaps by treating the hierarchy as a unified system rather than isolated tasks.

Methodology: Bidirectional Flow and Soft Coupling

The core of the framework rests on three architectural innovations:

1. Bidirectional Information Flow

Unlike previous "unidirectional" models that only passed information from the top down (protein to domain), this framework allows information to flow both ways. For instance, a high-confidence residue interaction can "vote up" a protein interaction, while a known protein interaction can act as a prior to narrow the search for residue-level contacts.

Conceptual comparison of unidirectional versus bidirectional information flow between hierarchical levels.

2. Soft Coupling via Training Set Augmentation

To prevent the propagation of errors from noisy data, the authors implemented "soft coupling." Using Support Vector Regression (SVR) to generate confidence scores, the model only incorporates the most certain predictions from one level as "auxiliary" training examples for the next. This ensures that only strong signals influence the learning process across levels.

Schematic of the soft coupling mechanism, showing how predictions from one level are used to augment the training set of another.

3. Local Modeling

To address data sparsity in large interaction networks, the authors utilized local models. Instead of a single global model, a specific model is built for each protein, focusing on its unique neighborhood in the interaction network. This allows the algorithm to learn the specific "interaction style" of individual proteins.

Results and Discussion

The multi-level approach was tested on the Saccharomyces cerevisiae (yeast) interaction network, utilizing gold standards from MIPS, DIP, and iPfam. The results demonstrated that bidirectional flow significantly outperforms independent models, particularly at finer granularities:

  • Residue Level: AUC increased from 0.5675 (independent) to 0.7361 (bidirectional).
  • Domain Level: AUC improved from 0.5214 to 0.7015.
  • Protein Level: AUC rose from 0.7153 to 0.7257.

A notable example is the hexokinase isoenzyme 2 (HXK2). The model identified interacting residues with high confidence (0.89), which subsequently boosted the confidence of the corresponding domain interaction from a mediocre 0.27 to a robust 0.72.

While the framework shows great promise, challenges remain regarding data sparsity at the residue level and the computational intensity of building local models for larger proteomes. However, the success of this bidirectional, multi-level approach suggests it could be generalized to other hierarchical problems in bioinformatics and beyond.

Summary updated

Abstract

Background Proteins interact through specific binding interfaces that contain many residues in domains. Protein interactions thus occur on three different levels of a concept hierarchy: whole-proteins, domains, and residues. Each level offers a distinct and complementary set of features for computationally predicting interactions, including functional genomic features of whole proteins, evolutionary features of domain families and physical-chemical features of individual residues. The predictions at each level could benefit from using the features at all three levels. However, it is not trivial as the features are provided at different granularity. Results To link up the predictions at the three levels, we propose a multi-level machine-learning framework that allows for explicit information flow between the levels. We demonstrate, using representative yeast interaction networks, that our algorithm is able to utilize complementary feature sets to make more accurate predictions at the three levels than when the three problems are approached independently. To facilitate application of our multi-level learning framework, we discuss three key aspects of multi-level learning and the corresponding design choices that we have made in the implementation of a concrete learning algorithm. 1) Architecture of information flow: we show the greater flexibility of bidirectional flow over independent levels and unidirectional flow; 2) Coupling mechanism of the different levels: We show how this can be accomplished via augmenting the training sets at each level, and discuss the prevention of error propagation between different levels by means of soft coupling; 3) Sparseness of data: We show that the multi-level framework compounds data sparsity issues, and discuss how this can be dealt with by building local models in information-rich parts of the data. Our proof-of-concept learning algorithm demonstrates the advantage of combining levels, and opens up opportunities for further research. Availability The software and a readme file can be downloaded at http://networks.gersteinlab.org/mll . The programs are written in Java, and can be run on any platform with Java 1.4 or higher and Apache Ant 1.7.0 or higher installed. The software can be used without a license.
BibTeX
@article{yip2009multi,
  author = {Kevin Y. Yip and Philip M. Kim and Drew McDermott and Mark Gerstein},
  title = {{Multi-level learning: improving the prediction of protein, domain and residue interactions by allowing information flow between levels}},
  journal = {BMC Bioinformatics},
  year = {2009},
  doi = {10.1186/1471-2105-10-241},
}
Figure