TL;DR
This paper introduces a probabilistic framework to refine existing biological pathway hypotheses by integrating diverse, "messy" data sources—like protein interactions, gene expression, and protein domains. By using a "compartment-specific" strategy, the model only looks at data where it is biologically relevant (e.g., ignoring DNA-binding data in the cell membrane), leading to more accurate models of how cells process signals.
What Problem Does This Paper Address?
In systems biology, we face a "Goldilocks" problem. On one hand, experimental biologists have specific, well-studied hypotheses about how a pathway works. On the other hand, high-throughput "omics" technologies generate massive, unbiased datasets. The challenge is that these two worlds don't talk to each other well.
Standard "de-novo" reconstruction methods try to build networks from scratch, often ignoring decades of expert knowledge. Conversely, traditional hypothesis testing is often too rigid to incorporate the noise and variety of modern genomic data. Furthermore, cellular pathways aren't monolithic; a protein's behavior and the relevance of a data type (like protein-protein interaction vs. transcription factor binding) change depending on whether the action is happening in the cell membrane, the cytoplasm, or the nucleus.
Methods
The authors propose a compartment-specific probabilistic graphical model to refine pathway hypotheses. Key technical highlights include:
- Local-move Gibbs Sampler: Instead of building a network from scratch, the algorithm starts with a "hypothesis pathway" and explores the space of possible refinements by adding or removing edges one at a time based on posterior probabilities.
- The Compartment Map ($C$): This is a crucial innovation. It’s a binary matrix that tells the model which data types are informative for which cellular locations. For example, it tells the model to weigh TF-DNA binding data heavily in the nucleus but ignore it when evaluating interactions in the cell membrane.
- Relation Regression: A novel methodology for integrating "node attributes" (like protein domains or phosphorylation states). It assumes that a protein’s properties are likely related to the properties of its neighbors in the network (e.g., proteins with similar domains are more likely to interact).
- Likelihood Integration: The model combines five distinct data types (PPI, TF-binding, knockouts, co-expression, and node attributes) into a single joint posterior distribution.
Datasets
The researchers validated their approach using the well-studied pheromone response MAPK pathway in the yeast S. cerevisiae. They utilized a "compendium" of heterogeneous sources:
- Physical Binding: Literature-curated data from BioGRID and Reguly et al.
- TF-DNA Binding: P-values from Harbison et al.
- Gene Expression: 20 datasets from 15 publications (including Roberts et al. and Gasch et al.) processed via the SPELL algorithm.
- Perturbation (Knock-out): Log-fold change data from Hughes et al. and Roberts et al.
- Node Attributes: Protein domains from PFAM and SMART; differential phosphorylation data from Gruhler et al.
Key Insights and Ideas
- Data is Context-Dependent: The "aha moment" here is that treating all data as globally relevant adds noise. By restricting data types to specific compartments (Membrane, Cytoplasm, Nucleus), the model filters out irrelevant signals.
- Nodes are as Important as Edges: Most network models focus only on edges (interactions). By including "node properties" (like whether a protein has a specific binding domain), the authors show that we can significantly improve the likelihood of identifying real biological interactions.
- Refinement over Reconstruction: The goal isn't to find the "correct" network, but to provide a list of edge probabilities that help experimentalists decide which parts of their current model are well-supported and which need revision.
Key Results
- Compartment Performance: The model showed high precision and recall in the nucleus and cytoplasm. Performance was lower in the membrane, primarily because the only informative data available there (PPI) was sparse.
- Node Data Impact: In the cytoplasm, node attribute data (protein domains) contributed the most to the differential likelihood of an edge, proving that protein features are a vastly underutilized resource in pathway modeling.
- Cross-talk Discovery: The model successfully identified potential "cross-talk" (interactions) between the pheromone pathway and other MAPK pathways (like starvation and osmolarity responses), suggesting the method can find novel components of known systems.
- Validation: Using a "leave-one-out" strategy, the model demonstrated strong predictive power, correctly assigning high posterior probabilities to known edges even when they were removed from the initial training hypothesis.
Limitations and Future Work
- Compartment Granularity: The current model uses only three broad compartments. The authors suggest that using the Gene Ontology (GO) to divide the cell into dozens of finer components could provide even greater control.
- Overfitting in Sparse Data: The "relation regression" for protein domains can suffer from overfitting because some domains are very rare. While they used Cauchy priors to mitigate this, more data from broader domain categories might be needed.
- Directionality: While the model handles undirected networks, the authors noted that imputing directionality (signal flowing from membrane to nucleus) is a natural next step for more complex signaling models.
Summary updated