Refining Protein Co-Membership in Reactome: A Weighted Scoring Approach Based on Hierarchy and Sequences of Patwhays

Technical Summary: Refining Protein Co-Membership in Reactome

Problem Statement
In network biology, the pathway co-membership approach is commonly used to link proteins belonging to the same pathway by projecting them into a clique (a fully connected subgraph). While effective for mapping pathway knowledge to proteins, standard unweighted co-membership methods are ill-suited for downstream applications relying on graph distances, such as random walks with restart (RWR). These methods fail to distinguish between proteins that are functionally closely related (e.g., within a specific sub-pathway) and those that are merely part of a broad, general pathway. Consequently, they do not accurately reflect the biological distance between proteins, treating all co-membership interactions as equal regardless of pathway specificity or sequential organization.

Methodology
The authors propose a weighted pathway co-membership method that incorporates biological distances between proteins by leveraging the hierarchical organization and sequential relationships of Reactome pathways. The methodology proceeds through the following steps:

1. Data Abstraction: Using Reactome version 96 (BioPAX format), the authors developed an abstraction framework to extract two types of relationships from the BioPAX graph:

   • Hierarchical Composition: Extracted via bp3:pathwayComponent relationships, modeled as abstraction:IsAComponentOf edges.
   • Sequential Relationships: Extracted via bp3:nextStep relationships between pathway steps, modeled as abstraction:NextStepPathway (NSP) edges.

2. Scoring System: The authors assign weights to these abstractions using Information Content (IC) based on protein annotations.

   • MICA Score (Hierarchical): For two pathways connected by a composition edge, the weight is derived from the Information Content of their Most Informative Common Ancestor (MICA). The formula used is $IC_{prot}(P) = -\log(\frac{I_{prot}(P)}{I_{prot}(T)})$, where $I_{prot}(P)$ is the number of proteins annotated by pathway $P$ and $I_{prot}(T)$ is the total number of proteins in the graph. Smaller, more specific pathways yield higher scores.
   • NSP Score (Sequential): For pathways linked by a sequence edge, the weight is calculated based on the IC of the union of their annotated proteins ($P1 \cup P2$).

3. Protein Co-Membership Scoring:

   • For any pair of proteins ($p1, p2$), the authors identify their respective pathway parents in the abstraction graph.
   • Two matrices are computed: one for MICA scores and one for NSP scores (considering ancestors).
   • The final co-membership score for the protein pair is defined as the maximum of the MICA score and the NSP score:
     $$score(p1, p2) = \max(score_{MICA}(P1, P2), score_{NSP}(P1, P2))$$
   • A score of zero indicates the proteins share only the root pathway without specific hierarchical or sequential links.

4. Implementation: The authors pre-computed weighted co-membership cliques for all 29 Reactome top-level pathways and the entire Reactome knowledge base (73 million interactions). These are available on NDEx.

Key Contributions

• Weighted Abstraction Framework: A novel method to abstract Reactome pathways into a graph that captures both compositional hierarchy and sequential order, assigning weights based on protein information content.
• Scoring System: A specific scoring algorithm for protein pairs that integrates hierarchical specificity (via MICA) and sequential context (via NSP), moving beyond simple binary co-membership.
• Resource Generation: The generation and public deposition of weighted co-membership cliques for the 29 Reactome top-level pathways and the entire Reactome knowledge base, facilitating exploration of protein interactions.
• Benchmarking: A comprehensive evaluation using Random Walk with Restart (RWR) on ten cancer hallmark gene sets, comparing the proposed method against state-of-the-art similarity measures (Cosine, Jaccard) and unweighted cliques.

Results
The evaluation focused on gene prioritization tasks using ten cancer hallmark gene sets (e.g., "Tissue invasion and metastasis," "Genome instability").

• Performance: The proposed weighted approach (incorporating both hierarchy and sequence) consistently achieved the highest Area Under the Curve (AUC) across all ten cancer hallmarks in RWR experiments with a low restart probability ($r=0.1$), designed for long-range network exploration. AUC values ranged from 0.78 to 0.90.
• Comparison: The method outperformed state-of-the-art methods using Cosine similarity (AUC 0.74–0.89) and Jaccard similarity (AUC 0.67–0.88). Unweighted cliques performed close to random (AUC 0.55–0.71).
• Component Analysis:
  • Removing sequential information (NSP) and relying only on hierarchy resulted in a slight performance decrease (AUC 0.77–0.89), indicating that while NSP scores constitute a small fraction of total scores (approx. 0.4% of interactions in the full clique), they provide significant informational value.
  • In local neighborhood exploration tasks (high restart probability $r=0.7$), all weighting methods showed equivalent performance, suggesting that for close neighbors, the specific weighting metric matters less than the general topology.

Significance and Claims
The paper claims that the proposed approach outperforms state-of-the-art methods for gene prioritization on cancer hallmarks, specifically in tasks requiring long-range network exploration. The authors argue that their method effectively captures the "biological distance" between proteins by accounting for the specificity of the pathways they share and the sequential order of pathway steps.

The authors emphasize that while the contribution of sequential (NSP) scores to the total weight distribution is numerically small, their inclusion leads to measurable improvements in performance, validating the importance of integrating pathway sequences into co-membership scoring. The work provides a scalable, pre-computed resource (available on NDEx) to facilitate more accurate protein interaction analysis in Reactome-based studies. The authors note limitations regarding the computational cost of re-generating cliques upon Reactome updates and the dependency on the richness of BioPAX descriptions in other pathway databases.