Network Intelligence

Research pipeline

Methodology

This platform is a presentation layer around a validated reproduction of Agrawal, Zitnik & Leskovec, Large-scale analysis of disease pathways in the human interactome (PSB 2018). It is a research tool for computational disease–protein association analysis, not a clinical diagnostic system.

Project workflow

This strip is a teaching diagram of the product story. The numbered list below is the research pipeline implemented in this reproduction.

  1. 01Disease
  2. 02Protein interactions
  3. 03Disease network
  4. 04Prediction methods
  5. 05Combined ranking
  6. 06Network patterns
  7. 07Candidate proteins

Decorative workflow — not a new scientific result.

  1. 1Disease Associations
  2. 2PPI Network
  3. 3Disease-Specific Network
  4. 4Structural Analysis
  5. 5Prediction Methods
  6. 6Graphlet Signatures
  7. 7Neural Embeddings
  8. 8Logistic Regression
  9. 9Candidate Proteins

Data sources

  • Human PPI network: 21,557 proteins, 342,353 interactions (Menche et al. / BioGRID via SNAP).
  • Disease–protein associations: 519 diseases with ≥10 genes (DisGeNET via SNAP).
  • Disease-level graphlet orbit p-values: 73 orbits per disease (paper supplementary).

Prediction methods

Neighborhood scoring, Random Walk with Restart (α=0.7), DIAMOnD (z-score approximation), spectral embeddings (truncated SVD, 64-d) with logistic regression, and NMF matrix completion. The combined score is the mean of min–max-normalized method scores.

Evaluation

Disease-centric 10-fold cross-validation. Metrics actually computed: Recall@25, Recall@100, MRR. ROC-AUC, PR-AUC, F1, MSE, and R² are not computed in this project and are not displayed.

Matrix Completion / NMF limitation

In the cross-validation evaluation, NMF is fitted on the full protein×disease matrix before fold splitting. Matrix Completion therefore has access to test-fold associations during factorization, which inflates its reported CV metrics relative to the other four methods. This does not affect live predictions, where all known associations are legitimately available.

Known deviations from the paper

  • Neural embeddings use truncated SVD, not node2vec.
  • DIAMOnD uses a z-score approximation rather than exact hypergeometric p-values.
  • Per-protein graphlet signatures use 12 structural proxies, not full 73-orbit ORCA counts.
  • Disease-level 73-orbit p-values from the SNAP dataset are shown as published.

References

  1. Agrawal M, Zitnik M, Leskovec J. Large-scale analysis of disease pathways in the human interactome. PSB 2018.
  2. Menche J, et al. Uncovering disease-disease relationships through the incomplete interactome. Science 2015.
  3. Ghiassian SD, Menche J, Barabási AL. A DIAMOnD module detection algorithm. PLoS Comput Biol 2015.
  4. Piñero J, et al. DisGeNET. Nucleic Acids Research 2017.