iDrug, re-implemented from scratch — benchmarked on the authors' real data

Cross-network weight β

Unobserved-cell weight w

α × β grid — AUPR

Degrading the bridge structure

The model — iDrug

iDrug predicts drug–disease and drug–target links by factorizing the two association matrices — drug–disease A₁ ≈ U₁V₁ᵀ and drug–target A₂ ≈ U₂V₂ᵀ — with a cross-domain consistency term coupling the drug factors over the shared anchor drugs (paper Eq. 1–3, multiplicative updates, up to 200 iterations). The coupling lets each domain borrow signal from the other — the effect the cold-start experiments stress-test. My idrug() uses one shared rank (90) in both domains, and the cross-domain weight w rather than w² — documented differences from the paper's code. idrug-official is a line-by-line port of the authors' MATLAB solver iDrug.m, run at their documented main(70, 70, 0.3, 0.01, 0.01, 0.01, 1, '2', 20) settings in Experiment 4 (my P@20 = 0.280 vs their printed 0.23).

Data — the authors' released matrices

The data is the authors' own release (clone github.com/Case-esaC/iDrug; ~290 MB, not shipped in this repo): CTD + DrugBank — drug–disease 1,321 × 3,966 with 111,481 links, drug–target 946 × 3,610 with 10,234 links, 469 shared anchor drugs — and the Gottlieb gold standard (PREDICT-style) — 593 × 313 with 1,933 links, 593 × 1,011 with 3,427 links, all 593 drugs shared. No synthetic data feeds any number or figure on this page.

Pair-prediction protocol

5-fold CV per domain: each fold hides 1/5 of the known links during training and scores them against an equal number of uniformly sampled unknown pairs. Metrics: AUROC and AUPR, computed per fold and reported as mean ± s.d. over the five folds. Settings follow the paper's pair-prediction values: ranks 90/70, α = β = γ = 0.001, w = 0.3, up to 200 iterations, seed 0, one run per domain.

Cold-start protocol

The scenario the paper cares about most: hold out every link of a fold of drugs (new drug) — or of diseases/targets (new column) — so the row is empty in the task domain, while the drug still exists in the other domain and the cross-network bridge can act. Metric: precision@5/10/15, averaged over all held-out rows of the five folds.

Robustness and sweeps

Stress tests on the bridge between domains: withhold a fraction of anchor drugs and/or of other-domain links and re-measure pair-prediction AUPR; sweep the α × β regularization grid and the cross-domain weight w; and run the paper's hypergeometric enrichment test. The atlas's section C plots these real-data sweeps.

Baselines

Five published baselines, re-implemented from their papers with default hyperparameters and no tuning: GRMF (graph-regularized matrix factorization), RLS-Kron (Kronecker-product regularized least squares), a simplified bi-random walk standing in for MBiRW, similarity-kNN, and a popularity (degree) baseline. TL-HGBI is not implemented. Progression and settings are described in each experiment section; treat every baseline number as a floor of unknown height.

iDrug: Integration of drug repositioning and drug-target prediction via cross-network embedding

Huiyuan Chen¹, Feixiong Cheng^2,3,4, Jing Li^1,4,*

PLOS Computational Biology 16(7): e1008040 · published July 15, 2020 · received November 22, 2019 · accepted June 10, 2020 · open access (CC BY 4.0) · editor: Avner Schlessinger (Icahn School of Medicine at Mount Sinai)

Authors' code, MATLAB solver and datasets: github.com/Case-esaC/iDrug ↗

Plain text — Chen H, Cheng F, Li J (2020) iDrug: Integration of drug repositioning and drug-target prediction via cross-network embedding. PLoS Comput Biol 16(7): e1008040. https://doi.org/10.1371/journal.pcbi.1008040

Cite it — BibTeX

@article{chen2020idrug,
  author  = {Chen, Huiyuan and Cheng, Feixiong
             and Li, Jing},
  title   = {iDrug: Integration of drug repositioning
             and drug-target prediction via
             cross-network embedding},
  journal = {PLOS Computational Biology},
  year    = {2020},
  volume  = {16},
  number  = {7},
  pages   = {e1008040},
  doi     = {10.1371/journal.pcbi.1008040}
}

This page is unaffiliated with the authors and with PLOS. The baseline methods (PREDICT, MBiRW, GRMF, RLS-Kron) are re-implementations from their own papers — details in the repo README.