The network-enhanced connectivity score compares the gene signature of a specific biological process—such as transcriptomics data from patients with a particular disease—to the gene signatures of reference drugs. It leverages interactomics by projecting this data into a network space encompassing protein-protein, protein-metabolite, and protein-microRNA interactions, and propagates signals across the network. The algorithm then calculates a similarity score between each drug and the process of interest, ultimately ranking the drugs based on their similarity. A visual representation of this pipeline is shown in Figure 1.
Figure 1
The algorithm requires two inputs:
- Disease Signature – Currently implemented for transcriptomic data [1], this includes paired transcriptomic data from healthy and diseased tissues across patients.
- Drug Signature – A collection of paired test-control L1000 experiments for small-molecule perturbagens across cell lines [2].
Raw data undergoes preprocessing (PP), including filtering and removing duplicate gene IDs, ensuring that only unique gene identifiers are retained for further analysis.
An LMM is applied separately to drug and disease signatures [3][4]. This model extends traditional linear models by accounting for experimental variability through random effects, specifically RNA plate and cell type. The output is a log fold change representing the difference between the condition and the control data.
A meta-analysis [5] is performed to account for multiple perturbation time points (6 and 24 hours) and produce a consolidated measurement for each drug, per time point and pooled across both.
The resulting fold change data from the LMM is processed using the MITHrIL pipeline [6], which propagates the signal across an extensive interaction network. This step generates a perturbation score for each network node and an accumulated perturbation for each pathway; the connectivity score can be computed on either.
Using the perturbation scores, a connectivity score is computed between the disease and each drug based on the Kolmogorov-Smirnov statistic [4][7]. The drugs are then ranked according to these scores. Drugs with the most negative connectivity scores are most likely to counteract the disease effect.
Step (d) is optional: the connectivity score can also be computed directly on the differential-expression signatures, which is the baseline NetCoS is compared against. See docs/pipeline.md for the full set of signature representations.
[1] M. Prudencio, J. Humphrey, S. Pickles, A.-L. Brown, S. E. Hill, J. M. Kachergus, J. Shi, M. G. Heckman, M. R. Spiegel, C. Cook et al., "Truncated stathmin-2 is a marker of TDP-43 pathology in frontotemporal dementia," The Journal of Clinical Investigation, vol. 130, 2020.
[2] L. C. for Transcriptomics (Broad Institute), L1000 Dataset -small molecule perturbagens- LINCS Trans-Center Project, 2014.
[3] D. Bates, M. Mächler, B. Bolker, and S. Walker, "Fitting Linear Mixed-Effects Models Using lme4," Journal of Statistical Software, vol. 67, pp. 1–48, 2015.
[4] K. K. M. Koudijs, S. Böhringer, and H.-J. Guchelaar, "Validation of transcriptome signature reversion for drug repurposing in oncology," Briefings in Bioinformatics, vol. 24, p. bbac490, 2023.
[5] M. Harrer, P. Cuijpers, T. Furukawa, and D. Ebert, Doing Meta-Analysis with R: A Hands-On Guide, Chapman and Hall/CRC, 2021.
[6] S. Alaimo, R. Giugno, M. Acunzo, D. Veneziano, A. Ferro, and A. Pulvirenti, "Post-transcriptional knowledge in pathway analysis increases the accuracy of phenotypes classification," Oncotarget, vol. 7, p. 54572, 2016.
[7] B. Chen, L. Ma, H. Paik, M. Sirota, W. Wei, M.-S. Chua, S. So, and A. J. Butte, "Reversal of cancer gene expression correlates with drug efficacy and reveals therapeutic targets," Nature Communications, vol. 8, p. 16022, 2017.
[8] A.-L. Brown, O. G. Wilkins, M. J. Keuss, S. E. Hill, M. Zanovello, W. C. Lee, A. Bampton, F. C. Y. Lee, L. Masino, Y. A. Qi et al., "TDP-43 loss and ALS-risk SNPs drive mis-splicing and depletion of UNC13A," Nature, vol. 603, pp. 131–137, 2022.
- Python — pipeline orchestration, connectivity score, validations, plotting. Dependencies in
requirements.txt(joblib,matplotlib,numpy,pandas,pyreadr,requests,scipy); package metadata insetup.py(drug_repurposing, v0.1.0). - R — LMM fitting, LINCS data handling, signature meta-analysis, edgeR for count-based case studies. Under
tsr/(legacy workflows) andnew_tsr/R/(refactored OOP version). - Julia — LMM optimization (
tsr/julia/,modules/drug_signature/julia/). - MITHrIL — external Java tool, run as a subprocess (
step2_disease_run_MITHrIL.py,step6_drug_run_MITHrIL.py).
Experiments are switched by which conf.py is active in src/ (there is no CLI flag); the CS entry point is cs_batch.run_cs_batch_for_conf(), and every run is logged in logs/cs_runs.tsv. Heavy R/Julia jobs can take hours.
docs/pipeline.md— precise stage-by-stage pipeline & architecture reference: the two pipelines (TSR/LMM and BinChen2017/ChEMBL) and the drug-collapse inversion, the signature representations (DEG / MITHrIL / pathways), the numbered steps, the CS methods and sign convention, disease-signature sources, the CS run registry, the config system, the analyses built on the pipeline, the repository layout and how to run it.