Reusability report: On Learning Shared and Context-Specific Regulatory Programs via Multi-Sample Constrained Inference
This repository includes the supplementary materials and code for the paper titled Reusability report: Evaluating the generality and robustness of the CORNETO framework for knowledge-driven multi-sample network inferece. The research evaluates CORNETO as a unified platform for knowledge-driven multi-sample network inference, assessing its reproducibility, robustness, and generality across diverse biological contexts. We benchmark multiple CORNETO-compatible inference paradigms and analyze how performance depends on data sparsity, sample availability, and prior-network completeness, providing practical recommendations for best practices in constrained network modeling. Additionally, we release a set of reproducible workflows and analysis scripts that serve as a baseline for future methodological comparisons and extensions. This framework is highly modular and extensible, enabling seamless integration of new inference tools, datasets, and refinement strategies as CORNETO continues to evolve.
Table of contents:
Understanding biological regulation requires methods that can infer molecular interaction networks from heterogeneous omics data while incorporating existing biological knowledge. Traditional network inference approaches are often applied on a per-sample basis, making it difficult to distinguish shared versus context-specific structure and leading to unstable or non-identifiable solutions. CORNETO (CONstrained optimization for the Recovery of NETworks from Omics) addresses these challenges by formulating knowledge-driven network inference as a multi-sample constrained optimization problem, enabling joint inference across conditions and improving interpretability and robustness. This work provides a comprehensive evaluation of CORNETO as an integrated framework, benchmarking its behavior across different inference paradigms, biological contexts, and input conditions. We systematically analyze how factors such as data sparsity, prior-network completeness, regularization strength, and sample availability affect solution quality, and provide practical guidance for selecting appropriate settings when applying CORNETO to real-world datasets.
This study evaluates CORNETO through four structured components::
- Reproducibility: Reproduction of previously published CORNETO analyses. Multiple independent runs to quantify stochastic variability/Comparison of single-sample vs multi-sample stability/Assessment of run-to-run consistency for stochastic tools (FBA/sFBA, KPNN)/Evaluation of sensitivity to optimization parameters and random seeds
- Reusability: Benchmarking CORNETO as a unified platform across tools and biological contexts: Application to two contrasting regimes: Dynamic system: hematopoietic differentiation/Homeostatic system: adult brain cortex. Unified contrast design applied to all tools. Consistent multi-sample vs single-sample comparisons. Regularization sweeps to identify optimal operating points
- CORNETO inference tools evaluated:iMAT – metabolic network inference/KPNN – knowledge-primed neural network modeling/Steiner Tree – regulatory subnetwork recovery/CARNIVAL – signaling pathway inference
- Characteristics and Limitations: Systematic analysis of CORNETO behavior under challenging input conditions: Impact of data sparsity and dynamic range/Effect of feature-weight skew and signal distribution/Sensitivity to prior-network completeness/Influence of sample availability on stability/Regularization-strength trade-offs/Identification of critical thresholds for reliable inference
- Enhancements: Post-inference strategies to improve robustness and interpretability: Bootstrap Aggregation/Ensemble resampling of conditions/Frequency-based edge consolidation/Reduction of stochastic variability/Improved recall in sFBA settings/Selective Node Reintegration/Targeted reintroduction of informative genes in KPNN/Recovery of performance lost to aggressive pruning/Preservation of model sparsity and interpretability
Following datasets consist only on the formatted transcripts and images files, go to corresponding autor references for original datasets.
| Dataset | Raw data link | Technology | scRNA-seq annotation reference data used |
|---|---|---|---|
| Hematopoiesis | GEO | scRNA-seq | Paper link |
| Brain | CELLxGENE | scRNA-seq | Paper link |
The evaluation was performed using the two datasets presented. The metrics employed were derived from the CORNETO proposed metrics. You can find the code for evaluation and visualization generation in this repository.
The findings demonstrate that CORNETO’s performance and robustness depend strongly on data structure, prior-network completeness, and sample coupling, and that multi-sample regularization and targeted post-inference refinements substantially improve stability and biological fidelity. Detailed analyses and comparisons are provided in the paper.
To run the code and reproduce the results, please follow the official installation instructions provided by the CORNETO package authors: https://corneto.org/dev/install.html
This documentation describes how to install CORNETO, required solvers, and dependencies needed to execute the optimization-based workflows used in this study. All analysis scripts in this repository assume a working CORNETO installation as described in the official guide.
This repository contains fully reproducible scripts for all analyses described in the paper.
Running each analysis component is possible through the tool-specific code provided in this repository:
-
Reproducibility analyses → reproducibility/
-
Reusability benchmarking → reusability/
-
Characteristics and limitations studies → limitations_analysis/
-
Enhancement strategies → enhancements/
Each folder includes standalone scripts and configuration files required to reproduce the corresponding experiments.
Contributions to improve this repository are welcome! If you have suggestions or improvements, please open an issue or submit a pull request.