Joint multi-sample network inference steadies omics reconstructions, but a prior cut to 20% leaves zero active reactions
Synopsis
Multi-sample regulatory network inference has long solved each sample separately and stitched shared structure from overlaps, which is unstable under noise. CORNETO instead solves several samples inside one mixed-integer constrained optimization with structured sparsity penalties. On haematopoietic and adult cortical single-cell data, joint inference generally improved robustness and reproducibility, while predictive gains depended on biological context and prior-network completeness.
Interpretation
Solving several samples inside one constrained optimization yields more stable and reproducible reconstructions than inferring each sample separately, and that robustness is the main benefit of multi-sample inference. Knowledge-driven methods were often applied per sample, with shared patterns identified by overlapping independently inferred sample-specific networks, an approach sensitive to instability under noisy or weakly constrained settings. Multi-sample sparse flux balance analysis produced smaller networks than single-condition runs, and on haematopoietic data joint inference recovered more reference-consistent edges at comparable error; CARNIVAL at λ=0.1 retained about 720 shared edges versus roughly 470–500 in single-sample runs.
The method writes network inference as a mixed-integer constrained optimization over network flows, using structured sparsity penalties to recover shared and condition-specific subnetworks together. Flux balance analysis, sparse flux balance analysis, iMAT, CARNIVAL and prize-collecting Steiner trees become special cases of one formulation, with an open-source implementation supporting modular extensions. The evaluation covers iMAT, CARNIVAL, Steiner tree and knowledge-primed neural networks on a haematopoietic differentiation dataset and an adult brain cortex dataset, keeping each method's original evaluation procedures and benchmark definitions.
The benefit is regime-dependent: more complete prior networks and more jointly modelled conditions both make solutions more stable. Progressively removing edges from the Human-GEM reaction graph simulated incomplete priors; at 20% completeness all runs returned zero active reactions, 40–60% recovered limited signal, and 80–100% gave higher precision and recall across a broader regularization range. Randomly subsampling 2, 4, 8 or 16 haematopoietic pseudobulk conditions for joint iMAT optimization raised pairwise Jaccard similarity between active-reaction sets from independent runs and reduced sensitivity to regularization strength.
Perspective
The analysis concerns the reusability, robustness and reproducibility of the CORNETO framework, not the biological correctness of the inferred regulatory mechanisms; the authors state that experimentally validated context-specific regulatory ground truths are unavailable. Evaluation covers a haematopoietic differentiation dataset and an adult brain cortex dataset, representing dynamic and comparatively stable settings, and keeps the framework-appropriate inputs, optimization objectives and evaluation criteria of iMAT, CARNIVAL, Steiner tree and knowledge-primed neural networks rather than ranking them against one ground truth. The authors propose that establishing generality will require systematic evaluation across several independent datasets, tissues, organisms and prior-knowledge networks.
The authors stress that the dynamic and comparatively stable settings are each represented by a single dataset, making them case studies rather than biological replicates, so observations about biological variability should be read as hypotheses about inference regimes. Because validated context-specific regulatory ground truths are lacking, the biological analyses assess consistency with established relations and do not experimentally validate causal mechanisms. CARNIVAL returned no feasible solutions on the cortical data, which the authors attribute to extreme intercell-type homogeneity and class imbalance producing a degenerate optimization space.