Do Adipose "Hub Genes" for Insulin Resistance Survive Cell-Composition Correction?

Insulin resistance means the body's tissues respond weakly to insulin, and it is the main driver of type 2 diabetes. Studies searching for its molecular drivers in fat tissue usually work the same way. They measure gene activity in a ground-up biopsy, build a network of genes that rise and fall together, and call the most connected genes "hubs," meaning candidate master regulators. A ground-up biopsy, however, averages over many cell types: fat cells, their precursor cells, immune cells, and blood-vessel cells. When the mix of cells changes, every marker gene of the expanding cell type moves at once. Those markers then look tightly co-regulated and highly connected even when nothing inside any individual cell has changed. Two practical problems make this worse. The best measures of insulin resistance need insulin assays or infusion tests, so most public gene-expression datasets have no usable insulin-resistance label. And the cohorts that do have labels are small, so hub lists rarely replicate.

This note proposes a replication audit together with a repair. The audit asks how many published-style adipose hub genes for insulin resistance are really proxies for cell-type proportions. The repair estimates those proportions computationally from the bulk data, which is called deconvolution, and builds the gene network only from co-variation that remains after the proportions are accounted for. A further test asks whether a cheap blood-panel score, the triglyceride-glucose (TyG) index, can stand in for insulin-based labels when ranking genes. If it can, many more cohorts with only routine labs could be pooled. The deliverable is evidence about which gene-ranking procedure produces candidates that hold up in independent cohorts, and at what sample size they start to hold up.

Core Claim

We hypothesize three things. First, in subcutaneous fat, a large share of the top hubs found by standard co-expression or protein-interaction centrality are marker genes of cell types whose proportions track insulin resistance. Second, conditioning the network on deconvolved proportions removes most of these markers and yields hubs that replicate better across independent cohorts. Third, gene rankings anchored on TyG or the metabolic score for insulin resistance (METS-IR) largely agree with rankings anchored on an insulin-based index. If all three hold, small-cohort discovery can be enlarged with routine-lab cohorts without degrading the molecular signal. If the first fails, the bulk-hub literature in this tissue is more trustworthy than feared, and that is also worth knowing.

Positioning Against Prior Work

The work that motivates the composition concern is The causal effect of obesity on prediabetes and insulin resistance reveals the important role of adipose tissue in insulin resistance. That study built single-nucleus reference profiles from human fat and applied MuSiC, a deconvolution method that accounts for variation across reference donors, to bulk data from METSIM, GTEx, MOSS and FTC. It found that cell-type proportions are independently associated with the Matsuda index, an insulin-sensitivity measure that requires timed insulin assays. Together with body fat and mitochondrial gene expression, these variables explained 44.39% of its variance. The association it reports is exactly what would contaminate bulk hub rankings. The study itself built no networks and used no surrogate scores. Age-dependent genes in adipose stem and precursor cells… applied the related Bisque method to the same cohorts and showed that precursor-cell proportions shift with age and obesity, which gives a second cell-type axis capable of producing false hubs. We reuse both deconvolution methods and the single-nucleus references.

Two hub-gene studies show the standard pipeline we audit. Shared senescence-associated gene networks in PCOS and T2DM ranked genes by protein-interaction centrality in bulk GEO data. It reported discovery AUC up to 1.0 and external AUC above 0.8, without any composition correction, and validated in blood rather than target tissue. Adipose-Derived Mesenchymal Stem Cells Accelerate Diabetic Foot Ulcer Healing… combined four feature selectors with Firth regression and reported a training-to-external drop from above 0.99 to 0.944. That drop is the generalization gap our sample-size analysis quantifies. The Parkinson's hub study Identification of of hub genes of Parkinson's disease through bioinformatics analysis explicitly flags small cohorts as the limit on its 0.90 AUC signature.

On the phenotype side, the surrogate literature establishes TyG and METS-IR as clinically useful. For example, The Predictive Value of Insulin Resistance Surrogates for Diabetic Kidney Disease… reports a TyG AUC of 0.741 and 0.776 for kidney disease, in males and females respectively. Insulin-free surrogates are also widely used, as in Association Analysis of… METS-IR and Gestational Diabetes Mellitus. No study in the file asks whether these surrogates recover the same tissue-level molecular signal as insulin-based indices. Insulin resistance prediction from wearables and routine blood biomarkers shows that routine labs carry real information about insulin resistance (AUROC 0.80 for classifying it, R² of 0.50 for predicting HOMA-IR), but it uses no molecular data. That paper constrains our expectations: surrogates are informative but imperfect, so agreement in gene rankings will be partial, and measuring how partial is part of the result.

Research Program

The core construction is a composition-conditioned co-expression network. Let $y_g$ be the log expression of gene $g$ across samples, and let $\hat{w}$ be the deconvolved cell-type proportions from MuSiC or Bisque using the METSIM single-nucleus fat reference. The method works in three steps. First, regress each gene on $\hat{w}$, transformed so the proportions are no longer forced to sum to one, and on standard technical covariates. Second, compute the gene-gene correlations from the residuals. Third, rank genes by their connectivity in the resulting network and by their association with the insulin-resistance label. A second variant fits cell-type-specific expression terms, gene by proportion by phenotype, for the two or three most abundant cell types, so that a hub can be attributed to a cell type. Propagating deconvolution error matters because $\hat{w}$ is noisy, particularly for related immune subsets. We therefore repeat the procedure over bootstrap draws of the reference donors and report hub stability across draws.

Phenotype labels are computed in each cohort from whatever is available. These include the Matsuda index or HOMA-IR, which need insulin measurements, and TyG and METS-IR, which need only triglycerides, glucose, HDL cholesterol and BMI. The question of whether an index is available at all is treated as the assessment barrier it is.

A semi-synthetic benchmark gives ground truth. We build pseudo-bulk samples from the single-nucleus data, impose phenotype-linked shifts in cell proportions matched to those observed in METSIM, and plant known within-cell-type co-regulated modules with designated hub genes. Any method's top-k list can then be scored for planted hubs recovered versus marker genes falsely promoted.

Baselines

The naive reference is ranking genes by plain differential expression against the insulin-resistance label in bulk data, with no network and no composition adjustment. This is what a hub analysis must at least beat to justify its complexity.

The strongest existing competitor for hub discovery is the standard bulk pipeline from the cited studies. That pipeline either selects intramodular hubs by module membership in an unadjusted co-expression network, or selects differentially expressed genes, maps them onto a STRING protein-interaction network, and ranks them by degree and betweenness, following Kong et al.. We also run the multi-selector ensemble with Firth regression of Cao et al.. It is the most overfitting-aware method in the file, so it is the hardest bar for replication. All methods receive the same discovery samples, the same gene filter and covariates, and the same tuning budget, with the same number of hyperparameter settings chosen by inner cross-validation. All are scored on identical held-out cohorts.

For predicting the insulin-resistance label itself, the strongest reported number is the 44.39% of Matsuda variance explained by cell proportions, body fat and mitochondrial expression in the obesity–IR causal study. We refit that model on our splits as the reference that hub genes must add to. For cross-cohort replication of hub lists in fat tissue, no reported result was found on Lacuna, so that baseline is established here.

The key ablation is the identical network pipeline with the composition-conditioning step removed. Any gain in replication can then be attributed to conditioning rather than to other choices. Beating the field means two things. On the semi-synthetic data, recovery of planted hubs must be higher at matched false-marker rate. On real cohorts, the top-50 hubs must replicate significantly more often in held-out cohorts than the unadjusted pipeline and the ensemble, with a confidence interval over cohort pairs that excludes zero.

Why not just…?

Why not just regress out cell proportions and run differential expression, skipping the network? This is the most serious threat. Adjusted differential expression may replicate just as well, in which case centrality adds nothing beyond the adjustment. We treat it as a full arm of the study rather than a straw man. If composition-adjusted network hubs do not replicate better than composition-adjusted differential-expression rankings, the paper's conclusion becomes "adjust, and drop the network step." That is still a useful correction to the hub-gene literature. The network claim survives only if planted hubs that do not change in average level, the bottleneck genes highlighted in the network literature, are recovered by centrality and missed by adjusted differential expression on the semi-synthetic benchmark.

Why not just use the proportion-based model from the obesity–IR study, since it already explains 44% of insulin-sensitivity variance? That model predicts the label but names no regulators inside cells. Our goal is candidate molecular targets, not a better predictor. We will nonetheless report whether adding adjusted hub genes raises the variance explained above the refitted 44.39% baseline. If it does not, we will state plainly that the hubs carry no information beyond composition for prediction.

Why not just use TyG clinically and forget the transcriptome? For screening, that is the right answer. The surrogate studies and the wearable study show routine labs suffice, and we make no claim to improve screening. The surrogate question here is narrower: whether TyG can substitute for insulin-based labels when ranking genes, which decides whether routine-lab cohorts can enlarge molecular discovery.

Proposed Experiments and Decisive Checks

The first experiment is the semi-synthetic benchmark described above. We vary how strongly composition shifts with the phenotype and the sample size from 30 to 240. All methods are scored on planted-hub recall and on the fraction of top-k lists made up of single-nucleus marker genes.

The second experiment is a real-data replication audit. We discover hubs in one of METSIM, FTC or MOSS and test them in the others. A replicated hub is one whose association with the insulin-resistance label has the same sign and a false-discovery rate below 0.05 after composition adjustment in the held-out cohort. GTEx fat, which has limited metabolic phenotyping, is used only to test whether network centrality is stable across cohorts. We report the marker-leakage fraction of each method's hub list. The audit is falsified if unadjusted hubs contain no more cell-type markers than random gene sets.

The third experiment is surrogate concordance within METSIM. We correlate gene rankings anchored on Matsuda, HOMA-IR, TyG and METS-IR. We then test whether TyG-anchored discovery replicates against Matsuda-anchored held-out labels nearly as well as Matsuda-anchored discovery does.

The fourth experiment is a learning curve. By subsampling discovery cohorts, we estimate the sample size at which top-50 replication plateaus, with and without conditioning. This tests whether conditioning lowers the sample size needed for stable hubs, and it puts a number on the small-cohort warnings from the cited studies.

As an external stress test, we apply the adjusted and unadjusted pipelines to the GEO datasets used by Kong et al. and Cao et al. Since those datasets lack surrogate labels and target fat tissue, we report only whether their published hubs are cell-type markers in our reference.