From Postprandial Muscle Biopsy to Druggable Node: a translational re-read of GSE231509

In 2023, Makhnovskii and colleagues published a beautifully designed paired study (GSE231509, PMID 37545425): vastus lateralis biopsies before and 1 hour after a mixed-meal tolerance test (MMTT) in 7 healthy, 7 obese, and 7 obese + T2D subjects. The headline finding was a receptor-tyrosine-kinase (RTK) module — ERBB3, MET, KIT — sharply suppressed after a meal in healthy muscle, and progressively unresponsive in obese and diabetic muscle. They also flagged a FOX/POU/Hippo transcription-factor programme.

Unusually clean paired design, but small N. A translational team looking at this paper would ask three questions: does the biology hold up under a modern, fully-traced reanalysis? What is the dataset not telling us that a richer pipeline could? And if we take the result seriously, what are the actionable, druggable nodes — and what is the realistic translational risk?

We rebuilt the entire analysis from raw counts using a paired DESeq2 + limma-voom pipeline, then layered on TF-activity inference (decoupleR / CollecTRI / DoRothEA), multi-collection GSEA, bulk-adapted ligand–receptor inference (LIANA, 2,100 LR pairs), Spearman clinical correlation against six metabolic traits, ensemble classifier-based stability selection (ElasticNet + RandomForest, nested CV), and a composite druggability score against DGIdb / ChEMBL / Open Targets.

The summary upfront: the original biology replicates, and the dataset has substantially more to give than was previously extracted — including a novel anabolic/inflammatory dichotomy in T2D, a discovery-grade biomarker shortlist, and a ranked list of druggable nodes led by MET (composite 0.792), ERBB3 (0.766), and KIT (0.659), all backed by approved Tier-1 drugs.

PC1 (15.9% of variance) cleanly separates Obese_T2D from Healthy/Obese, and intra-subject correlation (ρ = 0.684) confirms the paired design is real. The progressive blunting of the meal response is the first thing to fall out of the data:

Multi-collection GSEA (Hallmark, KEGG, Reactome) reveals that the disease gradient is not just signalling-blunting — it is a coordinated metabolic re-wiring that amplifies with disease severity:

• OXPHOS amplification scales with disease: HALLMARK_OXIDATIVE_PHOSPHORYLATION NES Healthy ~+1.3 → Obese +2.12 (FDR 2.2×10⁻⁶) → T2D +2.57 (FDR 1.4×10⁻¹⁹). Counter-intuitive: T2D muscle is mounting the strongest mitochondrial transcriptional response post-meal.
• Obese-specific PPAR / fatty-acid amplification (NES +2.26, FDR 2.7×10⁻⁴) with blunted cholesterol homeostasis (NES −1.86, FDR 1.8×10⁻³) — a transcriptomic fingerprint of metabolic inflexibility.
• T2D-unique anabolic reprogramming: HALLMARK_MYC_TARGETS_V1 NES +2.14 (FDR 4.6×10⁻⁹), REACTOME_TRANSLATION NES +2.10 (FDR 9.2×10⁻¹²), KEGG_AMINOACYL_TRNA_BIOSYNTHESIS NES +1.90.
• T2D-unique BCAA catabolism induction (KEGG_VAL_LEU_ILE_DEGRADATION NES +2.20, FDR 1.3×10⁻⁵) — consistent with the well-established BCAA-disposal defect in T2D.
• Blunted inflammatory dynamic range in T2D: IL6_JAK_STAT3 NES −2.05, INFLAMMATORY_RESPONSE NES −1.84, TNFA_VIA_NFKB NES −2.17 — chronic baseline inflammation likely exhausts the acute response window.

This is the single most important addition to the original paper for a biopharma audience. T2D muscle is not "less responsive" — it is responsive in a different mode: anabolic translation, BCAA breakdown, and OXPHOS amplification with simultaneously suppressed acute cytokine induction. Each of those axes has its own therapeutic correlates.

decoupleR-inferred TF activity (CollecTRI + DoRothEA, ΔVST per subject) extends the published FOX/POU programme. FOXA1 is the only TF that survives BH correction in the healthy-specific score, but MECP2 emerges as the strongest novel group-discriminator (KW H = 12.3) and EGR1 is the only TF hub in the analysis with a direct Open Targets T2D genetic-association score (0.163) — an important detail when prioritising hubs by genetic support.

A bulk-adapted ligand–receptor inference across 2,100 LR pairs from 5 LIANA-curated databases highlights phenotype-differential axes that the original paper's gene-level analysis could not see:

We ran two independent streams: (a) Spearman correlation of every gene Δ, TF activity, and ssGSEA pathway against six metabolic axes (HOMA-IR, HbA1c, fasting insulin/glucose, C-peptide, BMI), and (b) ensemble classifier-based stability selection (ElasticNet + RandomForest, 20×7-fold nested CV, 200 bootstrap resamples).

Classifier performance:

We computed a composite druggability score per candidate gene with explicit, weighted components, integrating DGIdb, ChEMBL, and Open Targets:

Three things sit in tension when a biopharma team looks at this dataset, and we want to put them on the table honestly.

1. The biology is convergent and reproducible. The RTK module replicates at gene, pathway, and (for MET) ligand–receptor levels. The FOX programme replicates. The progressive blunting of the meal response is a robust, monotonic phenotype. The novel anabolic/inflammatory dichotomy in T2D (induced MYC/translation/BCAA catabolism, blunted IL-6/JAK-STAT3 and TNFα/NF-κB) is a substantive extension of the original work and a plausible mechanistic model for why T2D muscle is metabolically inflexible and immunologically blunted at the same time.

2. The druggability layer turns the original observation into a prioritised, actionable list. MET, ERBB3, KIT — three approved-drug Tier-1 nodes — fall out at the top. KDR, NOTCH1 (nirogacestat), NOS3 (sapropterin), IGF1R, INSR follow with credible repurposing or known-target rationales. This is a concrete starting point for a target-prioritisation discussion that simply did not exist in the source paper.

3. The translational caveats are real and need to be flagged at the top of any internal memo.

• N = 21. No genes survive FDR<0.10 in the formal interaction model; no LR pairs survive FDR<0.20; no TFs survive FDR<0.05 for group discrimination. Everything here is discovery-grade Phase 1. The high classifier AUCs (0.985–0.993) are real, but the upper CI is constrained by sample size.
• Causal direction is unresolved. The healthy postprandial response is suppression of MET/ERBB3/KIT. In disease, that suppression is lost. Does therapy require inhibition (matching the healthy state by drug-induced suppression), agonism (rescuing the dynamic response), or upstream pathway rescue? Transcriptomic data alone cannot answer this. All Tier-1 drugs in the ranking are oncology kinase inhibitors developed for hyperactivation contexts.
• Genetic support is weak for the leading nodes. ERBB3, MET, and KIT are not GWAS-validated T2D risk loci. Of the entire druggable list, only INSR, IGF1R, NOS3, and EGR1 carry direct Open Targets T2D evidence.
• Bulk RNA-seq misses the cell-type story. NRG1/HGF non-replication is the canary: ligand biology in muscle is largely non-myofiber, and any next step should pair with single-nucleus RNA-seq.
• One time point (1h post-meal). Cannot distinguish a delayed response from an absent one.

What we would do if this were an internal target-ID program:

• External validation cohort. Independent muscle-biopsy MMTT cohort with N ≥ 30/group; pre-registered MET/ERBB3/KIT, MAPRE2, SGK1, BMPR1A, IRF8/EGR1/NPM1 as primary endpoints.
• Single-nucleus follow-up to resolve myofiber vs immune vs satellite-cell contributions to the LR signal — the BMP5→BMPR1A and DLL1→NOTCH1 progressive-blunting axes are particularly worth resolving.
• In vitro causal-direction studies in primary human myotubes with MET inhibition vs agonism under postprandial-like nutrient flux to settle the inhibition-vs-rescue question before any in vivo commitment.
• Multiple time-point MMTT (15min / 1h / 3h) to capture the temporal envelope.

The original paper identified a real biological signal in 21 muscle biopsies. A modern reanalysis pipeline — paired DESeq2, decoupleR, LIANA, ensemble stability selection, composite druggability scoring across DGIdb / ChEMBL / Open Targets — reproduces the core finding, extends it with a graded metabolic divergence, a discovery-grade biomarker panel anchored on MAPRE2, SGK1, BMPR1A, MECP2, NPM1, IRF8, EGR1, and turns the receptor-level finding into a ranked druggability list led by MET (0.792), ERBB3 (0.766), KIT (0.659) — all backed by approved Tier-1 drugs.

For a biopharma audience: this is exactly the use case where a small, well-designed paired clinical study becomes a tractable target-ID resource — if the analysis depth matches the design quality. The translational risk is well-defined and the next-step experiments are specific. The data warrant a Phase-1 discovery investment, not a clinical commitment.

Re-analysis based on GSE231509 (Makhnovskii et al. 2023; PMID 37545425). Methods: paired DESeq2 with apeglm shrinkage; limma-voom + duplicateCorrelation for the interaction contrast; fgsea on Hallmark / KEGG / Reactome; decoupleR on CollecTRI + DoRothEA regulons; LIANA-curated 2,100-pair LR set with bulk-adapted ΔL × ΔR scoring; nested-CV ElasticNet + RandomForest with 200-permutation null and 200-resample stability selection; composite druggability score integrating DGIdb, ChEMBL, Open Targets.