The research team, led by Hong Liu, Zicheng Liang, Jun Tan, and Chunyun Yuan from institutions in Changsha and Hangzhou, China, took an unusual approach to a familiar problem. Rather than running new laboratory experiments, they mined a publicly available dataset of skeletal muscle biopsies from ten subjects who underwent three distinct phases: an initial baseline, nine days of strict bed rest that reliably induces insulin resistance, and four weeks of exercise retraining that reverses it. This bidirectional design, in which the same individuals pass from normal metabolism into insulin resistance and back again, offered a rare opportunity to identify genes whose expression tracks the reversible dynamics of insulin sensitivity rather than static disease states.

To sift through the transcriptomic data, the researchers combined two complementary strategies. Weighted gene co-expression network analysis, or WGCNA, grouped thousands of genes into modules whose collective activity correlated with the metabolic phase, while differential expression analysis flagged individual genes whose levels changed significantly between phases. In the bed rest comparison, a module of 319 genes correlated strongly with the insulin-resistant state, and 66 genes were differentially expressed. In the exercise retraining comparison, two modules containing 237 and 417 genes tracked recovery, with 46 differentially expressed genes. Intersecting these results produced candidate gene pools of 46 and 42 genes for the two phases respectively.

The critical filtering step came from machine learning. The team applied LASSO regression, a technique that shrinks the coefficients of less informative genes to zero, and random forest analysis, which ranks genes by how much they improve classification accuracy across thousands of decision trees. Only genes that survived both algorithms were considered candidate targets. When the dust settled, WNT4 emerged as the sole gene selected by both methods in both the bed rest and exercise retraining phases. Its expression followed a striking bidirectional pattern: it rose after bed rest and fell again after exercise retraining, suggesting that its levels are tightly coupled to the metabolic status of the muscle rather than reflecting a one-way injury response.

To assess how reliably WNT4 could distinguish between metabolic states, the researchers used leave-one-out cross-validation, a rigorous scheme in which each sample is predicted by a model trained on all the others. This approach avoids the inflated accuracy estimates that plague small-sample studies using conventional neural networks. WNT4 achieved an area under the receiver operating characteristic curve of 0.88 in the bed rest comparison and 0.87 in the exercise retraining comparison, remarkably consistent values across two independent physiological transitions. Other candidate genes, including AQP7, GPD2, CTGF, and FOXM1, performed well in one phase but not both, underscoring WNT4’s unique position.

Gene set enrichment analysis then revealed what WNT4 might actually be doing. When samples were split into high- and low-WNT4-expression groups, the pathways positively correlated with high expression were dominated by energy metabolism, including oxidative phosphorylation, the TCA cycle, fatty acid metabolism, and the PPAR signaling pathway. This focused functional signature led the authors to propose what they call an endogenous compensation hypothesis: during the early phase of disuse, WNT4 may act as a protective factor that helps muscle stabilize its mitochondrial network and maintain metabolic homeostasis. WNT4 is known to inhibit the canonical Wnt signaling pathway and to keep muscle stem cells in a quiescent state, both of which could serve protective roles under metabolic stress.

But the hypothesis carries a darker corollary. If bed rest stress persists, the authors speculate, sustained WNT4 overexpression might excessively suppress the canonical Wnt and beta-catenin pathway, which previous studies have shown is needed for normal insulin-stimulated glucose uptake in skeletal muscle. In this scenario, the initial compensatory response would transition into a maladaptive state, contributing to intramyocellular lipid deposition and impaired insulin signaling through nodes such as AMPK and IRS-1. The authors describe this as a stage transition from endogenous compensation to pathological decompensation, a framework they acknowledge is entirely hypothetical and requires experimental validation through WNT4 knockdown and overexpression studies.

The regulatory architecture behind WNT4 adds another layer of complexity. Using database predictions, the team constructed a network linking transcription factors, messenger RNAs, microRNAs, and long non-coding RNAs. The transcription factor EGR1 emerged as the predicted upstream activator of WNT4 transcription, while the microRNA hsa-miR-1238-3p, buffered by the long non-coding RNAs LINC00969 and PRKCQ-AS1, was predicted to fine-tune WNT4 expression at the post-transcriptional level. Two circular regulatory loops involving these components scored above 0.90 on both random forest and support vector machine confidence metrics. The authors stress that these relationships are purely computational predictions and have not been experimentally confirmed.

Perhaps the most ambitious part of the study addresses a translational bottleneck: a search of the DrugBank database confirmed that no approved drugs currently target WNT4 directly. To fill this gap, the researchers built a computational drug discovery pipeline. They started with the amino acid sequence of human WNT4 and used AlphaFold to predict its three-dimensional structure, achieving an average confidence score of 93.12 on the pLDDT scale. Because no experimental crystal or cryo-EM structure of human WNT4 exists, the predicted binding pocket, located near residues CYS208, GLY210, and SER214 around the known lipid modification site at Ser212, remains a computational construct awaiting validation.

From there, the team deployed a deep learning docking model called DiffDock, which learns the conformational distribution of protein-ligand binding from data rather than relying on fixed sampling rules. After filtering more than 700,000 compounds from natural product and marine product databases using Lipinski’s rules and a quantitative drug-likeness score, 79,319 molecules were successfully docked against the predicted WNT4 pocket. Three compounds emerged at the top: CMNPD4923, identified as the marine sponge alkaloid Damirone B; CMNPD27319, the diaryl ether Aspergilol E from marine Aspergillus fungi; and CMNPD30941, a substituted isochromanone derivative. All three formed hydrogen bonds with the same pocket residues at distances between 1.70 and 2.42 angstroms.

One hundred nanosecond molecular dynamics simulations confirmed that all three protein-ligand complexes reached stable equilibrium, with root mean square deviations settling between 0.67 and 0.86 nanometers during the second half of the simulations. Binding free energy calculations using the MM-GBSA method ranked CMNPD4923 strongest at minus 20.08 kilocalories per mole, followed by CMNPD27319 at minus 17.13, while CMNPD30941 lagged at minus 5.80. Notably, this ranking diverged from the DiffDock confidence scores, suggesting that no single computational metric is sufficient and that multi-scale validation is essential. Density functional theory calculations added a quantum mechanical layer, revealing that the compounds possess active frontier orbital distributions with potential for charge transfer, though these gas-phase calculations approximate rather than replicate the protein pocket environment.

The authors are careful to frame every result as a hypothesis-generating prediction rather than a confirmed finding. The sample size of ten subjects is small, WNT4 was not cross-validated in an independent cohort, the protein structure is predicted rather than experimentally resolved, and the binding energies omit entropy contributions. Actual binding affinity would need to be confirmed by surface plasmon resonance or microscale thermophoresis, and functional effects by cellular assays. Nevertheless, the study demonstrates a replicable computational framework that spans from classical mechanics to quantum chemistry, and it offers a concrete chemical starting point for a target that has never been drugged. If future experiments validate the endogenous compensation hypothesis, WNT4 could become a new focal point for understanding how exercise reprograms muscle metabolism, and the three marine-derived compounds could serve as lead structures for developing the first targeted therapeutics aimed at this enigmatic signaling molecule.

Subject of Research: Computational identification of WNT4 as a biomarker and drug target in exercise-mediated skeletal muscle insulin sensitivity

Article Title: Computational identification of candidate biomarkers and compounds associated with skeletal muscle insulin sensitivity improvement mediated by physical activity: an integrative multi-omics and computational chemistry study

Article References: Liu, H., Liang, Z., Tan, J., & Yuan, C. (2026). Computational identification of candidate biomarkers and compounds associated with skeletal muscle insulin sensitivity improvement mediated by physical activity: an integrative multi-omics and computational chemistry study. Frontiers in Bioinformatics, 6, Article 1885995. https://doi.org/10.3389/fbinf.2026.1885995

Image Credits: AI Generated

DOI: 10.3389/fbinf.2026.1885995

Keywords: WNT4, insulin sensitivity, skeletal muscle, physical activity, insulin resistance, machine learning, WGCNA, molecular docking, DiffDock, molecular dynamics, natural products, drug discovery