AI Network Model Predicts How Aging Tissue Signals Reprogram Cells
A new computational framework maps how secreted factors from aged tissue rewire gene expression, pinpointing mitochondrial dysfunction as a central driver.
Summary
Aging tissues release a mix of signaling proteins that change how nearby cells behave, but tracing those effects is hard. Researchers built a computer model of gene activity in human knee cartilage and simulated how proteins secreted by young and aged mouse knee fat pads would ripple through it. They then compared the predictions with real RNA sequencing of aged cartilage cells treated with fat pad-conditioned medium. The model recovered many of the same biological pathways seen experimentally, especially those tied to mitochondrial energy production, a core hallmark of aging. If the approach holds up in other tissues, it could help scientists screen which aging-related signals drive tissue decline and which interventions might restore more youthful cell behavior.
Detailed Summary
Aging tissues release a complex mix of secreted proteins, often called the senescence-associated secretome, that can reshape neighboring cells. Tracing those downstream effects is difficult, particularly in slow-turnover tissues such as articular cartilage, where change accumulates gradually. This study introduces an in silico perturbation framework that simulates how extracellular signals propagate through a tissue-specific gene network, using knee cartilage and the infrapatellar fat pad (IFP) as the test case.
The team built a cartilage-specific co-expression network from public dataset GSE114007: 18 healthy cartilage samples (mean age 38) and 20 knee osteoarthritis samples (mean age 66). After filtering, 13,629 genes remained. Weighted gene co-expression analysis with a topological overlap matrix produced a network, trimmed to the top 50,000 edges. The dataset behaved as expected: 1,294 genes were differentially expressed (800 up, 494 down), with Sox9 down and Mmp13, Col1a1 and Fap up. The network was scale-free, and mitochondrial complex I genes Ndufb9 and Ndufab1 emerged as top hubs.
For inputs, IFPs from young (4–6 months) and aged (21–24 months) male mice were cultured for 72 hours and their secretomes profiled by mass spectrometry. Thirty proteins differed at a nominal p<0.05 (10 higher, 20 lower in aged IFPs). These were mapped to receptors through the NicheNet database, and the receptors seeded a Random Walk with Restart (restart probability 0.7) to rank genes by their network proximity to the signals.
Benchmarking used RNA-seq of aged mouse chondrocytes treated for 72 hours with young or aged IFP-conditioned medium (four samples per group), which yielded 84 differentially expressed genes (17 up, 67 down). The model's top genes included NDUFB9, NDUFAB1, ribosomal genes (RPS15, RPL32, RPS16) and protein-processing genes (CHCHD2, TOMM6, SPCS1). Using the top 50 genes, Gene Ontology enrichment found 46 pathways in the RNA-seq data and 50 in the model, with 18 shared. Semantic similarity grouping (threshold 0.7) raised the overlap to 27 pathways, dominated by mitochondrial processes. ROC, predictive value analyses and SDHA immunostaining supported a primary effect on mitochondrial respiration; stricter FDR cutoffs improved positive predictive value.
The framework offers a data-driven way to prioritize which age-dependent extracellular cues may drive cellular dysfunction, and it could be extended to other tissues and diseases. Several caveats apply. Protein inputs used uncorrected p-values and were few in number. The experimental arm had only four samples per group, used male mice only, and was done in vitro. The network came from human osteoarthritis versus healthy tissue, which conflates age and disease, and it was applied to mouse data. Validation occurred at the pathway level, and the semantic reclassification step may inflate apparent agreement. This is a proof of concept, not a clinical tool.
Key Findings
- Cartilage co-expression network built from 38 human samples (18 healthy, mean age 38; 20 knee osteoarthritis, mean age 66) covering 13,629 genes and the top 50,000 edges
- Osteoarthritis vs healthy cartilage had 1,294 differentially expressed genes (800 up, 494 down), with Sox9 down and Mmp13, Col1a1 and Fap up
- Mitochondrial complex I genes Ndufb9 and Ndufab1 were top hub genes in the scale-free network
- Aged vs young IFP secretome: 30 differentially abundant proteins (10 up, 20 down in aged) at nominal p<0.05, used as model inputs
- Aged chondrocytes treated with aged vs young IFP medium showed 84 differentially expressed genes (17 up, 67 down; FDR<0.05, |log2FC|>1)
- Top-50-gene analysis: 18 GO pathways shared between model and RNA-seq, rising to 27 after semantic similarity grouping (≥0.7), mainly mitochondrial
- Stricter FDR cutoffs increased positive predictive value, and SDHA staining was used to test the predicted mitochondrial effect
Methodology
The authors built a WGCNA/TOM co-expression network from human cartilage RNA-seq (GSE114007; n=18 healthy, 20 knee osteoarthritis) and simulated signal propagation with Random Walk with Restart (restart 0.7) seeded from receptors of IFP-secreted ligands (NicheNet). Ligands came from LC–MS/MS secretomes of IFPs from young (4–6 mo) and aged (21–24 mo) male C57BL/6 mice. Predictions were benchmarked against bulk RNA-seq of aged mouse chondrocytes treated with young or aged IFP-conditioned medium (n=4 per group) using GO enrichment overlap, semantic similarity, ROC/AUC, PPV and NPV, plus SDHA immunostaining.
Study Limitations
Secretome inputs used nominal p<0.05 without multiple-testing correction and involved only 30 proteins, and the experimental benchmark had four samples per group in male mice only. The network was built from human osteoarthritis versus healthy tissue (age and disease are confounded) but applied to mouse secretome and chondrocyte data, and validation was at the pathway level with a semantic-similarity step that may inflate overlap. No conflicts of interest were available in the supplied text.
Enjoyed this summary?
Get the latest longevity research delivered to your inbox every week.
Enter your email to subscribe: