Machine learning defines cellular senescence signatures in lung

Cellular senescence has long been one of biology's most paradoxical phenomena. In healthy tissue, senescent cells—those that have permanently exited the cell cycle in response to stress, DNA damage, or telomere shortening—act as a safeguard against cancer, halting division before damage can accumulate. Yet as these cells linger, they secrete inflammatory signals that can, over time, reshape the tissue environment in ways that fuel tumor progression. Researchers have struggled for decades to quantify senescence reliably, largely because the field has lacked a standardized, objective measure of how 'senescent' a given cell or tumor actually is. Now, a team led by Lifei Ma and Huiyang Li, with senior authors Gong-Hong Wei, Xiaoman Wang, and Hou-Zao Chen, has built exactly that: a machine learning model called the Predictive Cellular Senescence Model, or PreCSenM, which not only measures senescence with unprecedented consistency across datasets but has already uncovered a promising therapeutic vulnerability in lung adenocarcinoma, the most common form of lung cancer worldwide. The core problem PreCSenM addresses is methodological. Senescence has traditionally been assessed through a patchwork of markers—senescence-associated β-galactosidase staining, p16 and p21 expression, telomere-associated DNA damage foci—each of which captures only a partial slice of the phenomenon and each of which behaves differently depending on cell type, species, and experimental context. That fragmentation has made it genuinely difficult to compare senescent cell burden across studies, let alone across patients. To get around it, the team assembled an enormous training compendium: 888 transcriptomic profiles spanning diverse cell types and a wide range of senescence-inducing conditions, from replicative exhaustion to oncogene activation to drug treatment. After careful normalization and batch-effect correction, they applied the Boruta feature selection algorithm, an approach designed to identify genes whose predictive signal exceeds what random chance would produce, to distill a stable consensus cellular senescence-related gene signature, which they abbreviated CSGS. With the gene signature in hand, the researchers faced a second design decision: which algorithm should convert a tumor's expression profile into a senescence score? Rather than betting on a single method, they benchmarked ten machine learning algorithms—including logistic regression, support vector machines, random forest, XGBoost, partial least squares regression, artificial neural networks, and random survival forest models—within a rigorous cross-validation framework. The winning configuration became PreCSenM, which integrates the CSGS to output a continuous cellular senescence (CS) score: a single standardized number representing the senescence level of any given sample. Importantly, the model was tested against existing senescence quantification approaches using both area under the ROC curve and area under the precision-recall curve metrics, and it outperformed them on both normal and cancer transcriptomic datasets. A web portal at http://precsenm.bmicc.org/ makes the tool freely accessible to other researchers, lowering the barrier for labs that lack computational infrastructure. The most clinically significant results came from applying PreCSenM to lung adenocarcinoma, or LUAD. Using data from The Cancer Genome Atlas (TCGA) and other public cohorts, the team found that the CS score behaved as a robust predictor of clinical outcomes, stratifying patients into groups with distinctly different overall survival. Perhaps counterintuitively, higher senescence scores correlated with better prognosis—an observation the authors describe as revealing a 'pro-senescence' potential in this cancer type. This fits a growing body of evidence that inducing senescence in tumor cells can arrest their proliferation, provided the senescent cells do not persist long enough to promote inflammation or escape the state. The finding suggests that, in LUAD at least, pushing tumor cells into senescence may be therapeutically beneficial. Multi-omics analysis helped explain why. Patients with higher CS scores showed signatures of greater genomic stability, including patterns consistent with lower somatic copy-number alteration burden, and displayed enhanced immune-related features—infiltration of immune cells and activation of immune signaling pathways—across the tumor microenvironment. The team also validated the CS score's biological meaning experimentally, showing that it correlated with telomere-associated foci, one of the most reliable physical markers of deep cellular senescence. Together, these lines of evidence suggest that senescence-high tumors in LUAD exist in a state that is both genetically quieter and more visible to the immune system, a combination that aligns with their better clinical outcomes. Having established that the CS score could measure senescence meaningfully, the researchers turned the model around and used it as a drug-discovery engine. By mining drug-induced transcriptional signatures from resources such as the Library of Integrated Network-Based Cellular Signatures (LINCS), they searched for compounds whose effects on gene expression most closely resembled the high-senescence state. The screen converged on a clear winner: histone deacetylase inhibitors, or HDACis—a class of epigenetic drugs that includes the FDA-approved agent vorinostat (also known as SAHA) and the laboratory tool trichostatin A. HDAC enzymes remove acetyl groups from histone proteins, tightening chromatin and silencing genes; inhibiting them loosens that grip, broadly altering transcription. The prediction was that HDACis would act as potent inducers of senescence in LUAD cells. Laboratory experiments confirmed the computational prediction. When the team treated lung adenocarcinoma cells with HDAC inhibitors, transcriptional and epigenetic profiling—combining RNA sequencing with ATAC-seq, a technique that maps regions of open, accessible chromatin—revealed a coordinated shift toward the senescence program. More strikingly, the profiling pinpointed a single transcription factor at the center of the shift: FOSB, a member of the AP-1 family of immediate-early transcription factors. ATAC-seq data showed that chromatin accessibility changes upon HDACi treatment opened regulatory regions targeted by FOSB, implicating it as a core driver of the senescence program rather than a bystander. To test causality directly, the researchers used small interfering RNA to knock down FOSB expression, and the result was decisive: HDACi-treated cells with reduced FOSB underwent markedly less senescence, confirming that FOSB is required for the drugs to push tumor cells into the senescent state. The implications reach in two directions at once. For the senescence field, PreCSenM offers what has been missing: a multidimensional, standardized quantification tool that bridges computational prediction with clinical relevance and mechanistic validation. Instead of arguing about which single marker best defines senescence, researchers can now compute a continuous score grounded in a consensus gene signature and validated across hundreds of profiles. For oncology, the study provides a concrete pipeline—from model to drug screen to mechanism—that could be replicated for other cancer types. If high senescence is genuinely favorable in LUAD, then therapies that deliberately induce it, with careful attention to clearing senescent cells before they turn harmful, become an actionable strategy in precision oncology. The FOSB finding adds a specific molecular handle for that effort, and one that may help predict which patients will respond to HDAC inhibitors. The study, published in Genome Medicine, was a large collaborative effort involving institutions across China and the United States, including the Chinese Academy of Medical Sciences and Peking Union Medical College, Tianjin Institutes of Health Science, the National Cancer Center, the University of California, Riverside, Shanghai Jiao Tong University, Sichuan University, Peking Union Medical College Hospital, and Fudan University. It was supported by the National Natural Science Foundation of China, the National Key Research and Development Project, and other national funding programs. Because all analyses relied on publicly available, de-identified datasets, the computational core of the work is fully reproducible, and the web portal ensures that any researcher can score their own samples through PreCSenM without installing specialized software. Caveats remain, as they always do. The model was trained on bulk transcriptomic profiles, and applying it to single-cell data or to tissues where senescence and quiescence are hard to distinguish will require further validation. Clinically, the pro-senescence strategy must contend with the known risks of senescent cell accumulation, and prospective studies will be needed to determine whether HDACi-induced senescence in patients produces the same favorable immune and genomic landscape seen in retrospective cohort analysis. But the conceptual advance is hard to overstate: for the first time, senescence in cancer can be measured the way tumor mutation burden or microsatellite instability is measured—as a quantitative, standardized biomarker—and then acted upon. A phenomenon once relegated to petri dishes and staining assays now has a place in the clinical conversation, and lung adenocarcinoma may be the first battleground where that conversation changes treatment. Subject of Research: Cellular senescence quantification in lung adenocarcinoma using a machine learning model (PreCSenM), including identification of HDAC inhibitors and the transcription factor FOSB as drivers of pro-senescence therapy. Subject of Research: Medicine Article Title: Machine learning-based definition of cellular senescence reveals pro-senescence potential implications in lung adenocarcinoma Article References: Ma, L., Li, H., Li, Y., Lin, Z.-A., Li, J.-Q., Zhang, Y.-Z., Zhang, P., Yao, Z., Li, J., Xiong, M., Cao, Y., Li, R., Yang, C., Tang, X., Chen, M., Wang, H.-P., Zheng, W., Yang, J., Wang, X., ... Chen, H.-Z. (2026). Machine learning-based definition of cellular senescence reveals pro-senescence potential implications in lung adenocarcinoma. Genome Medicine. https://doi.org/10.1186/s13073-026-01686-y Image Credits: AI Generated DOI: 10.1186/s13073-026-01686-y Keywords: cellular senescence, machine learning, PreCSenM, lung adenocarcinoma, HDAC inhibitors, FOSB, AP-1, multi-omics, gene signature, precision oncology, transcriptomics, Genome Medicine