Finding a cell that researchers might want to remove is harder than it sounds. Senescent cells can persist after stress or damage, but they do not all display the same identifying signals. A test that misses some of them, or mistakes other cells for them, can make an experimental treatment difficult to evaluate.
A September 21 technical report in Nature Aging, highlighted by MIT, offers a way to connect two kinds of evidence: the chemistry of a cell and the genes it is using. Called RamanOmics, the approach combines Raman microscopy, RNA measurements and machine learning in mouse lung and skin. Its contribution is a richer map of senescence-associated cell states, not evidence that aging has been slowed or reversed.
What the researchers did
The team, involving MIT, Massachusetts General Hospital and Harvard Medical School, compared lung and skin from three 2-month-old mice with those from three 26-month-old mice. A separate wound-repair experiment examined skin from three 24-month-old mice, using samples collected at wounding and three days later. These were animal tissue experiments, not a human study or a test limited to cultured cells.
The imaging was performed outside the animals on removed tissue, cut into 14-micrometre sections and chemically fixed. Researchers first collected Raman spectra, then measured gene activity in the same sections using STARmap spatial transcriptomics. This method records RNA signals while retaining their locations in tissue. A panel of 890 genes supplied cell-identity and aging-related information.
Single-nucleus RNA sequencing from the same tissue blocks provided a broader gene-expression reference. Computational mapping used that reference to estimate additional, unmeasured genes in the spatial data. Those estimates should not be confused with direct measurements of every gene in every imaged cell. The team aligned Raman and spatial RNA information for 15,201 lung cells and 11,069 skin cells.
What they found
A shared signal emerged around Raman shifts of 1,131 to 1,135 inverse centimetres, a spectral region associated with lipids. These features were enriched in cells classified as p21-positive in both tissues. Other patterns depended on location: lung cells showed gene programs associated with tissue scaffolding and TGF-beta signaling, while skin cells showed programs involved in epidermal differentiation and barrier maintenance.
In wounded skin, increased p21 expression accompanied several of the same epidermal genes and lipid-associated Raman signals. This extended the observations to a repair setting. It did not establish that the lipid changes cause senescence or that removing the cells would improve healing.
The classification results were useful but modest. Combining Raman and gene-expression features raised test accuracy from 73.7% to 77.6% in lung and from 61.8% to 65.5% in skin, gains of about four percentage points. Raman alone achieved 48.7% accuracy in lung and 60.0% in skin on the balanced test data. Its area under the receiver operating characteristic curve was about 0.54 and 0.67, respectively; 0.5 represents chance-level ranking. The results support complementary information, not a highly reliable stand-alone optical detector.
Why senescent-cell measurement is difficult
Cellular senescence is a biological state commonly involving a durable halt to cell division, alongside changes in metabolism, gene activity and communication with surrounding tissue. It is one process associated with aging, not another name for aging itself. Senescence can also occur in young tissue and contribute to development or repair.
There is no single universally reliable marker because senescence varies with cell type, tissue, trigger and time. Proteins such as p16 and p21 help restrain cell division, but neither identifies every senescent cell, and their presence is not unique to senescence. Researchers therefore combine evidence such as cell-cycle changes, damage signals and altered secretion, interpreted in context. The SenNet detection recommendations address this tissue-specific challenge.
RamanOmics still depends on a working definition. Here, the reference labels were based on detectable Cdkn1a RNA, the gene encoding p21. The classifier learned patterns associated with p21-positive versus p21-negative cells. It did not independently establish every cell’s senescence status against a universal biological gold standard. The authors explicitly acknowledge that their definition represents a p21-positive subset.
What Raman microscopy and AI add
Raman spectroscopy shines light on a sample and measures tiny energy changes in the scattered light. Those changes reflect vibrations of chemical bonds. The resulting spectrum acts like a chemical fingerprint, with features associated with lipids, proteins and other molecular groups. Raman microscopy maps those fingerprints across a tissue. Overlapping signals mean a peak does not necessarily identify one particular molecule.
The machine-learning classifier was a random forest, an ensemble of decision trees that learns combinations of measurements associated with a label. Researchers compared models using Raman features, gene-expression features and both together. Feature-ranking analyses identified which measurements contributed most to the predictions. The resulting “barcode” is a compact representation of informative features, not a newly discovered universal senescence marker.
The terminology needs care. The Raman measurement is label-free: it reads intrinsic chemical signals without adding a senescence-targeting dye or antibody. It is nondestructive in the sense that the imaged section remained available for subsequent spatial RNA profiling. But the complete workflow included fixation, molecular probes and tissue processing. It was not entirely label-free, and it did not demonstrate repeated measurements of living cells.
Calling this a noninvasive test for people would go further than the experiment. “In situ” here refers to cells in their tissue context, not imaging inside a living patient. MIT’s announcement describes a possible future endoscope and efforts to adapt the approach to human tissue. Those are development goals.
What this could mean for senolytic and aging research
Our earlier SENESCENCE2030 roadmap coverage examined the need for better-defined targets and measurements. This study supplies a concrete experimental approach to part of that problem: connecting a cell’s molecular program with its chemistry and position in tissue.
If validated more broadly, such measurements could help researchers characterize which cell states a senolytic affects, compare responses across tissues, or investigate how senescence changes during repair. Senolytics aim to selectively eliminate senescent cells. A changed optical signature alone would not prove that cells were eliminated, that a harmful state was corrected, or that health improved. This study did not test a senolytic treatment or lifespan extension.
The limitations
The biological sample was small, with three mice per age group and only two organs in the aging comparison. Thousands of measured cells do not substitute for large numbers of independent animals. The young mice were all male; the older group contained two males and one female. Human tissue, broader disease settings and independent cohorts still require validation.
The model evaluation also leaves room for optimism bias. Researchers balanced the rare p21-positive cells against a downsampled p21-negative group, then randomly split cells 70:30 for training and testing. This was not a reported test in an entirely separate cohort of animals. Accuracy and precision from that balanced dataset should not be read as clinical performance in tissue where the target cells are uncommon.
Practical hurdles remain. MIT reports roughly 30 hours to analyze a tissue area of about one square millimetre. Estimated cell boundaries can admit signals from neighboring structures. Some gene values were computationally inferred, and chemical assignments need confirmation with independent methods. The observed associations do not show which biochemical changes drive senescence.
Bottom line
RamanOmics adds a chemical view to the molecular study of senescence, with evidence from fixed mouse lung, skin and wound samples. The strongest result comes from combining measurements. It is an early research and measurement advance, with substantial validation still needed before a stand-alone detector or clinical diagnostic test could follow. It offers no evidence that human aging has been slowed or reversed.
Sources and reporting note
This article is based on the full peer-reviewed Zhang and colleagues Nature Aging paper, including its methods and reported model performance, with MIT News used for the imaging-time estimate and development plans. The distinction between tissue preservation and noninvasive patient testing, and the implications of the cell-level test split, are The Lifespan Brief’s assessment of the reported design.