SenCat: Cataloging human cell senescence through multi-omic profiling of multiple senescent primary cell types

May combine two claims.

Senescent phenotypes are heterogeneous, varying with cell type, inducing stimulus, microenvironment, and time since induction, and distinct transcriptomic and proteomic profiles have been reported across models.

“Senescent phenotypes vary depending on the cell type, senescence-inducing stimulus, microenvironment, and time since senescence induction, with distinct transcriptomic and proteomic profiles reported across models.9–11”

• cites [9] sourcePartly supported

  The Results report distinct secretory protein profiles across inducing stimuli and between fibroblasts and epithelial cells; Fig. 3 also describes inducer-specific RNA and protein signatures. The cited paper does not establish that phenotypes vary with microenvironment or time since induction. Page numbers are not available in our copy. (LLM, conf=0.88)

  Experimental — accuracy not yet measured
• cites [10] sourceNot checked — no full text of the source
• cites [11] sourcePartly supported

  The Results and Discussion report distinct transcriptomic signatures by cell type and inducing stimulus, as well as changes over time after induction. They also note variability associated with laboratory culture conditions, but do not establish microenvironment-dependent phenotypes more broadly. This paper analyzes transcriptomes; it does not report distinct proteomic profiles across models. Page numbers were not available in our copy. (LLM, conf=0.86)

  Experimental — accuracy not yet measured

May combine two claims.

ML-defined senescent populations were enriched for other senescence signatures, with some lineage-specific disparities.

“the ML-defined senescent populations were enriched for other senescence signatures,12–18 although with some lineage-specific disparities for some lists (Figure 4G)”

• cites [12] sourceCould not verify

  The section “A need for senescence-specific markers” says senescent cells vary by cell type and stimulus, but it does not examine ML-defined populations. Searches for machine learning, signature enrichment, and lineage-specific lists found no passage reporting the claimed enrichment or disparities. (LLM, conf=0.76)

  Experimental — accuracy not yet measured
• cites [13] sourcePartly supported

  The “Translation to humans” section (page unavailable) describes senescence-burden signatures and says senescent-cell phenotypes vary by biological context. It does not report that ML-defined senescent populations were enriched for other signatures or document the claimed lineage-specific disparities. (LLM, conf=0.81)

  Experimental — accuracy not yet measured
• cites [14] sourceSupported
• cites [15] sourceSupported
• cites [16] sourceSupported
• cites [17] sourceSupported
• cites [18] sourceCould not verify

  In “Decoding senescence: A practical assessment of gene sets and resources” (page unavailable), the paper discusses differences among senescence gene lists and their applicability across cell types. Searches for ML-defined populations, enrichment for other signatures, and lineage-specific disparities found no passage reporting the claimed result. (LLM, conf=0.74)

  Experimental — accuracy not yet measured

May combine two claims.

The TGFβ pathway representation in non-senescent epithelial cells from old lungs is consistent with enhanced paracrine signaling from increased numbers of senescent neighbors and with established features of epithelial aging.

“consistent with enhanced paracrine signaling from increased numbers of senescent neighbors and with established features of epithelial aging.23,25,32–35”

• cites [23] sourcePartly supported

  In Results (‘VPF treatment reduces senescent cell burden in mice’; page unavailable), the paper reports age-associated TGF-β signatures in some lung epithelial cell types and TGF-β signaling from fibroblasts and capillary aerocytes affecting neighboring epithelial cells. This supports the proposed paracrine interpretation, but the paper does not establish that the epithelial cells are non-senescent or that increased numbers of senescent neighbors caused their pathway representation. (LLM, conf=0.82)

  Experimental — accuracy not yet measured
• cites [25] sourcePartly supported

  The ‘Response to EMT-activating factors and their links to fibrosis’ section says senescent cells release TGF-β and that extracellular signals could affect surrounding cells; its Figure 3 discussion also describes senescent-cell accumulation during aging. The paper does not establish increased TGF-β signaling in non-senescent epithelial cells from old lungs or identify the claimed features of epithelial aging. (LLM, conf=0.84)

  Experimental — accuracy not yet measured
• cites [32] sourcePartly supported

  The Introduction (page unavailable) says senescent cells increase with age in the lung, and the Discussion (page unavailable) says their secretions can influence surrounding cells. Those passages support the proposed paracrine explanation, but the paper does not establish TGFβ pathway representation in non-senescent epithelial cells from old lungs or link that specific finding to senescent neighbors. (LLM, conf=0.84)

  Experimental — accuracy not yet measured
• cites [33] sourcePartly supported

  The Discussion (page unavailable) identifies senescent cells as a source of profibrotic signaling, including TGFβ, and the Results report senescence markers in lung epithelial cells. The paper does not examine TGFβ pathway representation in non-senescent epithelial cells from old lungs or establish that their signaling comes from senescent neighbors. (LLM, conf=0.84)

  Experimental — accuracy not yet measured
• cites [34] sourcePartly supported

  The Fibrosis section describes damaged renal epithelial cells releasing paracrine signals, including TGFβ, and A brief history of EMT identifies TGFβ as an inducer of epithelial transition (page numbers unavailable). These passages support a general TGFβ-signaling connection, but do not address increased senescent neighbors, non-senescent cells in old lungs, or epithelial aging. (LLM, conf=0.83)

  Experimental — accuracy not yet measured
• cites [35] sourceSupported

• AssumptionThe authors state this as an assumption: Senescent fibroblasts drive epithelial dysfunction during aging.

May combine two claims.

Those earlier models show strong cross-dataset generalizability and robust scoring, yet they often come from mixed multi-study collections or narrower biological settings, which may restrict how well they capture the full range of senescent states across cell types and triggers.

“These models demonstrate impressive cross-dataset generalizability and scoring robustness, but they are often derived either from heterogeneous multi-study collections or from narrower biological settings, which can limit the extent to which they capture the full diversity of senescent states across cell types and triggers.14–17,20”

• cites [14] sourcePartly supported

  Split into 2 atoms — 0 confirmed, 2 partially supported, 0 disputed, 0 misattributed, 0 unverifiable. Expand the claim to see per-atom verdicts.

  Experimental — accuracy not yet measured
• cites [15] sourcePartly supported

  Split into 2 atoms — 1 confirmed, 1 partially supported, 0 disputed, 0 misattributed, 0 unverifiable. Expand the claim to see per-atom verdicts.

  Experimental — accuracy not yet measured
• cites [16] sourcePartly supported

  Split into 2 atoms — 1 confirmed, 1 partially supported, 0 disputed, 0 misattributed, 0 unverifiable. Expand the claim to see per-atom verdicts.

  Experimental — accuracy not yet measured
• cites [17] sourcePartly supported

  Split into 2 atoms — 0 confirmed, 2 partially supported, 0 disputed, 0 misattributed, 0 unverifiable. Expand the claim to see per-atom verdicts.

  Experimental — accuracy not yet measured
• cites [20] sourcePartly supported

  Split into 2 atoms — 1 confirmed, 1 partially supported, 0 disputed, 0 misattributed, 0 unverifiable. Expand the claim to see per-atom verdicts.

  Experimental — accuracy not yet measured

Classified as probably not a substantial claim.

Earlier work has produced RNA signatures of cellular senescence from various collections of human in vitro data.

“recent efforts have identified transcriptomic signatures of cellular senescence derived from a different panels of human in vitro datasets.14–17,20”

• cites [14] sourcePartly supported

  The Methods (Generation of SenMayo; page unavailable) describe a 125-gene senescence set assembled from 15 prior studies, supporting the claim that this work produced a senescence-related gene signature. But the genes were verified in human or mouse cells, and the Introduction describes validation in tissues from aged humans and mice; the copy does not establish that SenMayo was derived from a panel of human in vitro datasets. (LLM, conf=0.87)

  Experimental — accuracy not yet measured
• cites [15] sourcePartly supported

  The discussion (section labeled “1 3”; page unavailable) says earlier studies sequenced senescent cells cultured in vitro to derive signatures across several forms of senescence, and the results describe three in vitro-derived signatures from the literature. Our copy does not establish that the datasets used to derive those signatures were human, so it supports the claim except for that qualifier. (LLM, conf=0.82)

  Experimental — accuracy not yet measured
• cites [16] sourcePartly supported

  The Discussion (page unavailable) says previous studies generated transcriptomic senescence signatures from induced senescence in controlled in vitro environments. The Introduction (page unavailable) likewise describes signatures derived from cultured cells. These passages do not specify that the earlier signatures came from different panels of human datasets. (LLM, conf=0.82)

  Experimental — accuracy not yet measured
• cites [17] sourceSupported
• cites [20] sourceSupported

May combine two claims.

SenCat adds value beyond earlier computational and ML methods through its broad atlas, its in-house generation under tightly controlled and consistent conditions, its combination of RNA and protein data, and its capacity to derive both shared and context-limited senescence programs within one framework.

“the added value of SenCat lies in the breadth of the underlying atlas, the fact that it was entirely generated under strictly controlled and consistent settings in-house, the integration of transcriptomic and proteomic modalities, and the ability to derive both shared and context-restricted senescence-associated programs from a common framework.”

• cites [14] sourceCould not verify

  Split into 2 atoms — 0 confirmed, 0 partially supported, 0 disputed, 0 misattributed, 2 unverifiable. Expand the claim to see per-atom verdicts.

  Experimental — accuracy not yet measured
• cites [15] sourcePartly supported

  The Discussion (page not available) says the paper’s SenSig method identifies shared features of senescence while allowing tissue- and pathology-specific analysis. It does not establish those capabilities for SenCat, nor does it support the claimed broad, entirely in-house atlas generated under consistent conditions or integration of transcriptomic and proteomic data. The methods describe an RNA-based SenSig signature instead. (LLM, conf=0.84)

  Experimental — accuracy not yet measured
• cites [16] sourcePartly supported

  Split into 2 atoms — 0 confirmed, 2 partially supported, 0 disputed, 0 misattributed, 0 unverifiable. Expand the claim to see per-atom verdicts.

  Experimental — accuracy not yet measured
• cites [17] sourcePartly supported

  Split into 2 atoms — 0 confirmed, 1 partially supported, 0 disputed, 0 misattributed, 1 unverifiable. Expand the claim to see per-atom verdicts.

  Experimental — accuracy not yet measured
• cites [18] sourcePartly supported

  Split into 2 atoms — 0 confirmed, 1 partially supported, 0 disputed, 0 misattributed, 1 unverifiable. Expand the claim to see per-atom verdicts.

  Experimental — accuracy not yet measured
• cites [26] sourcePartly supported

  Split into 2 atoms — 0 confirmed, 1 partially supported, 0 disputed, 0 misattributed, 1 unverifiable. Expand the claim to see per-atom verdicts.

  Experimental — accuracy not yet measured
• cites [27] sourceCould not verify

  Split into 2 atoms — 0 confirmed, 0 partially supported, 0 disputed, 0 misattributed, 2 unverifiable. Expand the claim to see per-atom verdicts.

  Experimental — accuracy not yet measured

May combine two claims.

Senescence is complex, and the gene expression programs linked to it differ among cell types and inducing triggers.

“Our results underscore the complexity of senescence and the diversity of senescence-associated gene expression programs across cell types and triggers.”

May combine two claims.

For proteomic markers shared by at least 4 cell types, peroxisome was additionally enriched, and lysosome-related pathways matched transcriptomic results.

“When considering proteomic markers shared by at least 4 cell types, the results were similar except for the inclusion of “peroxisome” as an enriched “Hallmark” pathway (Figure 2E, top- and bottom-left graphs); for the “GO Cellular Components” analysis, most lysosome-related pathways matched the transcriptomic results (Figure 2E, top- and bottom-right graphs).”

May combine two claims.

The general processes metabolic reprogramming, tissue remodeling, and response to damage were conserved in the proteomic analysis, with some variation in internal composition.

“The general processes described for the transcriptomic analysis (“metabolic reprogramming”, “tissue remodeling”, and “response to damage”) were also conserved in the proteomic analysis (Figure 2F), although the internal composition of each of these categories varied somewhat.”

May combine two claims.

Proteomic markers declining with senescence showed reduction in processes related to cell cycle, nucleus, DNA, and RNA, matching transcriptomic results.

“Analysis of the proteomic markers declining with senescence matched the transcriptomic results, with a significant reduction in processes related to cell cycle, the nucleus, DNA, and RNA (Figure S5D–F).”

May combine two claims.

ML-derived senescence scores consistently and robustly stratified senescent versus proliferating cells across SenCat samples, whereas traditional markers such as p16, p21, IL6, GDF15, and LMNB1 showed variable performance.

“while traditional senescence markers such as p16, p21, IL6, GDF15, and LMNB1 showed variable performance depending on cell type and condition, the ML-derived senescence scores consistently and robustly stratified senescent versus proliferating cells across the SenCat samples profiled (Figure 3C).”

May combine two claims.

MX1, CXCL8, and GDF15 mRNAs, and FDXR, VDAC2 and TAGLN proteins showed strong positive coefficients, while TOP2A mRNA, some histone mRNAs, SERPINH1, STMN1 and several histones showed strong negative coefficients.

“Notably, MX1, CXCL8, and GDF15 mRNAs, and FDXR, VDAC2 and TAGLN proteins showed strong positive coefficients, while TOP2A mRNA and some mRNAs encoding histones, along with SERPINH1, STMN1 and several histones, showed strong negative coefficients.”

May combine two claims.

CDKN1A (p21) and GDF15 mRNAs were among the top 20 most shared upregulated markers, but their proteins were not among the top 20 shared elevated markers.

“Canonical markers often increasing in senescence, such as CDKN1A (p21) and GDF15 mRNAs, were among the top 20 most shared markers upregulated by all senescence models in our analysis (Figure 1E), but, remarkably, CDKN1A/p21 and GDF15 proteins were not among the top 20 shared elevated markers (Figure 1F).”

May combine two claims.

Metabolic pathways (adipogenesis, oxidative phosphorylation, fatty acid metabolism) and coagulation were associated with senescence, with higher variability across cell types than transcriptomic pathway analysis.

“Metabolic pathways (“adipogenesis”, “oxidative phosphorylation”, “fatty acid metabolism”), and “coagulation”, were also associated with cell senescence (Figure 2D), although with higher variability among cell types when compared to the transcriptomic pathway analysis.”

May combine two claims.

Reduced transcriptomic markers indicated inhibition of cell cycle progression, diminished nuclear functions, and broad inhibition of DNA replication, metabolism, and RNA metabolism and processing.

“Analysis of the reduced transcriptomic markers indicated that cell cycle progression-related pathways were inhibited in senescent cells, nuclear functions were diminished, and there was broad inhibition of DNA replication and metabolism, and RNA metabolism and processing (Figure S5A–C).”

May combine two claims.

CellPhoneDB analysis revealed numerous signaling axes potentially mediating crosstalk between senescent fibroblasts and non-senescent epithelial cells in lungs from old mice, with TGFB2 signaling standing out.

“This analysis revealed numerous signaling axes potentially mediating this crosstalk (Figure 4K), standing out among them was TGFB2 signaling, with pro-fibrotic and epithelium-remodeling roles (Figure 4L).”

• cites [31] sourceNot checked — no full text of the source

• AssumptionThe authors state this as an assumption: Senescent fibroblasts drive epithelial dysfunction during aging.

May combine two claims.

The proteomic score trajectories with age largely matched the patterns from snRNA-seq data, showing clear increases with age in several tissues and differences between tissues.

“Notably, the age-related trajectories of these proteomic scores broadly mirrored the trends observed in the snRNA-seq datasets (Figure 4P), with clear age-associated increases across multiple tissues and relative tissue-to-tissue differences (Figure 4O).”

May combine two claims.

In lung cell lineages, senescent cells displayed increased scores for p53, EMT, NF-κB, apoptosis, and hypoxia, and decreased association with E2F targets, compared to non-senescent counterparts.

“within each of these lung cell lineages, and to varying degrees, senescent cells displayed increased scores for pathways including p53, EMT, NF-κB, apoptosis, and hypoxia, compared to non-senescent counterparts, and decreased association with E2F targets”

May combine two claims.

Overlap between transcriptomic and proteomic senescence markers was low, with no cell type exceeding 20%.

“no cell type showed an overlap significantly higher than 20% (Figures S4E–G), indicating little concordance between transcriptomic and proteomic markers of senescence”

May combine two claims.

A mismatch remains between RNA and protein data across senescence models, reflecting multi-level gene regulation and complicating the definition of unified molecular signatures.

“First, there is still a mismatch between transcriptomic and proteomic data across senescence models, reflecting a multi-leveled regulation of gene expression and making it difficult to define unified molecular signatures.”

Probably not a substantial claim: may not pass S3 (load-bearing).

The refined senescence signature contained up to 400 genes.

“This strategy yielded a refined senescence signature of up to 400 genes.”

May combine two claims.

Across RNA and protein datasets, pathways related to p53, NF-κB, EMT, apoptosis, hypoxia, and lysosomes appeared repeatedly, although the particular RNAs or proteins varied with lineage or trigger.

“pathways such as those involving p53, NF-κB, EMT, apoptosis, hypoxia, and lysosomal function were recurrently represented across transcriptomic and proteomic datasets, even when the specific RNAs or proteins involved differed by lineage or trigger.”

May combine two claims.

The SenCat approach works well for senescence that is experimentally induced and for senescence that arises naturally.

“Importantly, this framework performs well in both acutely induced and naturally occurring senescence”

May combine two claims.

Regressing total UMI counts from AUCell scores yielded similar results, ruling out sequencing depth artifacts.

“To rule out potential artifacts caused by cell-specific differences in sequencing depth, total UMI counts (nCount_RNA) were regressed from AUCell scores, obtaining similar results (Figure S7B).”

May combine two claims.

The ML-derived senescence signatures are not overfitted to the original training dataset and are applicable to new biological contexts.

“Using transcriptomic and proteomic data, ML scores successfully distinguished senescent from non-senescent states, confirming that the signatures are not overfitted to the original training dataset and were applicable to new biological contexts.”

May combine two claims.

Increased transcriptomic markers indicate a tilt towards damage response pathways, while increased proteomic markers indicate tissue repair and metabolic reprogramming pathways.

“Together, the increased transcriptomic markers indicate a tilt towards damage response-related pathways, and the increased proteomic markers towards tissue repair and metabolic reprogramming pathways.”

Classified as probably not a substantial claim.

A cluster-by-cluster differential expression comparison of Day 0 cells against later time points kept only genes that did not show significant upregulation at Day 0 in any cluster.

“Specifically, we performed cluster-wise differential expression analyses between Day 0 cells and all other time points, retaining only genes that were not significantly upregulated in Day 0 in any cluster.”

Probably not a substantial claim: may not pass S3 (load-bearing).

Before fitting models for each tissue, the authors required SenCat markers to have enough detection across samples.

“Within each tissue, detection constraints were applied to retain only SenCat markers with sufficient coverage across samples before model fitting.”

Probably not a substantial claim: may not pass S3 (load-bearing).

Experimental samples from each cell line were collected and processed together to reduce batch effects.

“To minimize batch effects, we collected and processed experimental samples from each cell line at the same time.”

May combine two claims.

The ML-derived senescence score is robust, quantitative, and applicable to independent datasets.

“producing a robust, quantitative score applicable to independent datasets.”

Classified as probably not a substantial claim.

For every sample, cluster, and condition, senescent and non-senescent cells were paired deterministically by log10(nCount_RNA) with an optimal one-dimensional matching method.

“Specifically, within each sample, cluster, and condition, senescent and non-senescent cells were matched deterministically based on log10(nCount_RNA) using an optimal one-dimensional pairing strategy.”

Probably not a substantial claim: may not pass S3 (load-bearing).

Mass spectrometry samples were acquired in a block-randomized order to reduce batch effects.

“For MS acquisition, samples were run in a block-randomized design to minimize batch effects.”

Probably not a substantial claim: may not pass S3 (load-bearing). May combine two claims.

A small random portion of reference-condition cells was moved into the senescent class to prevent complete separation and make model fitting more stable.

“To avoid complete separation and stabilize model fitting, a small random subset of reference-condition cells was reassigned to the senescent class.”

• AssumptionRests on an assumption: Reference-condition cells are non-senescent.
• AssumptionRests on an assumption: All non-reference-condition cells are senescent, or at least initially senescent.

Classified as probably not a substantial claim.

When comparing senescent with non-senescent cells inside clusters, the authors used a depth-balanced matched design instead of comparing all cells directly.

“For downstream comparisons between senescent and non-senescent cells within clusters, we used a depth-balanced matched analysis rather than directly comparing all cells.”

Classified as probably not a substantial claim.

The same normalization and scoring strategy was applied to the external IMR-90 dataset without retraining the model.

“We applied the same normalization and scoring strategy to this dataset without retraining the model.”

Probably not a substantial claim: may not pass S3 (load-bearing). May combine two claims.

In the lung doxorubicin time-course snRNA-seq data, every Day 12 sample was dropped for technical problems, and one Day 1 replicate was also discarded as an outlier after sample-level preprocessing.

“For the lung doxorubicin time-course dataset, all Day 12 samples were excluded from downstream analyses because of technical issues, and one Day 1 replicate was additionally removed as an outlier after sample-level preprocessing.”

Probably not a substantial claim: may not pass S3 (load-bearing). May combine two claims.

Within each tissue, older-age scores were tested against the 3-month group with parametric and non-parametric methods, and multiple-testing correction was applied.

“For each tissue, scores at older ages were compared against the 3-month group using both parametric and non-parametric tests, followed by multiple-testing correction.”

• AssumptionRests on an assumption: 3-month-old mice are a valid young reference group for proteomic senescence score comparisons.

Classified as probably not a substantial claim. May combine two claims.

The analysis used only markers that rose in snRNA-seq, since AUCell is not yet able to detect decreased markers and could add noise.

“We used only markers that increased in the snRNA-seq analysis, as the AUCell approach is not yet adequate to identify reduced markers, and it may therefore introduce noise.”

Probably not a substantial claim: may not pass S3 (load-bearing). May combine two claims.

In bulk proteomic aging data, zeros were considered missing, abundances were log2-transformed, and each file was median-centered to lessen technical variation between assays.

“Zero values were treated as missing, abundances were log2-transformed, and median-centering was applied within each file to reduce technical differences across assays.”

Probably not a substantial claim: may not pass S3 (load-bearing).

Identifications at peptide and protein-group levels were filtered at a 1% false discovery rate.

“FDR was set to 1% for filtering identifications at the peptide and protein group levels.”

May combine two claims.

SenCat pairs RNA profiles with protein data from the same senescent cell types, narrowing its marker lists toward proteins that may be secreted or surface-exposed and detectable in body fluids.

“Second, SenCat complements transcriptomic profiles with proteomic datasets from the same senescent cell types, refining our lists to those markers likely to include detectable secreted or surface proteins in biological fluids.”

Classified as probably not a substantial claim.

Coefficients were forced to keep the original direction of each SenCat proteomic marker: markers linked to higher senescence could only add positively, and those linked to lower senescence could only add negatively.

“Importantly, model coefficients were constrained to preserve the original directionality of the SenCat proteomic markers, so that proteins originally associated with increased senescence could only contribute positively, whereas those originally associated with reduced senescence could only contribute negatively.”

Probably not a substantial claim: may not pass S2 (checkable).

More profiling work is essential to supplement the mainly RNA-based senescence signatures.

“strongly suggests that additional profiling efforts are critical for complementing the predominant RNA-based senescence signatures.”

Probably not a substantial claim: may not pass S3 (load-bearing).

The threshold for statistical significance was P < 0.05 unless stated otherwise.

“Unless otherwise indicated, statistical significance was defined as P < 0.05.”

May combine two claims.

SenCat is built on time-resolved in vivo snRNA-seq data, which allows senescence dynamics to be tracked across cell types and early changes to be separated from later ones, whether senescence occurs naturally or abnormally in organs.

“First, SenCat is grounded in time-resolved in vivo snRNA-seq datasets, enabling the identification of senescence dynamics across cell types and permitting the distinction of early changes in cell senescence from those occurring later on (Figure 4D), persisting both naturally and aberrantly within organs.”

Classified as probably not a substantial claim. May combine two claims.

Doublets were identified per sample using scDblFinder, and only singlets were retained for downstream analysis.

“Doublets were then identified on a per-sample basis using scDblFinder, and only singlets were retained for downstream analysis.”

Probably not a substantial claim: may not pass S3 (load-bearing).

Quality control for snRNA-seq was done per sample in a depth-aware way, rather than using one set of global cutoffs for RNA counts, detected features, or mitochondrial content.

“Rather than applying fixed global thresholds for RNA counts, detected features, or mitochondrial content, we implemented a depth-aware quality-control procedure independently for each sample.”

Classified as probably not a substantial claim. May combine two claims.

Features that every one of the 14 models kept as informative were treated as robust markers and placed in the final RNA and protein sets.

“Only features consistently retained as informative across all 14 models were considered robust markers and included in the final RNA/protein set.”

Classified as probably not a substantial claim. May combine two claims.

Preprocessing was done separately inside every training fold to avoid information leakage: imputation used training samples only, and feature scaling was learned from training data before being applied to the held-out sample.

“To prevent information leakage, all preprocessing steps were performed independently within each training fold: missing values were imputed using only the training samples, and feature scaling was also defined exclusively from the training data before being applied to the held-out sample.”

Classified as probably not a substantial claim. May combine two claims.

The matching ensured comparisons between cells with closely comparable sequencing depth and avoided biases from unequal depth distributions between groups.

“This matching ensured that comparisons were performed between cells with closely comparable sequencing depth and avoided biases introduced by unequal depth distributions between groups.”

Classified as probably not a substantial claim. May combine two claims.

The ML analysis used leave-one-cell-type-out cross-validation, with each of 14 binary classifiers holding out one cell type for testing while training on the others.

“We implemented a leave-one-cell-type-out strategy, training 14 independent binary classifiers to distinguish senescent from non-senescent samples. Each classifier excluded one cell type from training and used it for testing.”

Classified as probably not a substantial claim. May combine two claims.

Before modeling, genes that were very hard to detect were dropped, and the signature size was fixed by ranking genes on a combined measure of model coefficient strength and detection across cells.

“To further improve robustness against dropout, genes with extremely low detectability were removed before modeling, and the final signature was capped at a fixed size by ranking genes according to a combined score that incorporated both model coefficient strength and detectability across cells.”

Classified as probably not a substantial claim.

Senescent cells were called with a cluster-specific reference threshold: the cutoff for each cluster was the 99th percentile of corrected AUCell scores among reference-condition cells in that cluster, with a global reference cutoff used when a cluster had too few reference cells.

“Senescent-cell calls were then assigned using a cluster-aware reference thresholding strategy: for each cluster, the threshold was defined as the 99th percentile of corrected AUCell scores in the reference-condition cells from that same cluster; when insufficient reference cells were available in a given cluster, a global reference-based threshold was used instead.”

• AssumptionRests on an assumption: Reference-condition cells are non-senescent.

Probably not a substantial claim: may not pass S3 (load-bearing).

Reference-condition cells were downweighted during model training.

“In addition, reference-condition cells were downweighted during model training.”

Classified as probably not a substantial claim.

A gene became a valid tissue-level marker only if at least one leave-one-out run selected it and its mean coefficient kept the same sign as the original SenCat coefficient.

“Only genes selected in at least one iteration and whose average coefficient retained the same direction as the original SenCat coefficient were kept as valid tissue-level markers.”

Classified as probably not a substantial claim.

Tissue-specific leave-one-out gaussian elastic-net regression models were trained using sample age as the outcome variable and selected SenCat proteomic features as predictors.

“We then trained tissue-specific leave-one-out gaussian elastic-net regression models using sample age as the outcome variable and the selected SenCat proteomic features as predictors.”

• AssumptionRests on an assumption: Chronological age is a valid proxy for senescence in mouse tissues for bulk proteomic scoring.

Classified as probably not a substantial claim.

AUCell scores were corrected for sequencing depth using generalized additive models fitted separately within each cluster, with nCount_RNA as the sole covariate and all cells from all time points included.

“Because AUCell scores are influenced by sequencing depth, we corrected these values using generalized additive models fitted separately within each cluster, using nCount_RNA as the sole covariate and including all cells from all time points in the fit.”

Classified as probably not a substantial claim. May combine two claims.

Cross-validation was used for model fitting, and predictors with positive weights only were kept, so the final signature contained genes whose expression rose in senescent cells.

“Model fitting was performed using cross-validation, and only positively weighted predictors were retained, thereby restricting the final signature to genes associated with increased expression in senescent cells.”

Probably not a substantial claim: may not pass S3 (load-bearing).

The Benjamini-Hochberg method was used to correct for multiple testing.

“Multiple testing correction was performed using the Benjamini-Hochberg procedure.”

Classified as probably not a substantial claim.

Each sample-level comparison had to include at least a set minimum of matched senescent and non-senescent cells.

“A minimum number of matched senescent and non-senescent cells was required for each sample-level comparison.”

May combine two claims.

Senescence was verified through BrdU incorporation as a proliferation measure and increased SA-β-gal activity.

“The senescent state was confirmed by measuring BrdU incorporation to assess proliferation (Cell Signaling Technology, 6813) and increased SA-β-gal activity (Cell Signaling Technology, 9860) following the manufacturer instructions.”

Probably not a substantial claim: may not pass S3 (load-bearing). May combine two claims.

Some senescence markers might shift only briefly at early or late stages, whereas others could represent secondary adaptations instead of core senescence.

“accordingly, certain markers may only change transiently at early or late times, while others may reflect secondary adaptations rather than core senescence.”

• AssumptionRests on an assumption: Senescence is a dynamic process whose molecular features vary with the stage along the senescence trajectory.

Probably not a substantial claim: may not pass S3 (load-bearing). May combine two claims.

Mouse procedures were performed under Animal Study Proposal # 476-LGG-2027, with review and approval by the NIA, NIH Animal Care and Use Committee.

“All mouse procedures, including import, housing, experimental work, and euthanasia, were conducted under Animal Study Proposal # 476-LGG-2027, reviewed and approved by the Animal Care and Use Committee of the NIA, NIH.”

Probably not a substantial claim: may not pass S3 (load-bearing).

In bulk proteomic aging analyses, imputation and feature scaling were done inside each training fold to prevent information leakage.

“All preprocessing steps, including missing-value imputation and feature scaling, were performed within each training fold to avoid information leakage.”

May combine two claims.

BrdU incorporation and SA-β-gal activity were measured per the manufacturer's instructions and served as validation that senescence was induced across models.

“BrdU incorporation and SA-β-gal activity were quantified following the manufacturer’s instructions and were used to validate senescence induction across models.”

Probably not a substantial claim: may not pass S3 (load-bearing).

For analyses with multiple comparisons, P values were corrected by the Benjamini-Hochberg false discovery rate method.

“When multiple comparisons were performed, P values were adjusted using the Benjamini-Hochberg false discovery rate procedure.”

Classified as probably not a substantial claim.

The ML-derived transcriptomic and proteomic senescence signatures came from penalized logistic regression with leave-one-cell-type-out cross-validation.

“For machine-learning analyses, transcriptomic and proteomic senescence signatures were derived using penalized logistic regression models with leave-one-cell-type-out cross-validation.”

Probably not a substantial claim: may not pass S3 (load-bearing).

In ligand-receptor analyses, only interactions that met CellPhoneDB's statistical criteria were used for visualization and interpretation.

“Only interactions passing the statistical criteria implemented by CellPhoneDB were considered for downstream visualization and interpretation.”

Classified as probably not a substantial claim. May combine two claims.

The features kept by every model were then used to fit unregularized logistic regressions, and each marker's final weight was the average coefficient over the cross-validation models.

“Features retained across all models were used to train unregularized logistic regression models, and final marker weights were calculated as the mean coefficient across cross-validation models.”

Classified as probably not a substantial claim.

When comparing senescent and non-senescent cells in snRNA-seq data, the authors paired cells within each sample, cluster, and condition by log10-transformed nCount_RNA to reduce sequencing-depth bias.

“For comparisons between senescent and non-senescent cells in snRNA-seq datasets, cells were matched within each sample, cluster, and condition according to log10-transformed nCount_RNA to minimize sequencing-depth bias.”

Classified as probably not a substantial claim. May combine two claims.

A cell was called senescent if its corrected AUCell score exceeded a cluster-specific cutoff set at the 99th percentile of reference-condition cells; if a cluster had too few reference cells, a global reference cutoff was used.

“Senescent cells were defined using a cluster-specific threshold corresponding to the 99th percentile of corrected AUCell scores in the reference condition. When insufficient reference cells were available within a cluster, a global reference-condition threshold was used instead.”

• AssumptionRests on an assumption: Reference-condition cells are non-senescent.

Classified as probably not a substantial claim. May combine two claims.

In proteomic analyses, Benjamini-Hochberg correction was used, and proteins were kept as senescence-associated markers if they had nominal P < 0.05 and an absolute log2 fold change of at least 0.585.

“Multiple-testing correction was performed using the Benjamini-Hochberg procedure, and proteins with nominal P < 0.05 and an absolute log2 fold change ≥ 0.585 were retained as senescence-associated proteomic markers.”

Classified as probably not a substantial claim.

In snRNA-seq analyses, AUCell produced senescence scores, which were adjusted for sequencing depth by generalized additive models fitted per cluster with nCount_RNA as the covariate.

“For snRNA-seq analyses, senescence scores were calculated using AUCell and corrected for sequencing-depth effects by fitting generalized additive models within each cluster using nCount_RNA as the covariate.”