Detection of senescence using machine learning algorithms based on nuclear features

Senescence is a cellular program that limits replication of old, damaged, or cancerous cells and is characterized by stable cell cycle arrest, a bioactive secretome (SASP), metabolic reprogramming, enlarged and flattened morphology, increased lysosomal mass, chromatin rearrangements, and nuclear changes. Senescent cells accumulate with aging and in cancer and fibrotic lesions, contributing to disease. Senotherapies, including senolytics, aim to eliminate or modulate senescent cells, but progress is hampered by the lack of universal robust markers. Common markers such as SA-β-galactosidase and p16INK4a have important limitations and can yield false positives or be technically difficult in vivo. Recent imaging-based approaches show promise but their generalizability and ease-of-use remain unclear. This study asks whether interpretable nuclear morphology features, extracted with accessible software, can train machine-learning classifiers to reliably detect senescence across stressors, cell types, and tissues, and whether such classifiers can aid the characterization of senotherapies and detection of senescence in vivo and in patient samples.

The paper reviews limitations of established senescence markers: SA-β-gal activity reflects lysosomal mass but is not specific (e.g., macrophages stain positive) and is constrained to cryosections in vivo; p16INK4a is variable, often deleted in cancer, and difficult to detect in mouse tissues. Given the heterogeneity of senescence, no universal marker exists and multiple markers are typically needed. Imaging-based senescence classifiers have been reported, including deep learning approaches using nuclear morphology, but questions remain about portability to other labs, applicability across cell types and contexts, and computational demands. These gaps motivate an approach based on a small set of interpretable nuclear features measurable by open-source tools.

In vitro senescence induction and imaging: Human cell lines (e.g., A549, SK-HEP-1, SK-MEL-103, MCF7, HCT116) and primary fibroblasts (IMR90) were treated with senescence inducers (etoposide, doxorubicin, alisertib, barasertib; dosing per cell line) for 7 days; quiescence was induced by 0.5% FBS. Ionizing radiation (15 Gy) induced DNA damage. Senescence markers included SA-β-gal activity (X-gal and C12FDG), BrdU incorporation, p53, p21CIP1, γH2AX, 53BP1 foci. High-content microscopy (IN Cell Analyzer 2500HS) captured DAPI-stained nuclei. Nuclear features were extracted using IN Carta (Area, Form Factor, Elongation, Compactness, Chord Ratio, Gyration Radius, Displacement) and, in a parallel pipeline, CellProfiler (17 morphology features, including area, eccentricity, Feret diameters, perimeter, solidity, etc.). Quality control excluded artifacts (intensity thresholds, size filters, edge-touching objects). Classifier development: Initial classifiers were trained on A549 cells treated with etoposide versus DMSO controls, assuming all treated cells were senescent and all DMSO cells non-senescent. Two algorithms were built: AEM (classification tree) and AERFM (random forest). Performance was evaluated by ROC/AUC, precision, recall, accuracy, and F1, and validated on independent test sets. Classifiers were also trained using CellProfiler-derived features (AEMCP, AERFMCP). To reduce label noise, marker-based classifiers were trained using marker-defined cells: BAEM (SA-β-gal positive vs negative), BPEM (BrdU−/p21CIP1+), and PPEM (p53+/p21+). Generalization across models: Additional classifiers were generated per cell type/inducer and a General Model (GM) trained on pooled data from multiple cell lines (A549, SK-MEL-103, SK-HEP-1) and inducers (etoposide, doxorubicin, alisertib, barasertib). A voting consensus algorithm (VCA) aggregated predictions from eight base classifiers by equal-weight voting. Co-culture assays mixed varying ratios of senescent and non-senescent cells. Simultaneous C12FDG and DAPI staining enabled ground-truth senescence by SA-β-gal and feature extraction from the same cells. Senolytic characterization and screening: GFP-labeled senescent A549 (etoposide) were co-cultured with mCherry-labeled non-senescent A549 and treated with senolytics (ABT-263, ABT-737). Changes in GFP/mCherry counts and classifier-predicted senescence quantified selective killing. A screen of 676 compounds (10 μM; Target Selective and Protein Kinase Inhibitor Library II) was performed in parallel in A549 and IMR90 (triplicates). Toxic compounds (≤40% of control cell count) were excluded. Senescence percentages were predicted by the GM classifier and normalized via B-score (CellHTS2). Hits required B-score >15 in at least 2/3 replicates and were categorized by selectivity. In vivo tissue classifier and scoring: Mouse models included liver cancer initiation via hydrodynamic tail vein injection of transposons expressing NRasG12V (senescence-inducing) or NRasG12V,D38A (inactive), senolysis (NRasG12V plus senolytic vs vehicle), liver fibrosis (CCl4 vs oil), and aging (young ~90 days vs old ~600 days). Serial sections underwent IHC/IF for p21CIP1, uPAR, ORF1, GFP, and hematoxylin (or H&E). Whole-slide imaging was analyzed in QuPath (nuclear detection and features: area, perimeter, circularity, max/min caliper, eccentricity; and DAB intensities for IHC). A cell senescence score (CSS) was defined by comparing each nucleus’s features to idealized normal (mean of p21− nuclei) and idealized senescent (top-intensity p21+ nuclei) profiles. Because single-cell CSS consistency varied, a tissue senescence score (TSS) was defined as the percentage of cells with CSS between 1 and 5, which correlated with senescent marker prevalence. TSS was computed on hematoxylin- or p21-stained sections and cross-validated against marker frequencies. Human samples: Liver resections from patients with mild NAFLD were stained with H&E and IHC for p16INK4a. QuPath-derived nuclear features from H&E slides (with circularity threshold >0.7 to focus on hepatocytes) were used to compute CSS and TSS, and correlated with p16INK4a positivity. Statistical analyses included t-tests, one-way ANOVA with Tukey’s tests, Pearson correlations, and reporting of p-values; performance metrics (precision, accuracy, recall, F1) were computed on co-culture datasets.

- Nuclear morphology changes in senescent A549 cells: Compared to DMSO controls, etoposide-induced senescent cells had significantly larger area and altered gyration radius, compactness, chord ratio, displacement, and elongation (Kolmogorov-Smirnov tests, all p < 0.0001 except form factor). - Classifier accuracy in A549: AEM and AERFM predicted senescence at levels comparable to SA-β-gal staining in both training and independent test datasets, with high specificity (validated by ROC/PR analyses). - Specificity vs quiescence and confluency: Classifiers correctly identified senescent cells but not quiescent (0.5% FBS) arrested cells; predictions were robust across cell seeding densities. - DNA damage vs senescence: The AEM classifier identified most etoposide-induced senescent cells; it labeled <30% of irradiated (DNA damage) cells as senescent, and identified MLN8054-induced senescence despite minimal DNA damage, indicating detection of senescence-related nuclear features beyond generic DNA damage. - Strong correlation with markers in co-cultures: AEM predictions correlated with SA-β-gal (r = 0.8745, p < 0.0001), p21+/BrdU− (r = 0.9478, p < 0.0001), and p21+/p53+ (r = 0.8681, p < 0.0001). Marker-trained BAEM achieved r = 0.9969 (p < 0.0001) vs SA-β-gal. - Performance metrics comparing assumption- and marker-based models (median across 70–96 wells): AEM Precision 0.90, Accuracy 0.89, Recall 0.94, F1 0.89; BAEM 0.82/0.80/0.90/0.83; BPEM 0.84/0.83/0.90/0.84; PPEM 0.81/0.77/0.88/0.82. - Generalization across cell types/inducers: Nine classifiers (including GM, AEM, IEM, random forests, and VCA) showed high correlations between predictions and SA-β-gal across multiple co-cultures (A549 with doxorubicin or barasertib; SK-HEP-1, SK-MEL-103, MCF7, HCT116, IMR90 with etoposide). GM and VCA were among the most consistent across datasets; AERFM and VCA had the best recall. Some conditions (e.g., barasertib-treated A549) showed lower but still significant correlations. - Senolytics assessment: In A549 co-cultures of GFP-senescent and mCherry-control cells, ABT-263 and ABT-737 selectively reduced senescent cells by both GFP counts and AEM predictions (e.g., AEM predicted senescent cells decreased significantly vs DMSO at day 10; p = 0.0006–0.0007). - Drug screen (676 compounds at 10 μM): After excluding 69 toxic compounds, 607 were analyzed; 56 drugs met the senescence induction criteria (B-score >15 in ≥2/3 replicates). Selectivity: 27 induced senescence only in A549 (e.g., MLN8054), 11 only in IMR90 (e.g., AG-14361), and 18 in both. - Validation of selective inducers: MLN8054, ZM447439, ARQ621, and SC144 induced higher SA-β-gal positivity in A549 than IMR90; doxorubicin and niraparib induced senescence in both; AG-14361 was selective for IMR90. MLN8054 and ARQ621 induced arrest, p53/p21CIP1, and SASP in A549 but not in IMR90. Pre-treatment with MLN8054 or ARQ621 sensitized >75% of A549 to ABT-263 vs <50% of IMR90. - Tissue Senescence Score (TSS) in vivo: In NRasG12V liver cancer initiation, TSS from hematoxylin sections was higher than in NRasG12V,D38A controls, correlated with p21CIP1 positivity, and correlated between hematoxylin- and p21-based TSS (r = 0.7177, p = 0.0026). TSS tracked senolytic efficacy (senolytic vs vehicle: TSS lower, p = 0.0051; p21CIP1 positivity reduced, p = 0.0138). - Fibrosis and aging: CCl4-treated mice showed higher p21CIP1 positivity (p = 0.0195) and TSS (p = 0.0095) vs oil controls. Old mice (~600 days) had higher p21CIP1 positivity (p = 0.0427) and TSS (p = 0.0010) vs young (~90 days). - Human NAFLD: In 34 patients, TSS from H&E sections correlated with p16INK4a-positive cell percentage (r = 0.3862, p = 0.024).

The study demonstrates that senescence-associated nuclear morphology, quantified by a compact set of interpretable features, enables accurate machine-learning classification of senescence across diverse cell types and inducers. These classifiers distinguish senescence from quiescence and are robust to confounding factors like confluency, indicating they capture morphology changes linked to the senescent state rather than generic cell cycle arrest. While DNA damage can partially overlap with senescence signatures, classifier performance with MLN8054-induced senescence and limited labeling of irradiated cells indicate specificity for senescence-relevant nuclear changes. A family of models, including a pooled General Model and a voting consensus, perform well across datasets, supporting portability. Importantly, classifier predictions have sufficient recall to power practical applications even when precision is not perfect, enabling drug discovery and senolytic assessment. The approach extends to tissue by using a tissue senescence score that correlates with established senescence markers and captures senescence dynamics during oncogene-induced liver tumor initiation, after senolytic treatment, in fibrosis, and during aging. Applicability to human NAFLD further supports translational relevance. Compared with deep learning image-based classifiers, this feature-based strategy reduces computational demands, enhances interpretability, and leverages open-source tools (CellProfiler, QuPath), facilitating adoption. Given senescence heterogeneity, no single universal predictor is expected; nevertheless, this toolbox offers broadly useful, adaptable classifiers for detecting pathophysiological senescence and evaluating senotherapies.

The authors present a practical, interpretable machine-learning framework using nuclear morphology features to detect senescence at single-cell and tissue levels. Classifiers generalize across cell types and inducers, distinguish senescence from quiescence, and support applications such as senolytic characterization and high-throughput screening. A tissue senescence score extends detection to mouse models of cancer initiation, fibrosis, and aging, and correlates with senescence in human NAFLD. This work provides accessible pipelines compatible with open-source software, offering a scalable route to quantify senescence in research and translational contexts. Future work could expand training to additional cell types, stressors, and tissues, refine tissue scoring for broader pathology, and integrate multimodal markers to further improve specificity and generalizability.

Senescence is heterogeneous and context-dependent; nuclear morphology changes may be subtle or absent in certain models (e.g., constitutively active MEK-induced senescence), limiting classifier performance. Classifiers trained on specific inducers or cell types may perform variably across conditions (e.g., lower correlations in some barasertib or IMR90 settings). Ground-truth markers (e.g., SA-β-gal, p16INK4a, p21CIP1) are themselves imperfect, influencing training and evaluation. Some overlap with DNA damage responses can yield false positives. Tissue scoring showed variability at single-cell level (necessitating TSS aggregation), and adaptation may be needed for other tissues and disease contexts. Precision may not always be optimal, though recall is high for screening applications.