Senescence, a cellular response limiting the replication of old, damaged, and cancerous cells, is characterized by stable cell cycle arrest, a bioactive secretome (SASP), and various phenotypic changes including metabolic reprogramming, altered morphology, increased lysosomal mass, chromatin rearrangement, and nuclear changes. Senescent cells accumulate with age and are implicated in cancer and fibrosis. Their contribution to aging and disease progression has spurred interest in senolytics—drugs selectively eliminating senescent cells. Reliable identification of senescent cells is crucial for successful senolytic clinical trials and a deeper understanding of senescence biology. Current markers, such as SA-β-Gal and p16INK4a, have limitations: SA-β-Gal stains non-senescent cells like macrophages and is challenging to detect in vivo; p16INK4a is often deleted in cancer cells and difficult to detect in mouse tissue. The heterogeneity of senescence necessitates multiple markers, highlighting the need for alternative methods. Imaging-based approaches have shown promise, but their widespread applicability remains uncertain. This study leverages nuclear features to develop machine-learning algorithms for identifying senescence, aiming for broader accessibility, applicability across senescent cell types, and utility in diverse contexts.

The literature extensively documents the role of senescence in aging and age-related diseases. Studies have shown the accumulation of senescent cells in various tissues with age, and their association with age-related pathologies (10, 11). Senolytics, drugs that selectively eliminate senescent cells, are a promising therapeutic approach targeting these pathologies (12). Several markers have been used to identify senescent cells, including SA-β-galactosidase (2) and p16INK4a (1, 16). However, limitations associated with these markers have led researchers to explore alternative approaches (17). Recent studies have proposed image-based classifiers leveraging deep learning (18, 19, 20) to identify senescent cells. While showing promise, the accessibility and generalizability of these methods require further investigation. This research builds upon this prior work, proposing a more accessible and broadly applicable approach.

This study employed a multi-faceted approach to develop and validate machine-learning classifiers for senescence detection. Initially, senescence was induced in A549 human lung adenocarcinoma cells using etoposide, a topoisomerase II inhibitor. Senescence was confirmed by SA-β-Gal staining, BrdU incorporation, and immunofluorescence for DNA damage markers (γH2AX), p53, and p21CIP1. High-throughput automated microscopy and image analysis software (IN Carta and CellProfiler) were used to extract nuclear features (area, gyration radius, compactness, etc.) from DAPI-stained nuclei. These features were then used to train random forest and classification tree-based machine-learning algorithms (AEM, AERFM, AEMCP, AERFMCP) to predict senescence. The classifiers were validated using test datasets and assessed using precision-recall and ROC curves. The researchers also investigated the classifiers' ability to distinguish senescent cells from quiescent cells and those undergoing DNA damage. To evaluate performance, co-cultures of different ratios of senescent and non-senescent cells were analyzed, and additional classifiers were trained using marker-identified senescent cells (BAEM, BPEM, PPEM). To assess the classifiers' generalizability, they were applied to other cell types (SK-HEP-1, SK-MEL-103, MCF7, HCT116, and IMR90) and senescence inducers (doxorubicin and barasertib). Finally, a high-throughput drug screen was performed using a library of 676 drugs to identify those selectively inducing senescence in A549 cells, and a tissue senescence score (TSS) was developed and validated in mouse models (liver cancer initiation, aging, and fibrosis) and human liver samples from patients with NAFLD. Immunohistochemistry, immunofluorescence, and image analysis software (QuPath) were used in the in vivo studies.

The key findings demonstrate the successful development and validation of a family of machine-learning algorithms capable of accurately predicting cellular senescence based on nuclear morphology. 1. **Accurate Senescence Prediction:** The classifiers (AEM, AERFM, AEMCP, AERFMCP, BAEM, BPEM, PPEM, GM, etc.) accurately predicted senescence in various cell types (A549, SK-HEP-1, SK-MEL-103, MCF7, HCT116, IMR90) and upon induction by different stressors (etoposide, doxorubicin, barasertib). The algorithms showed high correlation with SA-β-galactosidase activity and other senescence markers (p21CIP1, p53, BrdU). Precision, accuracy, recall, and F1 scores supported the classifiers' robust performance. The general model (GM) classifier, trained on multiple senescence models, consistently performed well across diverse datasets. 2. **Distinguishing Senescence from Quiescence and DNA Damage:** The classifiers successfully distinguished senescent cells from quiescent cells and those undergoing DNA damage, highlighting their specificity for senescence-related nuclear changes. The results showed that nuclear changes detected by the algorithm were not merely indicative of a DNA damage response, but rather a combination of features specific to senescent cells. 3. **Senolytic Drug Characterization:** The classifiers were used to characterize the effects of senolytic drugs (ABT-263, ABT-737). In co-culture assays, these drugs selectively reduced senescent cell populations, which was accurately reflected by the classifiers' predictions. 4. **Senescence-Inducing Drug Discovery:** A high-throughput screen of 676 drugs identified compounds (including aurora kinase inhibitors and Eg5 inhibitors) preferentially inducing senescence in cancer cells (A549) over normal fibroblasts (IMR90). This demonstrates the utility of the classifiers in identifying potential therapeutic agents. 5. **Tissue Senescence Score (TSS):** A TSS was developed to assess senescence in tissue sections. This score showed strong correlation with the percentage of senescent cells (p21CIP1-positive cells) in mouse models of liver cancer initiation, aging, and fibrosis. Importantly, the TSS correlated with senolytic drug efficacy and identified senescence in human liver samples from patients with mild NAFLD. The TSS performed comparably to p16INK4a staining in human samples.

This study significantly advances the field by providing a robust and accessible method for detecting cellular senescence. The use of readily available image analysis software and easily interpretable nuclear features enhances the accessibility and broader adoption of these classifiers. The accuracy and generalizability of the classifiers across various cell types and senescence inducers validate their utility in diverse research settings. The successful application of the classifiers to characterize senolytics, identify senescence-inducing drugs, and develop a tissue-based senescence score demonstrates their wide-ranging applicability. The study’s findings address the critical need for reliable senescence detection methods, opening new avenues for investigating the role of senescence in aging and disease. The TSS, particularly, provides a valuable tool for preclinical and clinical studies evaluating senotherapies and assessing disease progression in various conditions involving senescent cell accumulation.

This research successfully developed and validated a family of machine-learning algorithms for identifying cellular senescence based on nuclear features. These classifiers accurately predict senescence across diverse cell types and senescence inducers, distinguish senescence from other cellular states, and enable the characterization of senolytics and the discovery of novel senescence-inducing drugs. The development of a tissue senescence score further extends the utility of this approach to in vivo studies, including preclinical models and human patient samples. Future research could focus on refining the TSS for improved accuracy and wider tissue applicability, exploring the mechanisms underlying drug selectivity identified in the screen, and applying these classifiers to larger clinical cohorts to further validate their clinical utility.

While the classifiers demonstrated robust performance across various cell types and senescence inducers, some limitations exist. The classifiers' performance might vary based on the type of senescence, potentially being less effective in identifying senescence induced by specific mechanisms or in certain cell types. Further improvements to the TSS might be needed to optimize its accuracy and ensure consistency across different tissues and staining protocols. The study used a relatively small sample size in some in vivo experiments, although the results show a significant correlation. The generalizability to other tissues and species remains to be fully explored.