Cellular senescence, a hallmark of aging, contributes significantly to age-related diseases and is a potential therapeutic target. While molecular markers like SA-β-gal, p16, and p21 are used to identify senescent cells, these methods are often laborious and lack the high-throughput capability needed for large-scale drug screening. Senescent cells exhibit unique morphological changes, making image-based analysis a promising alternative. Convolutional neural networks (CNNs), a powerful tool in deep learning, have demonstrated exceptional accuracy in image classification tasks. This research leverages the advancements in CNNs to develop a novel, morphology-based system for high-throughput anti-senescent drug screening. The study aims to overcome limitations of existing methods by providing a quantitative, unbiased, and automated approach to identify senescent cells and assess the efficacy of potential anti-senescent drugs. The successful development and validation of such a system would significantly accelerate the discovery and development of treatments targeting age-related diseases. The increasing prevalence of age-related diseases underscores the urgency for such a methodology.

Existing methods for identifying senescent cells primarily rely on molecular markers such as SA-β-gal, p16, and p21. While effective, these methods are often time-consuming, expensive, and not well-suited for high-throughput screening. Recent advancements in deep learning, specifically CNNs, have shown remarkable success in image classification across various fields, including medical imaging. Several studies have explored the application of CNNs in biological image analysis, demonstrating the potential for automated cell identification and classification. However, there's a need for a robust system that can quantitatively assess cellular senescence based on morphology for high-throughput drug screening. The authors cite previous work using CNNs to identify cell types in images and highlight the potential of quantitative, unbiased evaluation systems in biological research. The importance of addressing cellular senescence as a therapeutic target for age-related diseases is emphasized, given the contribution of senescent cells to the inflammatory environment and disease progression. The study builds upon this foundation by developing a system capable of directly quantifying the senescence state of cells, offering a valuable tool for drug discovery.

The researchers induced cellular senescence in human umbilical vein endothelial cells (HUVECs) using three stressors: hydrogen peroxide (H2O2), camptothecin (CPT), and replicative stress. They generated a large dataset of phase-contrast microscopy images of senescent and control cells. A CNN was trained to classify these images with high accuracy (accuracy >0.9, F1 score >0.85, AUC >0.95). The robustness of the CNN was confirmed by testing it on independent datasets with different senescence induction methods and datasets acquired at a different institution (Kyoto University). To enable quantitative evaluation, a senescence score was developed based on the probability output from the CNN. This score, termed Deep Learning-Based Senescence Scoring System by Morphology (Deep-SeSMo), strongly correlated with the degree of senescence induction (Pearson correlation coefficient >0.9). Deep-SeSMo's effectiveness was validated using known anti-senescent drugs (NMN and metformin). A kinase inhibitor library was screened using Deep-SeSMo, identifying four potential anti-senescent compounds (terreic acid, PD-98059, daidzein, and Y-27632-2HCl). The anti-senescent effects of these compounds were confirmed using traditional methods (SA-β-gal staining and Western blotting). Finally, RNA sequencing was performed to investigate the underlying mechanisms of action, revealing that these compounds commonly suppress senescent phenotypes by inhibiting the inflammatory response pathway, particularly NF-κB signaling. Detailed protocols for cell culture, microscopy imaging, RNA isolation, RT-PCR, SA-β-gal staining, Western blotting, CNN training, senescence scoring, and RNA sequencing are provided in the methods section. The CNN architecture, training parameters, and data augmentation techniques are described thoroughly. The specific computational resources used are also noted.

The CNN developed in this study exhibited high accuracy in classifying senescent and control cells, achieving accuracy, F1 score, and AUC values consistently above 0.9, 0.85, and 0.95, respectively, across various senescence induction methods and datasets from different institutions. The CNN's generalizability was demonstrated by its consistent performance across different senescence induction methods and datasets from different laboratories. The Deep-SeSMo system, based on the CNN, accurately quantified the degree of senescence, exhibiting a high Pearson correlation coefficient (>0.9) with the strength of senescence-inducing stressors. The Deep-SeSMo-based drug screening successfully identified four novel anti-senescent compounds: terreic acid, PD-98059, daidzein, and Y-27632-2HCl. The anti-senescent effects of these four compounds were further validated using conventional methods such as SA-β-gal staining and Western blotting, showing significant reductions in senescence markers. RNA sequencing analysis revealed that these four compounds commonly suppressed senescence by inhibiting the inflammatory response pathway, with a significant negative enrichment in genes related to the inflammatory response and NF-κB signaling. Terreic acid showed unique upregulation of genes related to the positive regulation of ATPase activity in the mitochondria, suggesting a potential mechanism for its anti-senescent effects.

This study successfully demonstrates the use of a deep learning-based morphology scoring system, Deep-SeSMo, for high-throughput anti-senescent drug screening. The findings address the research question by providing a novel, accurate, and quantitative approach for identifying senescent cells and assessing the efficacy of drug candidates. The results highlight the potential of Deep-SeSMo as a powerful tool for drug discovery, particularly in the field of aging research. The identification of four novel anti-senescent compounds with a shared mechanism of action in inhibiting the inflammatory response pathway provides valuable insights into the therapeutic targets for age-related diseases. The study’s contribution to the field lies in the development of a robust, unbiased, and automated system for drug screening, offering a significant advance over traditional methods. The unique upregulation of genes related to mitochondrial function observed with terreic acid warrants further investigation into its mechanism of action.

This research established a novel, high-throughput drug screening platform, Deep-SeSMo, for identifying anti-senescent compounds based on deep learning-driven analysis of cellular morphology. Deep-SeSMo efficiently screened a kinase inhibitor library and identified four compounds with anti-senescent and anti-inflammatory properties. These findings demonstrate the potential of Deep-SeSMo as a powerful tool for drug discovery in age-related diseases and other fields requiring high-throughput screening of cellular morphology. Future research could focus on validating the in vivo efficacy of these compounds and further elucidating their mechanisms of action, particularly the unique effects observed with terreic acid on mitochondrial function.

The study primarily focuses on in vitro experiments using HUVECs and HDFs. The findings may not be directly generalizable to other cell types or in vivo settings. Further research is needed to confirm the in vivo efficacy and safety of the identified compounds. While Deep-SeSMo demonstrates high accuracy, there is a small percentage of misclassifications, which could potentially influence the overall senescence score. The precise mechanism by which terreic acid exerts its anti-senescent effects remains to be fully elucidated.