Anti-senescent drug screening by deep learning-based morphology senescence scoring

The study addresses whether cellular senescence—a key mechanism in ageing and age-related diseases—can be identified and quantitatively evaluated from label-free cell morphology using convolutional neural networks, and whether such a system can be leveraged for anti-senescent drug discovery. Prior approaches rely on molecular markers (SA-β-gal, P16, P21), but senescent cells also exhibit distinctive morphology (flat, enlarged cells, heterochromatin aggregation). Conventional methods struggle to quantify morphology-based senescence across large cell populations. The authors propose a robust CNN-based classifier for senescent endothelial cells and introduce a quantitative senescence score derived from the CNN’s output probabilities (Deep-SeSMo) to enable unbiased high-throughput evaluation and drug screening.

CNNs have rapidly improved image classification performance and are widely used in medical and biological imaging. Prior work includes CNN-based diagnostic tools in clinics and label-free identification of endothelial cells from phase-contrast images. While CNNs excel at qualitative classification, versatile biological systems benefit from quantitative measures. Cellular senescence is implicated in age-related diseases and is commonly assessed via molecular markers (SA-β-gal, P16, P21) and characteristic morphology, but unbiased, scalable quantification based on morphology is lacking. The study builds on this context to develop a quantitative morphology-based approach.

- Cell models: Human umbilical vein endothelial cells (HUVECs) and human diploid fibroblasts (HDFs, TIG-114). Senescence induced by three stressors: hydrogen peroxide (H₂O₂), camptothecin (CPT), and replication (serial passages). For HUVECs training: 0.25 mM H₂O₂ for 4 days or 100 nM CPT for 2 days; replication >10 passages. Controls matched durations with low passage (<3). For quantitative scoring: HUVECs treated with graded H₂O₂ (0–0.25 mM), CPT (0–100 nM), or passages (1–12). HDFs treated with H₂O₂ (0–0.025 mM) or CPT (0–100 nM). - Imaging and preprocessing: Phase-contrast microscopy images (2776×2074 px RGB TIFF). Automated single-cell cropping using OpenCV: binarization, size thresholding to remove noise, center-of-mass detection, extraction of 50×50 px patches per cell; conversion to numpy arrays. - Datasets: For HUVEC training, multiple inductions with 10 images each; example dataset sizes: H₂O₂-induced senescence 92,242 patches, H₂O₂ control 41,207; CPT-induced senescence 134,097, CPT control 64,535. Additional datasets acquired at Kyoto University for external validation. HDF datasets prepared similarly. - CNN architecture and training: 4 convolutional layers with ReLU activations, 2 max-pooling layers, 2 fully connected layers; softmax output (senescent vs control). Dropout (rate 0.5) after first and second max pooling and first dense layer. Input normalization Y = ((X/255) − 0.5) × 2. Data augmentation (rotation, width/height shifts, horizontal/vertical flips). Optimization: stochastic gradient descent (learning rate 0.032), cross-entropy loss; Glorot uniform weight initialization. Trained three models: (i) control vs H₂O₂-senescence, (ii) control vs CPT-senescence, (iii) control vs mixed H₂O₂+CPT senescence. Performance metrics: accuracy, precision, recall, F1 score, ROC-AUC; Grad-CAM for interpretability. - Traditional ML baseline: HOG features (vector length 2916) with logistic regression, random forest, and linear SVM; evaluated by accuracy, F1, ROC-AUC. - Generalisability tests: External test sets with senescence induced by H₂O₂, CPT, or replication; evaluated by the three pre-trained CNNs. Cross-institution validation using datasets from Kyoto University and Keio University (mixed training vs single-institution). Cross-cell-type generalisation tested on HDFs. - Deep-SeSMo scoring: Uses the softmax senescence probability per cell; senescence score = average probability across all cells in a sample. Correlation with stress strength (H₂O₂ or CPT concentration, passage number) quantified by Pearson correlation. Inference speed measured (0.08–0.1 ms per image). Models trained on combined-institute data and on HDFs were also used to compute scores. - Validation with known modulators: Metformin (0–1000 μM) and nicotinamide mononucleotide (NMN, 0–1000 μM) co-treated during senescence induction; assessed by SA-β-gal staining, Western blotting (P21, P53, p-P53 Ser15, P16INK4a), and Deep-SeSMo scores. Senolytic assay: mixture of young and old HUVECs treated with ABT263 (0.25–2.5 μM) analyzed by Deep-SeSMo. - Drug screening: 80-compound kinase inhibitor library (10 μM) screened under three senescence inductions (H₂O₂, CPT, replication). Senescence scores computed by three CNNs and normalized; converted to rankings per evaluation; compounds prioritized by median ranking across evaluations. Heatmap and surface plots used to identify anti-senescence clusters. - RNA sequencing: Moderately senescent HUVECs treated for 4 days with terreic acid (10 nM), Y-27632-2HCl (500 nM), daidzein (5 μM), or PD-98059 (100 nM). Library prep (NEBNext Ultra II), Illumina HiSeq 2×150 bp; QC (FastQC), trimming (Trimmomatic), alignment/quantification (HISAT2-StringTie-Ballgown). Differential expression, GSEA (MSigDB gene sets), and GO analysis (DAVID). Data deposited in DDBJ SRA (DRA010959). - Compute environment: NVIDIA GTX1080Ti GPUs; Ubuntu 16.04, CUDA 8.0, cuDNN 6.0; Python 3.5, TensorFlow 1.4.0, Keras 2.1.2. Code available on GitHub (https://github.com/Dai-Kusumoto/Deep-SeSMo).

- CNN classification of senescent vs control HUVECs achieved strong performance. With mixed H₂O₂+CPT training: accuracy 0.93, F1 score 0.88, ROC-AUC 0.98. Traditional ML (HOG + SVM/RF/LR) underperformed relative to CNN. - Generalisability: Pre-trained CNNs (trained on H₂O₂, CPT, or mixed) accurately classified senescence across new datasets induced by H₂O₂, CPT, or replication; average accuracy >0.9, F1 >0.85, ROC-AUC >0.95 (often ~1.00). Models trained on combined Keio+Kyoto data generalized better across institutions; HUVEC-trained CNNs also classified senescent vs healthy HDFs. - Interpretability: Grad-CAM indicated healthy-cell predictions relied on peripheral regions, whereas senescent-cell predictions focused on heterogeneous intracellular regions. - Deep-SeSMo quantitative scoring: Although single-cell probabilities tended to be near 0 or 1, the average senescence score strongly correlated with senescence strength across gradients of H₂O₂, CPT, and passage number. Pearson correlations typically >0.9 across networks and stressors; networks trained on both H₂O₂ and CPT showed correlation coefficients >0.9 including replication stress. Inference speed per image was 0.08–0.1 ms. - Validation with known modulators: Metformin and NMN reduced SA-β-gal positivity, decreased P21, p-P53 (Ser15), P53, and P16INK4a activation levels, and lowered Deep-SeSMo senescence scores in a dose-responsive manner. Deep-SeSMo also detected senolytic effects of ABT263 in mixed young/old HUVEC populations. - Drug screening (80 kinase inhibitors) identified four anti-senescent compounds by median ranking across conditions: terreic acid, PD-98059, daidzein, and Y-27632-2HCl. These formed an anti-senescence cluster in ranking heatmaps. - Experimental validation of top hits: All four compounds reduced SA-β-gal-positive cell fractions and suppressed P53–P21 axis activation and P16INK4a expression across induction methods. Four drugs deemed non-effective by Deep-SeSMo (SC-514, TYRPHOSTIN51, Indirubin, SU4312) showed minimal effects on P53–P21 signaling. - Transcriptomics: RNA-seq revealed common downregulation of inflammatory response and NFκB signaling gene sets across the four compounds (negative enrichment in GSEA). Among top differentially expressed genes, NFκB activator TBL1XR1 was downregulated, while NFκB inhibitors SIGIRR and ASCC1 were upregulated. Terreic acid uniquely upregulated genes related to positive regulation of ATPase activity and oxidative phosphorylation, suggesting maintenance of mitochondrial function under stress. BTK expression was minimal in HUVECs, implying terreic acid acts via BTK-independent mechanisms.

The study demonstrates that label-free morphology encodes sufficient information for accurate identification of cellular senescence and that a CNN can not only classify senescent cells but also yield a quantitative senescence score (Deep-SeSMo) by averaging output probabilities. This quantitative index correlates tightly with senescence severity induced by oxidative, DNA-damaging, or replicative stress, enabling robust, unbiased assessment at scale. Application to pharmacology showed Deep-SeSMo captures known anti-senescent and senolytic effects and can screen compound libraries to discover anti-senescent candidates. The identified compounds converge on suppression of inflammatory pathways (notably NFκB/SASP), aligning with senescence biology, and terreic acid may additionally ameliorate mitochondrial dysfunction. These findings support morphology-based deep learning as a powerful platform for senescence phenotyping and drug discovery, with potential applicability across cell types and institutions.

A rapid, accurate CNN-based system was developed to identify senescent cells from phase-contrast images and to compute a quantitative senescence score (Deep-SeSMo). The score correlates with senescence-inducing stress strength and enables high-throughput, non-biased drug screening. Screening of 80 kinase inhibitors identified four anti-senescent compounds—terreic acid, PD-98059, daidzein, and Y-27632-2HCl—which suppress senescence and inflammatory signatures, with terreic acid additionally linked to improved mitochondrial pathways. This approach establishes a generalizable framework for morphology-driven phenotyping and drug discovery. Future work should elucidate the precise mechanisms (especially for terreic acid), extend validation to in vivo models and additional cell types, and explore broader disease contexts where morphology-based CNN scoring can accelerate therapeutics.

- The CNN, while highly accurate, still exhibited rare misclassifications; the biological basis of a small fraction of apparent senescent cells in control conditions (and vice versa) remains unclear. - The quantitative score arises from a binary classifier’s probabilities; the observed digital-like transition of cells between states and the mechanistic underpinnings of varying senescence thresholds among cells are not yet explained. - Mechanism of terreic acid’s anti-senescent effect in endothelial cells is unresolved (BTK is minimally expressed); further proteomic/metabolomic studies and in vivo validation are needed. - Most results are from in vitro endothelial and fibroblast models; generalizability to tissues and organisms requires additional validation.