Date: 13th January 2021
Cellular senescence is a hallmark of ageing and is important for the pathogenesis of ageing-related disorders and diseases such as cancer. It is a state in which cells can no longer divide and is a potential therapeutic target. Now, researchers have developed a morphology-based deep learning system that can successfully classify cellular senescence and evaluate the effects of anti-senescent agents.
Endothelial cells serve many functions in homoeostasis and disease, and senescence in endothelial cells often appears to be an initial step in the cascade of events that lead to the development of age-related pathologies such as cardiovascular disease. Whilst, specific biological markers can be used for screening cellular senescence, morphology can also be diagnostic – for example the cells have flat and enlarged cell bodies and heterochromatin aggregations. However, quantitative assessment of morphology changes are subject to bias and therefore, up until now, large scale drug screens based on morphological changes have been limited.
Now, researchers from Keio University School of Medicine, Japan, led by Shinsuke Yuasa, have developed a robust, morphology-based deep learning system to identify senescent cells. The addition of a non-bias quantitative scoring system, or Deep Learning-Based Senescence Scoring System by Morphology (Deep-SeSMo), was used to screen for drugs that control cellular senescence and identified four candidates which showed anti-senescent and anti-inflammatory effects.
In deep learning, convolutional neural networks (CNN) are a class of deep neural networks, which are particularly well suited for image analysis, and are becoming increasingly employed as a diagnostic tool in the clinic. Using a CNN, the team here successfully trained the network to identify senescent cells in image datasets from different senescence induction methods, which were separate from those used for training purposes. Furthermore, the CNN exhibited a high performance in datasets from different universities, and in different cells types which suggested that cellular senescence shows a unique morphologic characteristic, and that a morphology-based CNN system could reliably identify senescent cells.
However, whilst this algorithm could detect senescent from control cells, the output was a non-linear prediction with only two values, and no intermediate states. The team therefore, wanted to add complexity to the prediction, so they developed Deep-SeSMo, which could calculate a senescence score for each phase-contrast image, rapidly – in only 0.08–0.1 ms.
To test Deep-SeSMo, the team first examined the effects of well-known anti-cellular senescence reagents, and Deep-SeSMo was able to correctly assess the effects of these different compounds.
However, the team wanted to expand its use further for drug discovery, so they used Deep-SeSMo to screen for drugs that controlled cellular senescence using a kinase inhibitor library. Senescence scores were calculated by Deep-SeSMo and normalised to a control sample, which then allowed the drugs to be ranked for senescence suppression.
The top four candidates identified by the algorithm, were then reinvestigated using conventional methods and all showed a decreased cellular senescence activity and could also supress inflammatory phenotypes. Gene ontology (GO) analysis showed that the top compound, terreic acid, showed particular promise as a potential drug against senescence and age-related disease.
The team here have developed a drug-screening system for cellular senescence using a deep learning CNN. Deep-SeSMo could correctly evaluate the effects of well-known anti-senescent reagents, and identified four potential anti-senescent drugs in a kinase inhibitor drug screen.
As cellular senescence has a pivotal role in age-related diseases such as diabetes, heart failure, atherosclerosis, and cancer, such non-biased methods could be valuable tools to identify morphological differences in research and drug screening for these types of diseases. Furthermore, as morphological changes are often associated with other diseases Deep-SeSMo could be adapted to recognise, score and rank other disease-treating drug targets.
We are starting to see artificial intelligence applications reach many parts of our research and translate into clinical tools. Their use in drug screening, repurposing and image analysis shows particular promise. We have recently seen researchers develop a machine learning model that accurately predicts how combinations of different cancer drugs kill various types of cancer cells, the use of deep learning models for the identification of COVID-19 pneumonia on high resolutions computed tomography (CT) scans or to detect very early sign of breast cancer. The field is moving quickly, and is offering us many methods to save valuable time, provide unbiased analysis and perform large-scale screens. It is poised to transform healthcare – allowing us to provide better treatments, diagnostics and develop novel therapies.
Kusumoto, D., T. Seki, H. Sawada, A. Kunitomi, T. Katsuki, M. Kimura, S. Ito, J. Komuro, H. Hashimoto, K. Fukuda and S. Yuasa (2021). “Anti-senescent drug screening by deep learning-based morphology senescence scoring.” Nature Communications 12(1): 257.