
A long-awaited AI approach can now be implemented in screening for anti-senescence drug candidates.
Developing a novel therapeutic drug can cost between $161 million and $2 billion. It can take 12 long years from lead identification and through clinical trials. Even so, only 5 in 5000 drugs make it to the market [1].
Longevity.Technology: Artificial intelligence (AI) has been booming within the pharmaceutical sector over the past five years. In fact, Business Insider has found that AI can curb up to 70% of drug development costs [2]. While pharma has taken a long time to adopt a more digital approach, the benefits have become too obvious to ignore and investments into AI have increased in all fields. The Longevity sector is no different.
Professor Shinsuke Yuasa from the Keio University School of Medicine in Tokyo has been focusing on AI development for drug development in Longevity. There are many approaches to machine learning but this research group focuses on convolutional neural networks (CNN) – deep neural networks commonly applied for visual imagery analysis that expresses how one shape is modified by another.
A CNN approach has already allowed for endothelial cell identification derived from pluripotent stem cells, using phase-contrast microscopy images [3]. However, their most recent – and possibly most exciting – breakthrough is the development of DeepSeSMo.
So what is DeepSeSMo? It is a CNN based scoring system that was trained to quantify the number of senescent cells based on biological microscopy slides. Senescent cells are cells that have an arrested cell cycle found during aging. They have distinct morphologies with enlarged and flat cell bodies and distinct aggregation heterochromatin, a tightly-packed form of DNA.
Healthy cells were triggered into senescence through three different stressors associated with senescence: reactive oxygen species via H2O2 exposure, the anti-cancer reagent camptothecin and replication stress through repetitive passage of cells.
Three stressors were explored to determine if the mechanisms that triggered cells into a senescence affected the developed algorithm. It was no surprise that when the algorithm was trained on a combined set of senescent cells from all three stressors, it gave more generalisable and accurate results.
The training set consisted of human umbilical vein endothelial cells and the evaluation of just how well the algorithm was trained was done on human diploid fibroblasts. Not only did the results accurately predict the quantity of senescent cells, but it did so at an impressive 0.08-0.1ms.
Once validated, this algorithm was tested on tissue treated with various drugs in an attempt to find senolytic or senotherapeutic drugs – ones removing senescent cells or ones preventing cells from entering the senescence phase. Indeed, four targets were identified: terreic acid, PD-98059, daidzein and Y-27632·2HCl [4].
Longevity properties have been identified for all of the drugs picked out by the algorithm. Terreic acid extends yeast’s lifespan [5], PD-98059 and daizein suppresses cellular senescence and associated phenotypes [6] [7] and Y-27632·2HCl regulates the cell cycle more generally [8].
Deep-SeSMo is likely to be the beginning of the machine-learning boom within aging research for both research and industry, cutting down costs and saving lives…
All results were cross validated and confirmed with the accepted biological quantification of senescence using senescence-associated beta galactoside. Even so the authors point out the possible limitations. Cells in the intermediate state transitioning into senescence lead to false negatives or positives, yet within this specific study this did not change results when accounted for [4].
Such an approach could be used to eliminate the human bias and error in biological methods and applied on a large scale. Senolytic drugs could be screened at a faster and cheaper rate leading to the development of only the best candidates in the pipeline.
Furthermore, this particular algorithm could further Longevity research though elucidation of senescent mechanisms. Deep-SeSMo picks up characteristics unique to senescent cells and can thus be trained further to identify how and when they arise, as well as exploring the senescence threshold.
Deep-SeSMo is likely to be the beginning of the machine-learning boom within aging research for both research and industry, cutting down costs and saving lives.
[1] https://www.medicinenet.com/script/main/art.asp?articlekey=9877
[2] https://tinyurl.com/y22w8pcj
[3] https://pubmed.ncbi.nlm.nih.gov/29754958/
[4] https://www.nature.com/articles/s41467-020-20213-0
[5] https://tinyurl.com/y5bqfyy9
[6] https://www.nature.com/articles/onc2013477
[7] https://www.mdpi.com/1422-0067/15/11/21419
[8] https://www.sciencedirect.com/science/article/pii/S0003996915000680?via%3Dihub