Research published today in Nature Aging has described a machine learning algorithm that finds senescent cell-removing drugs (senolytics) and compared the algorithm’s discoveries to existing compounds.
A quest for effective therapeutics
After covering familiar territory regarding senescent cells, this paper begins with a discussion of existing senolytic compounds, such as the well-known dasatinib and quercetin combination. Most of these compounds were found through bioinformatics approaches that investigated how these cells stay alive long after they should have died by apoptosis [1].
While some senolytics have shown a certain amount of effectiveness in animal models [2], there are relatively few of these drugs, some were found to be ineffective against the diseases they were meant to target, and some have significant side effects. For example, a Phase 2 clinical trial has found that navitoclax causes blood platelets and some immune cells to decline [3].
Even with these disappointing results in mind, researchers have still found merit in the base principle of senolytic development, that senescent cells are valid targets and that removing them with the right compounds would show clinical effectiveness in treating diseases. Finding these compounds, however, is the problem.
Machine learning methods have been used for drug discovery in other areas, including antibiotics [4], but this team notes that there has been no previous progress in using them to find senolytics. Therefore, they sought to meet that need, training an artificial intelligence system from the ground up to find them.
Working on a very large dataset
The researchers took a well-established graph model of machine learning and trained it with detailed information on 2,352 compounds, which they tested themselves for senolytic activity against human lung fibroblasts that had been chemically driven senescent with etoposide. 45 of these initial 2,352 were selectively effective against senescent cells. The researchers then applied this trained model to a full 804,959 compounds.
The results the model returned were highly diverse; it determined that some compounds were highly likely to be senolytic and that others were not. After applying a filter against compounds that are too similar in structure to existing senolytics, the researchers selected 216 compounds that the algorithm chose and that they had on hand along with an additional 50 as negative controls.
The preliminary analysis was favorable: 25 of the initial 216 were found, by initial experiment, to have senolytic properties in the real world. While this is a relatively small percentage, it is clear that the algorithm had effectively narrowed down a very large search space. None of the negative controls had senolytic properties.
Comparison to a gold standard
The team then compared the effectiveness of these candidates to ABT-737, a drug that has significant senolytic properties but is unsuitable for clinical use due to its low bioavailability and side effects. At the 10-micromole dose, the researchers narrowed down the field further to three particular compounds that are roughly as effective and specific as ABT-737. Critically, none of these compounds decreased the viability of control cells, which ABT-737 is known to do.
The researchers note some appealing characteristics of these compounds: they are druglike compounds with no current clinical use, they do not resemble the compounds used in the training dataset, and their chemical properties make them good candidates for oral administration. Further testing revealed that they do not excessively harm healthy liver cells and that they are senolytic against cells that were driven senescent with doxorubicin.
All three of these compounds were found to bind Bcl-2, a mechanism of action that is common to multiple senolytics. This is a critical achievement in AI-guided drug discovery: despite the fact that they don’t look like other senolytics, these algorithmically discovered compounds have been determined to work the same way.
Finally, one of these compounds was tested in naturally aged mice, and the results were positive. The treated mice did not seem to suffer from side effects, and their expression of the senescence markers SA-β-gal and p16 were substantially lower in the kidneys. This, again, represents a critical achievement: an artificial intelligence had successfully discovered a compound that reduces a key biomarker of aging in an animal model.
Conclusion
There are, of course, limitations to the model used in this study and the data used to train it. The initial training data consisted only of a specific cell population that was driven to senescence in a specific way. Senescent cells are highly heterogenous, and other types of senescent cells might be vulnerable to entirely different approaches that this trained model cannot discover.
However, this was an effective proof of concept, and it certainly appears that AI drug discovery applies to senolytics. It is likely that a more robustly trained model on different types of senescent cells can offer even more useful information, discovering potential therapeutics that unassisted people would never have found themselves. Determining whether or not any of these drugs are truly effective in human beings is, as always, a matter of clinical trials.
Literature
[1] Zhu, Y. I., Tchkonia, T., Pirtskhalava, T., Gower, A. C., Ding, H., Giorgadze, N., … & Kirkland, J. L. (2015). The Achilles’ heel of senescent cells: from transcriptome to senolytic drugs. Aging cell, 14(4), 644-658.
[2] Xu, M., Pirtskhalava, T., Farr, J. N., Weigand, B. M., Palmer, A. K., Weivoda, M. M., … & Kirkland, J. L. (2018). Senolytics improve physical function and increase lifespan in old age. Nature medicine, 24(8), 1246-1256.
[3] Rudin, C. M., Hann, C. L., Garon, E. B., Ribeiro de Oliveira, M., Bonomi, P. D., Camidge, D. R., … & Gandhi, L. (2012). Phase II Study of Single-Agent Navitoclax (ABT-263) and Biomarker Correlates in Patients with Relapsed Small Cell Lung Cancer. Clinical Cancer Research, 18(11), 3163-3169.
[4] Stokes, J. M., Yang, K., Swanson, K., Jin, W., Cubillos-Ruiz, A., Donghia, N. M., … & Collins, J. J. (2020). A deep learning approach to antibiotic discovery. Cell, 180(4), 688-702.
