Recent research published in Nature Communications has used machine learning algorithms to find new compounds that can eliminate senescent cells [1].
Searching for new senolytics
Senolytics are molecules that destroy senescent cells. Only a small number of such molecules have been identified, and only two have shown efficacy in clinical trials: dasatinib and quercetin in combination [2]. One of the biggest challenges is that senolytics often only work against specific types of cells. Additionally, some senolytics may work well for one cell type while being toxic to other, non-senescent cell types [3].
There is also a group of senolytics that are used in cancer therapies. However, most of them target pathways that are mutated in cancer. Therefore, they cannot be used as therapeutic agents in different contexts.
These limitations highlight the need to identify new senolytics that can be safely applied in therapies. Sometimes, this search involves panel screens [4]. Other times, it involves targeting the proteins upregulated in senescence. The authors of this paper used a different approach: AI-based computational screens that detect hidden patterns in chemical data.
From big datasets to a few hits
First, the authors assembled an extensive dataset for training the algorithms. Based on published data, they identified 58 senolytics. They also added 19 senolytics reported in commercial patents.
The authors then combined their list of senolytic compounds with a wide variety of compounds that were never described in the literature as having senolytic properties, with the final list consisting of 2,523 compounds. They used this list to train machine learning models to predict whether or not a compound may be a senolytic.
First, researchers focused on two models that both showed poor performance, but the errors they displayed were different. The first model, while returning a few false positives, gave a high number of false negatives. The second model returned the opposite. Other models showed even worse performance.
These researchers then decided to choose a model that turns fewer false positives, with the reason being that false positives are worse than false negatives for early-stage drug discovery, as higher number of false positives will increase the number of predicted hits. More predicted hits result in more compounds that need to be experimentally validated, thus increasing the cost and time.
After choosing their initial model, the researchers used different tools to optimize its performance. Following optimization, they screened a 4,340-compound library and identified 21 hits, then experimentally validated them, first in cells with oncogene-induced senescence.
Three natural products found
The initial experiment identified three compounds with senolytic properties: the natural products ginkgetin, oleandrin, and periplocin, which can be found in traditional herbal medicines. Ginkgetin is a product of Ginkgo biloba, commonly known as ginkgo or maidenhair tree. Oleandrin is a product of Nerium oleander, known as oleander or nerium, and periplocin is a product of Periploca sepium, a Chinese silk vine.
These identified compounds, when added in proper concentrations, have minimal effect on normal cells but decrease the number of senescent cells.
This paper’s authors further validated their hits using a second cell line. This time, they induced senescence in human cancerous cell lines by adding etoposide, a cancer treatment medication, and treated them with the 21 candidates. Again, the same three compounds showed senolytic activity toward the senescent cells but not the normal cells.
The results of this screening suggested that oleandrin is a more potent senolytic than the known senolytic ouabain. Therefore, the researchers decided to focus on this compound and investigate it more thoroughly.
One experiment compared periplocin, oleandrin, and ouabain using low concentrations of compounds. In the cell lines used here, ouabain and periplocin was not cytotoxic. Oleandrin, on the other hand, displayed highly specific senolytic activity. Further testing of oleandrin has shown that it has greater potency and activity regarding the molecule it targets when compared to similar senolytic compounds.
Authors caution that compounds from the same chemical group as oleandrin are known for their toxicity, so oleandrin may be found to be toxic as well. Further research is necessary to determine how and when this compound can be safely used.
Utilizing artificial intelligence in the screens for new compounds
These researchers have shown how machine learning can utilize published screening data and find new therapeutic molecules. This approach reduced the number of potential hits by 200-fold. That efficiency resulted in fewer compounds that proceeded to experimental testing, thus reducing time and cost.
The authors applied some new approaches in the presented research. For example, they used only already published data to train their model. This saves the time and cost that would have been invested in conducting experiments to create training data.
There are relatively few compounds that are known senolytics, so the authors only had a limited number of compounds they could use to train the model. Despite this limitation, they were able to train the model to identify new senolytics. This shows that artificial intelligence can take full advantage of even a small data set.
While this is not the first published paper on AI and senolytics, the authors believe their work created a ”methodological groundwork for a new open science approach to drug discovery and repurposing.”
Literature
[1] Smer-Barreto, V., Quintanilla, A., Elliott, R. J. R., Dawson, J. C., Sun, J., Campa, V. M., Lorente-Macías, Á., Unciti-Broceta, A., Carragher, N. O., Acosta, J. C., & Oyarzún, D. A. (2023). Discovery of senolytics using machine learning. Nature communications, 14(1), 3445.
[2] Hickson, L. J., Langhi Prata, L. G. P., Bobart, S. A., Evans, T. K., Giorgadze, N., Hashmi, S. K., Herrmann, S. M., Jensen, M. D., Jia, Q., Jordan, K. L., Kellogg, T. A., Khosla, S., Koerber, D. M., Lagnado, A. B., Lawson, D. K., LeBrasseur, N. K., Lerman, L. O., McDonald, K. M., McKenzie, T. J., Passos, J. F., … Kirkland, J. L. (2019). Senolytics decrease senescent cells in humans: Preliminary report from a clinical trial of Dasatinib plus Quercetin in individuals with diabetic kidney disease. EBioMedicine, 47, 446–456.
[3] Zhu, Y., Doornebal, E. J., Pirtskhalava, T., Giorgadze, N., Wentworth, M., Fuhrmann-Stroissnigg, H., Niedernhofer, L. J., Robbins, P. D., Tchkonia, T., & Kirkland, J. L. (2017). New agents that target senescent cells: the flavone, fisetin, and the BCL-XL inhibitors, A1331852 and A1155463. Aging, 9(3), 955–963.
[4] Yousefzadeh, M. J., Zhu, Y., McGowan, S. J., Angelini, L., Fuhrmann-Stroissnigg, H., Xu, M., Ling, Y. Y., Melos, K. I., Pirtskhalava, T., Inman, C. L., McGuckian, C., Wade, E. A., Kato, J. I., Grassi, D., Wentworth, M., Burd, C. E., Arriaga, E. A., Ladiges, W. L., Tchkonia, T., Kirkland, J. L., … Niedernhofer, L. J. (2018). Fisetin is a senotherapeutic that extends health and lifespan. EBioMedicine, 36, 18–28.
