The entire article is summed up in the title! they do not tell us what they are, this is as close as they come to a revelation:
“In addition, these compounds – which all bind Bcl-2, a protein that regulates apoptosis and is also a chemotherapy target – exhibited favorable toxicity profiles in hemolysis and genotoxicity experiments. Testing these compounds in 80-week-old-mice (an age corresponding to 80-year-old humans), the researchers found that they efficiently removed senescent cells and reduced the expression of genes associated to senescence in the kidneys.”
Unless I’m missing something this is not “artificial intelligence”. A computer was programmed to scan a database (or research articles) to find compounds that met certain criteria which were provided by the programmers. It is extremely useful, but not intelligent.
The three compounds appear to be named BRD-K20733377, BRD-K56819078 and BRD-K44839765. These were all compared to a molecule the researchers were previously aware of called ABT-737. This can be read about in the description of the figure here:
https://www.researchgate.net/figure/Validation-of-identified-compounds-in-a-model-of-therapy-induced-senescence-a-d_fig3_370525031
Which is linked to from the study’s Research Gate page:
https://www.researchgate.net/publication/370525031_Discovering_small-molecule_senolytics_with_deep_neural_networks
Thank you for digging that up 
Regarding the methods of the study, I also had the impression that they had simply scanned a database from the article from the news site, but the study’s page in Nature suggests they used a neural network named Chemprop.
It is confusing how in one article they say they “screened” compounds and in another they say the computer “predicted” compounds. Not being a computer expert myself, it sounds fishy.
Edit: The computer predicted (selected) 2000+ compounds out of 800,000. Then the researchers screened the 2000 to come up with three likely candidates. I still don’t think it is A.I.
Ah - thats poor wording by the author of the article. “Predicted” to me suggests that its predicting a molecular structure that should work, one that is not a natural molecule they’ve just found.
I’m sure there are many different approaches with AI / machine learning - but screening compounds doesn’t sound that sophisticated.
I think the more interesting area is actually designing new drugs and biologics to accomplish a specific task.
there is a really good book about the early days of structured drug design I recommend if anyone is interested:
https://www.amazon.com/Billion-Dollar-Molecule-Companys-Perfect/dp/0671510576
Join journalist Barry Werth as he pulls back the curtain on Vertex, a start-up pharmaceutical company, and witness firsthand the intense drama being played out in the pioneering and hugely profitable field of drug research.
Founded by Joshua Boger, a dynamic Harvard- and Merck-trained scientific whiz kid, Vertex is dedicated to designing—atom by atom—both a new life-saving immunosuppressant drug, and a drug to combat the virus that causes AIDS.
You will be hooked from start to finish, as you go from the labs, where obsessive, fiercely competitive scientists struggle for a breakthrough, to Wall Street, where the wheeling and dealing takes on a life of its own, as Boger courts investors and finally decides to take Vertex public.