Discovering future drugs from tens of billions of molecules is becoming much less expensive thanks to a new method based on artificial intelligence.
Before manufacturing and testing a molecule in the laboratory, researchers want to simulate its interaction with—and therefore its effectiveness against—a target protein. This step, known as virtual screening, makes it possible to quickly rule out many candidates. The problem arises when the library contains billions of molecules: the calculations become very long and costly.
A set of test tubes.
Illustration image from Unsplash
An international team has developed AdaptiveFlow to reduce this burden. The platform has a library of 69 billion molecules already prepared for simulations. Rather than testing every molecule with the same level of precision, it first organizes this vast collection according to their properties.
Specifically, AdaptiveFlow distributes the molecules across a grid defined by 18 characteristics, such as their mass or their ability to dissolve in water. A few representatives from each group are tested first. Groups that produce the most promising results can then be explored much more extensively.
Artificial intelligence can step in at this stage. A model learns from the initial calculations to identify the molecules that deserve more attention. According to the researchers, this strategy can reduce computing costs by up to 1,000 times compared with an exhaustive examination of all 69 billion molecules.
To verify that the system was not merely producing good results on a computer, the team applied it to two proteins linked to cancer. One, PARP1, plays a role in DNA repair and is already a target of drugs. The other, FSP1, helps certain cells resist a particular form of cell death.
Molecules selected by the calculations were then manufactured and tested in reality. Several strongly blocked the targeted proteins as intended. For PARP1, one of the candidates achieved binding efficacy comparable to that of olaparib, a drug already used against certain cancers. This is not, however, a new drug ready to be administered: these molecules remain starting points for further research.
AdaptiveFlow is published as open access, and its code is freely available. Future applications will therefore be able to test this approach on other proteins and libraries, then verify in the laboratory whether the selected candidates retain their properties in cells and, later, in living organisms.