Whereas a "one shot" AI-generated drug discovery approach failed, this time, with further insights and two exploration modes, it worked:

"Yoshimori and colleagues discussed the distinction between structure and pharmacophore-steered molecular generation in their reinforcement learning-based approach75. They compared two agent networks, each guided by rewards based on either molecular similarity or pharmacophoric similarity to a known ligand associated with the target of interest. One notable observation was that the agent guided by the molecular similarity reward successfully generated a larger number of molecules exhibiting topological similarities to the reference ligand, but essentially failed to produce any molecules with a satisfactory pharmacophoric score."

"This finding implies the inherent limitations of molecular similarity-constrained methods when it comes to exploring the local chemical space. In our study, we discovered that methods focused on generating structurally analogous compounds could yield molecules that share similarities in both topological structure and ligand pharmacophores. This finding is rational since the concept of ligand pharmacophore is rooted in molecular structure. Moreover, we made an intriguing observation that these two modes of exploration can be complementary, covering distinct regions within the local chemical space. They can also overlap when a fine-grained pharmacophoric representation is employed along with a high molecular similarity cutoff."

Thus, will a poly agent approach be the key to a "one shot" path towards the discovery of small molecule drugs for myriad diseases?

https://www.nature.com/articles/s41467-025-56349-0