APIs from the Algorithm

Researchers at ETH Zurich’s Department of Chemistry and Applied Biosciences have developed an AI algorithm that can design new APIs to match a protein’s 3D surface structure. “Drug-target interactome-based generation of novel biologically active molecules” – or DRAGONFLY – is built using deep learning (with a graph transformer neural network) and chemical language models (CLM). 

Chemists have often turned to computers to search for potential new drug molecules, but the resulting molecules can be too difficult or even impossible to synthesize. And when it comes to using standard CLMs, there have been challenges with fine-tuning molecules.

The ETH Zurich team, led by Gisbert Schneider, specifically focused on designing an algorithm that could avoid manual work and ensure the molecules could be synthesized. Their paper, published in Nature Communications, explains that DRAGONFLY “operates on diverse chemical alphabets and does not require fine-tuning through transfer or reinforcement learning specific to a particular application. Furthermore, it enables the incorporation of desired physical and chemical properties into the generation of output molecules,” (1). 

The algorithm is designed to only suggest APIs that interact with the specified protein and location to help reduce side effects and “enable ‘zero-shot’ construction of compound libraries tailored to possess specific bioactivity, synthesizability, and structural novelty.” It was trained using existing information on known interactions between chemical molecules and 3D protein structures.

DRAGONFLY was tested in collaboration with Roche and other partners – and was able to design molecules in the PPAR class and increase the activity of currently available PPARs. Tests from Roche showed that the new APIs were stable and non-toxic. DRAGONFLY is also being used for a project with the Children’s Hospital Zurich for the treatment of medulloblastomas. According to the researchers, their paper shows that DRAGONFLY outperforms standard CLMs for molecular design. The researchers say, “Results suggest that DRAGONFLY is not limited to recapitulating the training data but also has the capacity to explore and generate molecules in previously uncharted regions of chemical space, albeit the extent of this exploration warrants further investigation.”
 

The algorithm and its software have been published openly so that researchers worldwide can use them for projects.