There are many tools for generating new molecules needed for new drugs. Some are open source and other’s are made by companies as paid subscriptions or paid collaborations. These AI tools help in docking of a new chemical entity to the protein of interest, or they can also help in designing the chemical compound, DNA/RNA sequence, or protein of interest. The technique of docking molecules have been around for sometime with companies like Schrodinger, Atomwise, Insilico Medicine and Chai Discovery performing much of the activities for their clients.
However, the open source tools are getting numerous. Consider NVIDIA, the company that is creating the GPU’s is creating a NVIDIA BioNeMo Framework to help develop tools in Biology, AlphaFold2 helps predict structure, Boltz2 predicts protein-ligand binding and RDKit is the core cheminformatics toolkit that is used as a backbone for drug discovery pipelines.
One particularly interesting one is from researchers at University of Florida and NYU, New York that has developed a method called PropMolFlow. According to them, ultimately, the needs is to get a molecule with specific properties rather than molecules that have a certain formula. PropMolFlow specifically invents new molecules that fit a certain property. This has been derived from DALL-E a diffussion model that was used for generating new photo’s. Now this method of using diffussion models did exist and were used in drug discovery since 2022 but what the researchers have done is to make them much faster with 100 inference steps per molecule. This property guided molecule generation can be used beyond the small molecule QM9 dataset and will be very useful for all future drug developers.
Direct link to the paper here: https://arxiv.org/pdf/2505.21469
