User Webinar: Automated Design of Kinase Inhibitors Using AlphaFold 2 Models
Tuesday, 25th June at 3 p.m. (BST)
Speaker: Chris Radoux, Associate Director, Structural Bioinformatics at Exscientia
AI design is well suited for automation. While it is difficult to capture the inspiration that leads to a manually designed compound, capturing generative method input parameters, filters and compound selection algorithms is straightforward. Emulating the way our designers at Exscientia use our design tools, we have developed a fully automated end-to-end structure-based method for the discovery of novel compounds. The approach does not rely on 2D models trained on known active compounds, and is therefore particularly useful for targets with less data, as well as identifying novel starting points for well-studied targets.
In the post AlphaFold2 era, we now have access to many more high quality protein models that can be utilised for structure-based approaches. To assess the useability of AlphaFold2 models for this approach, we designed, made and tested compounds against DYRK1B and PKD1, which represent relatively high and low quality models compared to the whole kinome. Nine compounds were made and tested for each target, achieving a hit rate of 44%, with IC50s in the range of ~50 nM to 1 μM. The majority of Bemis Murcko scaffolds were not seen in ChEMBL, and all compounds showed selectivity for their target over the four other kinases they were tested against.
This level of automation, combined with the positive results for PKD1 which represented one of the lowest quality AlphaFold2 kinase domain models, opens up the possibility of kinome wide design. This will be particularly impactful at the target assessment stage, where we can judge the target based on the quality of the designs or compare designs to the published literature to assess opportunities for novelty.