Benjamin Cossins began his presentation by introducing Exscientia's automated high throughput physics-based platform, which had been integrated into their AI drug discovery process. The company, being AI-first, relied heavily on data-driven methods, gathering all available data at the project's outset and integrating it into their platform. They built machine learning models to predict important properties in the compounds they aimed to generate, using these models to drive their generative design platform.
Cossins explained that Exscientia's core technology involved using AI to generate compounds that met their project objectives. They employ active learning to make selections based on different strategies, followed by chemical synthesis and testing. The process is iterative, with data and models being continuously updated. The goal is to produce candidates in about 10 to 15 design cycles, aiming for efficiency and fewer cycles.
Cossins highlighted the integration of physics-based tools into their process, using them as an additional filter to direct compound selection. He discussed the use of alchemical binding free energy calculations to predict affinities, noting that these calculations, while computationally intensive, are effective and often outperform data-driven ML models. Exscientia's physics-based platform was built using open source software and toolkits, such as Open Forcefield, Amber, Gromacs, and OpenMM. This approach is cost-effective and provides a wide range of features, though lacking dedicated support.
The platform includes an automated, cloud-based high throughput workflow system, allowing Exscientia to scale their operations significantly. They could use up to 1,000 GPUs at a time and process thousands of compounds daily. Cossins also discussed the use of active learning and machine learning to improve throughput and accuracy, building ML models to predict important properties and guide compound selection.
In conclusion, Cossins provided an example of the platform's real-world application in the NLRP3 programme, a complex project involving multiple enantiomers and charge states. The tools had been used to filter compounds based on affinity and improve predictions, demonstrating the platform's practical utility and impact on drug discovery efforts.
