Artificial intelligence (AI) is revolutionizing our world by predicting future behaviours from large datasets. Recent excitement has grown around AI that requires small (<100) datasets, guiding the investigation of vital areas like pharmaceuticals. "Active learning" (AL) techniques use experiment outcomes to make recommendations for new experiment designs based on areas of the state space where it is less certain of its predictions. The results of these predictions feed into the model for continual improvement. Bayesian Optimisation is being explored for material discovery tasks, however this process is limited in that models attempt to find the optimal material given a target profile, compared to AL, which focusses on building a robust and interpretable model. This project will aim to develop an inexpensive robot formulator with AL-driven decision-making to accelerate medicine manufacture. It is envisioned that a robot that is able to perform routine laboratory tasks, such as handling liquids and taking analytical measurements, could be guided by a regression AL algorithm such that it not only performs tasks, but learns and executes the next logical step, ultimately developing high quality, safe, and efficacious liquid medicines. Integrating AI, robotics, and automated analysis is an enormous challenge, however the outcomes could be phenomenal. Robotic formulators could drive drug candidates through pharmaceutical bottlenecks rapidly with quality data, using a large design space, with little waste. This will be demonstrated in the project by challenging the robot with complex drugs which are likely to be core medicines of the future. It is envisioned that this approach will be able to identify complex and unintuitive combinations of drug and additives which traditional formulation approaches would not.
It is anticipated that the project will have step-wise impact on future innovations. The robot formulator is inexpensive in comparison to current robotic formulation streams (such as those used in the Materials Innovation Factory) and the algorithms can be run on standard PCs using open-source software. Thus, the approach can be adopted in lower-resource environments for local priority medicines. The focus on algorithm integration timely to make best use of recent regression AL principles, and the blueprint proposed amenable to future developments in AI.
In order to achieve the ambitious aims of this project, the following process will be followed. Firstly, an inexpensive liquid-handling robot (£9k, owned by the PI) will be instructed to develop mixtures of drug and additive (in solution) with a single read-out (e.g. absorbance). An Xarm 5 robotic arm will be interfaced with the liquid-handling robot to allow the formulations to be transferred into analytical instruments. A regression AL algorithm will then analyse which conditions led to solubility and generate predictions on formulations with improved solubility that the robot will automatically investigate. This process will be optimised and evaluated to demonstrate that the robot is "learning" how to make these medicines better. The study will then move on to exploration of multiple product attributes at the same time, akin to "real world" medicine formulation.
The project will match processes the robot performs to those used by industry, to ensure the findings are translatable, guided by collaboration with Bayer. Furthermore, the technology will be designed to use industry-standard software, QBDvision, for high-quality handling and reporting of data. Thus, the robot scientist also provides immaculate reporting of results that are needed for approval of new medicines.
|