(361j) Integrating Automation and Machine Learning for Accelerated Closed-Loop Molecular Discovery | AIChE

(361j) Integrating Automation and Machine Learning for Accelerated Closed-Loop Molecular Discovery

Authors 

Koscher, B. A., Massachusetts Institute of Technology
Jensen, K. F., Massachusetts Institute of Technology
The discovery of functional materials has traditionally been driven by chemists’ intuition and manual experimentation; the automation of the design-make-test-analyze (DMTA) cycle presents a promising alternative for accelerating multi-functional materials discovery.

Integrating advancements in automatization and machine learning (ML), we developed a closed-loop, autonomous molecular discovery platform to both accelerate hit-to-lead (exploration) and lead-optimization (exploitation) workflows for molecules with multiple desired properties. The platform was designed for general chemistry with general, adaptable synthesis, workup, isolation, and characterization protocols. These capabilities can accommodate the diversity of molecules and multi-step syntheses recommended by its molecular generation and retrosynthesis algorithms. We demonstrate this platform in exploration and exploitation case studies without human intervention (beyond providing consumables and error recovery) on dye-like molecules targeting wavelength of maximum absorption, water–octanol partitioning, and photo-oxidative stability.

In the exploration study, the platform, guided by property-prediction ML models, automatically performed three iterations of the DMTA cycle. Using inexpensive syntheses to diverse targets as its means to explore and learn the structure–function space of four rarely reported scaffolds, the platform realized 312 unreported molecules and substantially improved the predictions of the guiding property-prediction models in these local chemical spaces. In the exploitation study, the same platform leveraged models trained on pre-existing examples of a fifth scaffold to discover 6 top-performing molecules within their structure–property space.

The platform demonstrates the potential for automated platforms which integrate the steps of the molecular discovery cycle (prediction, synthesis, measurement, and model retraining) to autonomously understand a local chemical space to aid in the discovery of functional molecules. As the components of this automated DMTA cycle are modular, with alternate property or retrosynthesis models, additional functionalities could be explored, such as biological activity, requiring only the deployment of the corresponding characterization hardware and/or protocols.

This platform is the culmination of many collaborators across multiple laboratories, and as such, this presentation will focus on the design of the experimental platform and its integration with its governing models to achieve autonomous operation.