Numerous challenges in science and engineering can be framed as optimization tasks, including the maximization of reaction yields, the optimization of molecular properties, and the development of drug formulations. Increasingly, machine learning-driven strategies are adopted to navigate high-dimensional spaces and solve these tasks efficiently. In this talk, I will present novel tools and algorithms that we recently developed with the goal of creating a unified and flexible framework for experiment planning in chemistry and drug discovery. Based on the principles of Bayesian optimization, we are building software tools tailored to the needs of chemical and pharmaceutical research. For instance, to enable the efficient optimization of categorical variables, such as molecules, we have developed Gryffin, an algorithm capable of leveraging domain knowledge in the form of physicochemical descriptors to accelerate molecular optimization. To support Gryffin with external information from, e.g., simulations or cheaper assays, we have developed Gemini, a deep-learning approach that produces accurate property predictions based on approximate computations or assays, calibrated against small sets of accurate measurements. To increase the robustness and reproducibility of experimental protocols and processes, we have developed Golem, a general-purpose algorithm that identifies optimal solutions when significant variability is present in the experimental conditions. In summary, I will provide an up-to-date overview of our efforts to expand the repertoire of data-driven experiment planning algorithms tailored to chemical and pharmaceutical research. Thanks to the flexibility of machine learning techniques, these algorithms may be deployed for the efficient optimization of experimental protocols, processes, and for the design of functional molecules and materials across chemistry, drug discovery and development.

Data-driven algorithms for experiment planning in chemistry and drug discovery

ACS Fall 2021

The official archive and repository for all ACS Meetings & Expositions abstracts and select posters or presentations.