Low Data Machine Learning for Sustainable Chemical Sciences

ongoing

Innovation in the chemical sciences is bound to iInnovation in the chemical sciences is bound to impact on Healthcare and Society. Supported by improved analytical methods and automation, brute force and large-scale experimentation have been playing an important role in generating volumes of chemical and biological data. These data now enable the support to decision making through machine learning/artificial intelligence (ML/AI) algorithms. In doing so, such algorithms help in the design and prioritization of experiments. As a result, we are witnessing a renaissance of ML/AI for accelerating chemistry, as in planning retrosyntheses, predicting reaction products, designing drug leads and materials de novo, and deconvoluting drug targets among others. Despite the chemistry advances leveraged by ML/AI, one can argue that not all research questions and findings benefit from the availability of big data (e.g., some discoveries are serendipitous). Here we argue that the current ML/AI toolkit excels in scenarios (big data) that do not entirely map onto daily practice (low and sparse data, often out of training set distribution). Thus, the scientific potential of ML/AI is not fully realized when state-of-the-art tools are implemented and primed for big data and highly charted search spaces. The disconnect between what is feasible and generally needed is apparent and is impacting our ability to advance the chemical sciences at a faster pace. In LowDataML we propose a suite of intersectional and transdisciplinary research projects that will create a new breed of scientists. Through an extensive training program, we will deliver 10 PhDs to the European R&D ecosystem, who are experts in low data ML with domain knowledge in synthetic organic chemistry and drug discovery. Further, the projects seamlessly integrate together and focus on: i) real world needs; ii) sustainability by minimising cost, time and materials consumption, with special focus on ecofriendly research practices.