Nano-scale to Meso-scale: Practice and Pedagogy of Deep Learning Applied to Molecular Systems

Abstract

Deep learning has found its way into the chemical sciences, and brought with it successes that have been observed in “traditional” applications such as computer vision and sentiment analysis.1–3 The integration of these methods into modern industry, pharma, and materials discovery makes it clear that the next generation of chemical and biological scientists and engineers will be operating at the intersection of data science and their respective fields. Unfortunately, nuances in these chemical systems mean that often methods leveraged in computer science applications are not out-of-the box translatable; expert knowledge is required to design a machine learning solution at the intersection of these fields.4–6 Here, this nuanced design process is detailed for two deep learning applications to the prediction of reaction rate constants, critical for understanding and design of reactive systems.7 Additionally, an optional learning event, designed to give undergraduate students an opportunity to learn and apply modern methods at this intersection, is detailed. Finally, work to leverage deep neural machine translation to improve protein thermal stability is presented, and the nuances highlighted. The translator has the ability to produce thermally stable variants of existing proteins or score already generated variants, allowing them to be leveraged at high temperatures where many proteins lose their applicability compared to other materials.8