Messenger RNA (mRNA) vaccines are a new alternative to conventional vaccines with a prominent role in infectious disease control. These vaccines are produced in in vitro transcription (IVT) reactions, catalyzed by RNA polymerase in cascade reactions. To ensure an efficient and cost-effective manufacturing process, essential for a large-scale production and effective vaccine supply chain, the IVT reaction needs to be optimized. IVT is a complex reaction that contains a large number of variables that can affect its outcome. Traditional optimization methods rely on classic Design of Experiments methods, which are time-consuming and can present human bias or based on simplified assumptions. In this contribution, we propose the use of Machine Learning approaches to perform a data-driven optimization of an mRNA IVT reaction. A Bayesian optimization method and model interpretability techniques were used to automate experiment design, providing a feedback loop.IVT reaction conditions were found under 60 optimization runs that produced 12 g • L −1 in solely 2 h. The results obtained outperform published industry standards and data reported in literature in terms of both achievable reaction yield and reduction of production time. Furthermore, this shows the potential of Bayesian optimization as a cost-effective optimization tool within (bio)chemical applications.

No abstract

No abstract

Abstract. Most agent-based models use uniform random distributions to configure the values of initial conditions in simulations. Moreover, random values are often used to distribute objects spatially, to determine unmeasured exogenous factors, and sometimes to determine aspects of the agents' behaviour. An alternative approach to the design and initialisation of an agent-based simulation is to adapt the principles of microsimulation using external sources of contextual data. In this approach, quantitative data are used in several ways: sample surveys for the initial conditions, calculated regression equations for the evolution of variables, and empirically based distributions for the calculation of new values. In this paper, we consider some of the advantages and difficulties of this alternative approach.

No abstract