How is ART used?

ART is typically used when an algorithm that performs well with small data sets (e.g., Fig 6 from Radivojevic et al, 2020), provides uncertainty quantification, and delivers actionable recommendations is required:

 ART has been successfully used to predict bioengineering outcomes and guide active learning processes that accelerate fundamental tasks in synthetic biology:

ART has guided an active learning process that optimized media to increase the production of flaviolin (60% and 70% increases in titer, and 350% increase in process yield) by suggesting unexpected media formulations (A). ART has been used to design pathways by recommending which combination of pathway promoters to use to increase tryptophan productivity ~105% (B). ART has been used for host engineering by pinpointing which combinations of genes to downregulate through CRISPR interference (CRISPRi) to increase isoprenol titer by ~500%.

ART outperforms other state-of-the-art approaches for guiding active learning, in a simulated media optimization process:

We used ART, JMP, gpCAM to guide three simulated active learning processes where the fifteen input variables and responses were used to recommend the inputs for the next cycle. Response was simulated through three different functions that present different levels of difficulty to being “learnt”. The starting inputs were the same for all algorithms and each process was run for ten Design-Build-Test-Learn (DBTL) cycles. To counter the usual stochasticity in active learning processes, each process was run ten different times, and the lines and shaded areas represent the mean and standard deviation of the highest production for each cycle. We also tested a new recommendation algorithm for ART that improves on its original parallel tempering approach, by using differential evolution (ART_DE).

While this is not a thorough comparison, it shows that for the large phase spaces often encountered in synthetic biology, ART can outperform other popular choices.