Exploration of Reservoir Properties in Molecular Computing Systems

In this study, we apply the PEN DNA toolbox (Polymerase-Exonuclease-Nickase Dynamic Network Assembly toolbox), a modular framework for molecular computing, to Reservoir Computing. Reservoir Computing is a computing approach initially developed to quickly train Recurrent Neural Networks but has since been adapted to other complex systems, as long as they have appropriate recurrent properties. Previous work has shown that chemical reaction networks (CRNs), both in silico and in vitro, are among such appropriate targets. However, the full range of reservoir characteristics of CRNs, such as memory capacity or computational capabilities, has not been established yet. Moreover, while past studies have shown that large-sized chemical reaction networks are expressive enough to deal with complex behaviors, their wet-lab implementation is more challenging. Here, we aim to evaluate the range of reservoirs that can be generated with the PEN DNA toolbox and clarify the trade-off between reservoir properties and network size. We do so through an automatic exploration of the reservoir property space using a Quality-Diversity algorithm. Finally, to mitigate the high number of simulations required by the algorithm paired with their computational cost, we developed a neural-network-based surrogate model which we apply for either the first 20% or 50% of evaluations as a way to quickly seed good potential solutions. We showed that the surrogate model reduced the run time of the algorithm with no loss in the quality of the best solutions found. Such approach can thus be applied to design a wide range of molecular reservoirs with good characteristics.