Neural networks built from enzymatic reactions can operate as linear and nonlinear classifiers

The engineering of molecular programs capable of processing patterns of multi-input biomarkers holds great potential in applications ranging from in vitro diagnostics (e.g., viral detection, including COVID-19) to therapeutic interventions (e.g., discriminating cancer cells from normal cells). For this reason, mechanisms to design molecular networks for pattern recognition are highly sought after. In this work, we explore how enzymatic networks can be used for both linear and nonlinear classification tasks. By leveraging steady-state analysis and showing global stability, we demonstrate that these networks can function as molecular perceptrons, fundamental units of artificial neural networks—capable of processing multiple inputs associated with positive and negative weights to achieve linear classification. Furthermore, by composing orthogonal enzymatic reactions, we show that multi-layer networks can be constructed to achieve nonlinear classification.