Efficient design of stochastic molecular circuits with differentiable programming

Existing methods for the inverse design of gene regulatory circuits typically rely on hand-crafted rules, evolutionary algorithms, and problem-specific techniques, often disregarding stochastic effects. Developing more general and efficient techniques would have a significant impact on many fields, including synthetic biology, developmental biology, and tissue engineering. In this work, we explore the use of general gradient-based optimization approaches to tackle this problem. We rely on automatic differentiation algorithms and hardware acceleration provided by neural network libraries to efficiently compute high-dimensional gradients through entire simulations. To retain full access to the stochastic effects that influence the system dynamics, we employ Stochastic Simulation Algorithms (SSAs), which in turn introduce additional complexities in gradient computations and optimization. We discuss the applicability of this framework to the design of genetic circuits with potentially nonlinear interactions and to the control of inherently stochastic effects, such as fluctuations in gene expression levels that cause readout errors in the circuit's outputs.