An Inverse Model for De-novo Design of Synthetic Gene Circuits

Designing a $N$-node gene circuit that realizes a target dynamics means searching for the appropriate regulatory parameters between every pair of nodes which in turn parameterizes the forward gene regulatory network model and simulates the dynamics, hoping to match the target dynamics. However, the $N^2$ increasing dimension of design space poses a significant challenge to find the right design at large $N$. To mitigate this challenge, here we introduce an inverse model— ADSGRN—predicting the $N^2$ regulatory parameters given target dynamics and show that once trained to design at small $N$ it generalizes to design at larger $N$. We trained ADSGRN on a 3-node dataset and evaluated it at about 10-node interval up to 100 nodes from two perspectives: 1. ADSGRN design by a single forward pass itself attained a dynamics-matching mean squared error roughly 50 − 60% lower than random baseline consistently across $N$. 2. Using ADSGRN design as the initialization for an iterative optimization approach—gradient descent (GD) here— can lower the necessary convergence time by 15 − 50% than random initialization, with advantage margin decreasing as $N$ grows. Together, our study supports the promise of developing more powerful inverse models to serve as foundational model that can design gene circuits at any size within milliseconds only by limited compute budget for training at scale of small $N$.