A Unified Framework for Benchmarking Machine Learning Methods in Genetic Circuit Design

Paper presented at BRACIS 2026 and published by Springer (Lecture Notes in Computer Science), where I am the first author. It proposes a unified benchmark for designing genetic circuits under molecular noise and compares four search methods - Cartesian Genetic Programming, Policy Gradient, Bayesian Optimization and a Neural Surrogate - on the same stochastic simulation oracle (Gillespie SSA), under an equalized budget. All three learning methods outperform the evolutionary baseline, and Bayesian Optimization leads on three of the five logic gates. The key finding is methodological: Policy Gradient's apparent advantage vanishes once every method receives the same simulation budget.