Harnessing Nature's Algorithm: From Test Tubes to Autonomous In Vivo Evolution.

Directed evolution (DE) enables the engineering of biomolecules without prior structural knowledge. However, traditional step-wise DE is constrained by limited screening throughput. To more efficiently navigate epistatic fitness landscapes, the field is increasingly adopting autonomous, continuous in vivo evolution systems. This review critically examines the molecular architectures and engineering principles driving this transition. We evaluate strategies for continuous genetic diversification-ranging from orthogonal replication systems (e.g., OrthoRep, T7-ORACLE) to CRISPR-guided mutagenesis (e.g., EvolvR)-with a focus on the fundamental trade-off between mutational load and host viability. Furthermore, we analyze the biophysical constraints of screening and the kinetic demands of coupling real-time selection with ultra-fast mutagenesis, as exemplified by phage-assisted continuous evolution (PACE). Crucially, we explore the functional integration of machine learning (ML), highlighting how active learning models and zero-shot predictions via protein language models (PLMs) can resolve epistatic complexities and mitigate the latency of next-generation sequencing. Finally, we discuss the multidimensional hardware and algorithmic bottlenecks currently impeding the realization of fully closed-loop biofoundries, and assess the strategic implications of these technologies for accelerating the engineering of complex therapeutics.