Directed evolution is a cornerstone of synthetic biology, yet its outcome is heavily constrained by the starting genotype. This latent capacity to innovate—termed evolvability—varies drastically across closely related proteins and microbial strains due to biophysical trade-offs and historical contingency. Financed by the ANR ProtEvol project, this thesis aims to systematically decipher and engineer the determinants of evolvability across two biological scales. At the macromolecular level, we will develop a novel, plasmid-free genomic diversification strategy to map the adaptive pathways of diverse ROK kinase homologs, leveraging AI and machine learning collaborations to extract predictive sequence features of functional promiscuity. Simultaneously, at the organismal scale, we will utilize the automated continuous-culture GM3 platform to evaluate how Escherichia coli strains with divergent pyruvate assimilation backgrounds evolve toward synthetic C1-formatotrophy. By combining high-throughput sequence diversification, machine learning, and automated evolution, this work will transform evolvability from an abstract evolutionary concept into a predictable, steerable parameter for industrial bioproduction.