Integrating machine learning tools in protein design: a case of MHETase engineering for PET biodeconstruction

The integration of machine learning tools into protein engineering offers substantial promise, yet linking computational predictions to experimental performance remains challenging. Here, we applied accessible computational platforms to engineer Ideonella sakaiensis MHETase, a key enzyme in the biodeconstruction of polyethylene terephthalate (PET). Homologue identification with EnzymeMiner, followed by solubility-focused redesign using AggreProt and ProteinMPNN, enabled the finding of MHETase-like enzymes and the generation of is-MHETase variants with increase in soluble expression, as supported by predictive scoring and experimental validation. However, retaining catalytic activity proved considerably more difficult. Structural modeling and kinetic analyses revealed that ProteinMPNN introduced substitutions incompatible with MHETase function, reflecting biases toward sequence patterns common in broader esterase families rather than constraints specific to this narrow enzyme subgroup. These findings highlight both the utility and the current limitations of accessible ML tools for enzyme engineering, underscoring the continued necessity of experimental validation to guide and refine computational predictions.