Limitations of the refolding pipeline for de novo protein design

With the emergence of powerful deep learning-based tools, computational protein design has become a widely accessible technique. Nowadays, it is possible to perform both sequence and structure design in a matter of minutes, making the technology attractive to the broader scientific community. In protein design campaigns, one of the most common in silico strategies to evaluate how well a sequence encodes a target structure is the so-called self-consistency or refolding pipeline. In this approach, a structure prediction model is used to refold the designed sequence to probe whether it is compatible with the intended structure, and is evaluated via two metrics linked to experimental success: the confidence score of the predicted structure (pLDDT) and the self-consistency root-mean-square deviation (scRMSD), which measures how closely the refolded structure matches the target. In this work, we systematically evaluate how different models and structure prediction settings impact these metrics, and to what extent they can be used to reliably filter sequence design candidates. We show that evolutionary information can obscure folding models abilities to assess sequence-structure compatibility, reducing the predictive performance of refolding metrics for experimental success, particularly for designs that share homology with natural sequences. We further highlight limitations of refolding metrics, including their sensitivity to structural features, such as flexibility. Our findings raise awareness of potential pitfalls in refolding-based evaluation and support more informed use of these metrics in protein design campaigns.