Figure 5.

Sequence- and structure-based folding free energy change (ΔΔG_folding in kcal/mol) predictions for CDKL5 missense variants. Violin plots illustrate the distribution of predicted ΔΔG_folding (kcal/mol) predictions for pathogenic and benign CDKL5 variants across five sequence-based (A,B) and five structure-based (C,D) computational methods. ΔΔG_folding values are plotted for variants located within the full-length protein (residues 1–960) (A,C) and the kinase domain (residues 1–302; (B,D). Blue and orange violins represent benign and pathogenic variants, respectively, as classified by germline classification. Sequence-based methods include SAAFEC-SEQ, I-Mutant2.0, INPS, DDGun, DDGemb, and structure-based methods include I-Mutant2.0, INPS, DDGun, mCSM, and DDMut. The figure highlights overall trends in destabilization, with pathogenic variants generally exhibiting more negative ΔΔG_folding values, particularly in the kinase domain (1–302) and in predictions from I-Mutant2.0, DDGemb, and mCSM. Among sequence-based methods, I-Mutant2.0, DDGemb, and SAAFEC-SEQ moderately distinguish between benign and pathogenic variants, with the clearest separation observed in the kinase domain (B). Structure-based methods such as I-Mutant2.0, mCSM, INPS, and DDMut show even stronger separation, particularly within the kinase domain (D). These results indicate that structure-based tools offer superior sensitivity in detecting the destabilizing effects of variants, with I-Mutant2.0 (structure) and mCSM demonstrating the strongest discriminatory performance between pathogenic and benign variants.