Background: Mutations in intermediate filament genes cause over 100 human diseases, yet predicting clinical severity from genotype remains unreliable due to phenotypic variability and the prevalence of variants of uncertain significance. Methods: FoldX folding free energy changes (ΔΔG) were calculated on AlphaFold2-predicted structures for 323 pathogenic missense mutations across four intermediate filament genes: DES (n = 96), LMNA (n = 94), KRT5 (n = 83), and KRT14 (n = 50). Predictions were validated against manually curated clinical phenotypes from 52 published case reports with variant-specific PubMed references. Feature importance was assessed by random forest analysis and LASSO regression. Results: ΔΔG correlated strongly with observed severity (Pearson r = 0.893, p = 5.74 × 10⁻¹⁹; ROC AUC = 0.967), with a threshold of 2.44 kcal/mol achieving 100% sensitivity and 86.7% specificity for severe phenotypes. Thermodynamic destabilization was the dominant predictor (71.9% feature importance), and five-fold cross-validation confirmed generalizability (mean AUC = 0.964 ± 0.056). ΔΔG did not discriminate pathogenic from benign variants (AUC = 0.520), indicating that it predicts severity rather than pathogenicity. Conclusions: Thermodynamic destabilization is a primary determinant of intermediate filament disease severity across protein types. This framework provides quantitative severity estimation applicable to clinical variant interpretation.