Benchmarking Deep Learning Predictions of Mutation-Induced Fold Switching
1. The study introduces a quantitative, NMR-characterized benchmark for mutation-induced fold switching using the GA/GB metamorphic model system, enabling residue-level evaluation of whether structure predictors can capture shifts among 3α (GA), 4β+α (GB), mixed, and unfolded states.
2. They performed systematic scanning mutagenesis at three key sites (GA98 L45X, GB98 T25X, and GB98-T25I L20X), generating 60 soluble monomeric variants and measuring fold-state populations by 2D 1H-15N HSQC NMR, with no additional folded states detected beyond the canonical GA and GB folds.
3. The mutational landscapes were highly position-dependent: at L45 in GA98, only L45Y fully switched from 3α to 4β+α, while many other substitutions unfolded the protein; mechanistically, the unique success of Tyr is attributed to stabilizing the GB-like hairpin via a specific sidechain H-bond (Y45 OH to D47).
4. At T25 in GB98, fold preferences were more distributed: I/L/V/Y at position 25 produced mixed populations of 4β+α and 3α, consistent with branching/steric effects destabilizing the GB-like state while improving packing in the GA-like core; several aromatics and charged substitutions drove unfolding.
5. At L20 in GB98-T25I, nearly any substitution disrupted the 3α state (small buried core sensitivity), often shifting to 4β+α or to mixtures with unfolded; a subset (A/C/F/W/Y) switched fully to 4β+α, while others produced partial unfolding, highlighting how local core geometry constrains fold-switchability.
6. They benchmarked multiple predictors (AlphaFold2/3, AF-Cluster, AFSample2, CF-random, ESMFold2, ESM3, Boltz-1x) without templates, generating 1000 models per mutant and classifying outputs by TM-align to GA vs GB reference structures; performance was quantified via RV coefficients (capturing mutation-to-mutation variability) and Brier scores (absolute classification error).
7. Results showed uneven capability: residue 45 mutants were the most predictable (best RV around ~0.49 for AlphaFold2 and ~0.48 for AFSample2), while residue 25 and 20 were generally difficult (most methods RV < 0.2), with AF-Cluster strongest at residue 20 (RV ~0.46) and CF-random best at residue 25 (RV ~0.28), underscoring strong context dependence.
8. A deeper look at AlphaFold2 on L45 mutants revealed a strong overall bias toward the 3α fold, yet the distribution of predicted states still tracked experimental outcomes; notably, recycling mattered: increasing recycling steps reduced the universal 3α bias and amplified differences between folded vs unfolded mutants, and for the true switch mutant 45Y the 4β+α fraction rose substantially with more recycling.
9. Physics-based and design-style methods were also tested: Rosetta fixed-backbone ΔΔG showed limited discrimination and often struggled to flag unfolding mutations; inverse folding methods (ProteinMPNN, ESM-IF1, Frame2Seq) largely failed to separate fold outcomes from unfolded using 4β+α backbones, but showed some discrimination with 3α backbones, again emphasizing dependence on which structural state is used as the reference.
10. By leveraging MegaScale experimental stability data, they found unfolded mutants generally had higher ΔΔG, supporting the NMR classifications; strikingly, for residue 45, AlphaFold2’s predicted “unfolded model fraction” correlated strongly with experimental ΔΔG (r ~ 0.89), whereas residues 20 and 25 showed weak/no correlation—suggesting some site-specific sensitivity to mutational energetics.
11. Generalization to other metamorphic systems was limited: for Sa1 mutants, most methods predicted almost exclusively one state and did not capture mutation-driven shifts; for KaiB, removing MSA information allowed AlphaFold2/3 to correctly reflect a designed shift in state favorability (WT resting state vs designed sequence producing fold-switched/unfolded mix), highlighting that MSA usage can qualitatively alter conclusions.
12. The take-home message is that current deep learning predictors show partial, position-dependent sensitivity to mutation-induced state changes, but struggle broadly with multi-state energetics; the paper argues that future progress may require explicit modeling of ensembles and thermodynamics (e.g., hybrid approaches, NMR-informed training, or models trained on ΔΔG/ensemble data).
📜Paper: https://t.co/KxNfSw3MS2
#ProteinFolding #MetamorphicProteins #FoldSwitching #NMR #AlphaFold #ProteinDesign #ComputationalBiology #Benchmarking #StructuralBiology #DeepLearning