StabilityDesigner — ΔΔG saturation mutagenesis

Predicted protein stability change on point mutations · scan over all 164 positions · 164-residue protein · structural features: yes

0.330

Held-out Pearson r (FireProtDB test)

2.19

Held-out RMSE (kcal/mol)

0.369

S669 Pearson r (full, independent)

+0.88

Ssym antisymmetry bias (0 = ideal)

Saturation-mutagenesis heatmap

Blue = stabilizing (ΔΔG<0), red = destabilizing (ΔΔG>0). Convention: ΔΔG > 0 destabilizes the fold.

Top suggested stabilizing mutations

Mutation	Predicted ΔΔG (kcal/mol)	Call
D10L	-1.797	STABILIZING
A49Y	-1.758	STABILIZING
A82Y	-1.752	STABILIZING
D70L	-1.705	STABILIZING
D10I	-1.657	STABILIZING
A134Y	-1.649	STABILIZING
S36Y	-1.645	STABILIZING
G12Y	-1.629	STABILIZING
G156Y	-1.624	STABILIZING
E11L	-1.603	STABILIZING
A112I	-1.592	STABILIZING
D70I	-1.586	STABILIZING
S117I	-1.555	STABILIZING
A49I	-1.550	STABILIZING
A41Y	-1.547	STABILIZING

Model & data

A RF regressor (chosen by grouped, protein-disjoint cross-validation) over 68 interpretable biophysical features: amino-acid property deltas (hydrophobicity, volume, charge, flexibility, polarity, helix/sheet propensity, side-chain H-bond capacity), BLOSUM62 and Grantham distance, substitution-type flags, local-window composition, and — when a PDB is supplied — relative solvent accessibility, secondary structure, Cβ/Cα contact number, burial depth, backbone H-bond count, plus engineered structure×chemistry interactions (e.g. cavity creation in the buried core).

Trained on FireProtDB curated experimental ΔΔG values (protein-disjoint train/val/test splits via the ThermoMPNN benchmark), with light antisymmetric reverse-mutation augmentation. Evaluated on the full independent S669 benchmark and on Ssym for forward/reverse antisymmetry.

Top feature importances (permutation, held-out)

Feature	Importance
rel_sasa	0.107
pos_frac	0.077
d_sheet_prop	0.070
buried_x_d_volume	0.066
win_hydro_mean	0.063
win_helix_mean	0.056
d_polarity	0.040
d_flex	0.039
buried_x_abs_d_hydro	0.036
d_volume	0.030

Query sequence

MNIFEMLRIDEGLRLKIYKDTEGYYTIGIGHLLTKSPSLNAAKSELDKAIGRNCNGVITKDEAEKLFNQDVDAAVRGILRNAKLKPVYDSLDAVRRCALINMVFQMGETGVAGFTNSLRMLQQKRWDEAAVNLAKSRWYNQTPNRAKRVITTFRTGTWDAYKNL

StabilityDesigner · CPU-only · for research/portfolio use. ΔΔG predictions are computational estimates; experimental validation (e.g. thermal/chemical denaturation) is required before drawing conclusions. Data: FireProtDB (Stourac et al. 2021, NAR), S669 (Pancotti et al. 2022, Brief. Bioinform.), Ssym (Pucci et al. 2018, Bioinformatics), splits via ThermoMPNN (Dieckhaus et al. 2024, PNAS).