StableProt: Structure-Aware Deep Learning for Protein Thermostability (Tm) and Environmental Adaptation (OGT) Prediction

StableProt, a structure-aware model using Foldseek 3Di tokens, predicts protein Tm and organismal OGT with calibrated uncertainty, achieving CRPS of 4.52°C and 8.71°C.

• Why it matters: Accurate, calibrated predictions are needed for enzyme engineering and metagenomic bioprospecting, but sequence-only methods, gradient conflicts, database skew, and uncalibrated outputs limit reliability.
• What they did: StableProt couples amino acid sequences with Foldseek 3Di structural tokens using a frozen SaProt backbone, uses disjoint heads for thermodynamic and ecological context, and applies 14% mesophilic subsampling plus calibrated ensemble modeling.
• The result: It reduces calibration error to 1.92%, recovers 92.0% of hyperthermophiles (≥80°C), and in prospective validation covers enzyme stability (75.3%) and host growth temperatures (78.1%).

AI summary of the abstract · may contain errors · not medical advice · read the paper.