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

StableProt is a novel structure-aware deep learning framework designed to predict protein melting temperature (Tm) and organismal optimal growth temperature (OGT). This advancement in computational prediction tackles key challenges in enzyme engineering, biocatalyst design, and metagenomic bioprospecting. Unlike traditional sequence-only representations, StableProt incorporates tertiary packing information by coupling amino acid sequences with Foldseek 3Di structural tokens through a frozen SaProt backbone.

This unique approach prevents destructive gradient conflicts caused by dual-task coupling and eliminates the issue of natural database skew favoring mesophilic organisms. The framework's two disjoint heads enable the conditioning of thermodynamic stability on ecological context, eliminating gradient conflict and improving prediction accuracy.

To address the issue of mesophilic collapse, StableProt employs targeted 14% mesophilic subsampling to expand thermophile representation, thereby recovering 92.0% of hyperthermophiles at temperatures greater than or equal to 80°C. Calibrated ensemble modeling further refines predictions, reducing calibration error to just 1.92%, and providing reliable 95% confidence intervals.

Evaluating StableProt across a range of temperatures highlights its robust accuracy, maintaining high precision even at thermal extremes. In comparison with existing tools, StableProt demonstrates superior performance on ProThermDB and zero-shot FireProtDB, achieving state-of-the-art probabilistic accuracy with CRPS values of 4.52°C and 8.71°C.

Moreover, its calibrated intervals effectively capture the environmental growth temperatures of both enzymes and hosts, showcasing its versatility and reliability. StableProt is now freely accessible as an interactive web server at https://stableprot.dixitlab.org/.