StableProt: Structure-Aware Deep Learning for Protein Thermostability (Tm) and Environmental Adaptation (OGT) Prediction
- 1Institute of Life Sciences
Abstract
Predicting protein melting temperature (Tm) and organismal optimal growth temperature (OGT) is essential for enzyme engineering, biocatalyst design, and metagenomic bioprospecting, yet computational prediction faces several fundamental bottlenecks: sequence-only representations miss tertiary packing; dual-task coupling induces destructive gradient conflict; natural database skew (>87% mesophiles) drives mesophilic collapse; and existing tools produce uncalibrated point predictions. To address these challenges, we present StableProt, a structure-aware framework that couples amino acid sequences with Foldseek 3Di structural tokens via a frozen SaProt backbone to capture 3D contacts. Disjoint heads condition thermodynamic stability on ecological context without gradient conflict, while targeted 14% mesophilic subsampling expands thermophile representation to eliminate mesophilic collapse. Calibrated ensemble modeling reduces calibration error to 1.92%, providing dependable 95% confidence intervals. Evaluating models across 10{degrees}C temperature intervals reveals severe baseline collapse at thermal extremes, whereas StableProt maintains robust accuracy across regimes. On ProThermDB and zero-shot FireProtDB, StableProt achieves state-of-the-art probabilistic accuracy (CRPS = 4.52{degrees}C and 8.71{degrees}C) while recovering 92.0% of hyperthermophiles ([≥]80{degrees}C), and in prospective validation, its calibrated intervals capture both enzyme stability (75.3% coverage) and host environmental growth temperatures (78.1% coverage). StableProt is freely accessible as an interactive web server at https://stableprot.dixitlab.org/.
Related
- Understanding environmental adaptation in intermediate wheatgrass
- Rapid trait adaptation has contrasting effects on different dimensions of stability under environmental change
- BZR1 and BES1: From Brassinosteroid Signaling Termini to Multifunctional Integrators of Plant Growth and Environmental Adaptation
- Convergent genome- and gene-level constraints shape repeated environmental adaptation in grasses
- Population Genomic Insights into Genetic Differentiation and Environmental Adaptation of the Asian Honey Bee (Apis cerana) in Sichuan, China.
BibTeX
@misc{behera2026stableprot,
title = {{StableProt: Structure-Aware Deep Learning for Protein Thermostability (Tm) and Environmental Adaptation (OGT) Prediction}},
author = {B. P. Behera and A. Dixit},
howpublished = {bioRxiv},
year = {2026},
month = nov,
doi = {10.64898/2026.11.06.764833},
url = {https://www.biorxiv.org/content/10.64898/2026.11.06.764833v1},
note = {CC BY-NC-ND},
}
RIS
TY - UNPB
TI - StableProt: Structure-Aware Deep Learning for Protein Thermostability (Tm) and Environmental Adaptation (OGT) Prediction
AU - B. P. Behera
AU - A. Dixit
PY - 2026
DA - 2026/11/06
PB - bioRxiv
DO - 10.64898/2026.11.06.764833
UR - https://www.biorxiv.org/content/10.64898/2026.11.06.764833v1
AB - Predicting protein melting temperature (Tm) and organismal optimal growth temperature (OGT) is essential for enzyme engineering, biocatalyst design, and metagenomic bioprospecting, yet computational prediction faces several fundamental bottlenecks: sequence-only representations miss tertiary packing; dual-task coupling induces destructive gradient conflict; natural database skew (>87% mesophiles) drives mesophilic collapse; and existing tools produce uncalibrated point predictions. To address these challenges, we present StableProt, a structure-aware framework that couples amino acid sequences with Foldseek 3Di structural tokens via a frozen SaProt backbone to capture 3D contacts. Disjoint heads condition thermodynamic stability on ecological context without gradient conflict, while targeted 14% mesophilic subsampling expands thermophile representation to eliminate mesophilic collapse. Calibrated ensemble modeling reduces calibration error to 1.92%, providing dependable 95% confidence intervals. Evaluating models across 10{degrees}C temperature intervals reveals severe baseline collapse at thermal extremes, whereas StableProt maintains robust accuracy across regimes. On ProThermDB and zero-shot FireProtDB, StableProt achieves state-of-the-art probabilistic accuracy (CRPS = 4.52{degrees}C and 8.71{degrees}C) while recovering 92.0% of hyperthermophiles ([≥]80{degrees}C), and in prospective validation, its calibrated intervals capture both enzyme stability (75.3% coverage) and host environmental growth temperatures (78.1% coverage). StableProt is freely accessible as an interactive web server at https://stableprot.dixitlab.org/.
ER -