StableProt

Structure-aware Tm and OGT prediction with a calibrated interval

Tm and OGT from sequence, with an interval

StableProt couples SaProt 3Di conformational tokens with disjoint heads for melting temperature (Tm) and optimal growth temperature (OGT). It returns a calibrated predictive interval rather than a point estimate. Training and every evaluation split are separated by a bidirectional homology audit (<30% sequence identity), with mesophilic downsampling on the OGT set.

In distribution (ProThermDB)
n = 3,340 decontaminated holdout targets. Point MAE 6.16 °C with continuous ranked probability score (CRPS) of 4.52 °C (Pearson r = 0.788, Spearman ρ = 0.517). Hyperthermophile screening recall (≥80 °C) reaches 92.0%.
Out of distribution (FireProtDB)
n = 322 zero-shot targets sharing <30% identity to training. Achieves MAE 11.85 °C and CRPS 8.71 °C (Pearson r = 0.432, Spearman ρ = 0.350), maintaining 81.8% detection recall on hyperthermophiles (≥80 °C).
Calibrated Uncertainty
Empirical variance scaling reduces expected calibration error (ECE) from 14.4% to 1.9% with 67.5% empirical central-interval coverage. Reports ±1σ calibrated predictive intervals (~68% nominal coverage) reflecting aleatoric and epistemic uncertainty.
0 AA
0 AA

What this server runs

StableProt is a structure-aware model for protein melting temperature (Tm) and organismal optimal growth temperature (OGT). The Tm and OGT heads are disjoint: the OGT prior enters the Tm path only. Predictions average a 5-seed ensemble and include a scaled predictive interval.

Data and evaluation

Training corpus comprises 28,739 clean Tm records (curated from 29,300 after permanently purging 561 sequences sharing ≥30% identity with evaluation constructs) and 940,000 curated OGT records with mesophilic downsampling. Held-out evaluation benchmarks include ProThermDB (n = 3,340), FireProtDB (n = 322), and BRENDA / BacDive OOD test sets under strict bidirectional homology decontamination (<30% sequence identity).

On bin-balanced OGT scoring across the thermal spectrum, StableProt maintains robust accuracy across all environmental regimes without degradation outside the mesophilic band. On FireProtDB single-point mutations (n = 3,649 physical experimental mutations), ΔTm MAE is 5.05 °C with 56.8% directional classification accuracy (ROC AUC = 0.585). As a sequence-level stability model not trained on mutational assay differentials, its predicted interval widens appropriately to reflect single-mutation uncertainty.

Loop design tab

Loop highlighting uses Chou–Fasman propensities, then re-scores Tm for edited sequences. It is a sequence-side aid, not a wet-lab-validated design method.

Institute of Life Sciences (ILS) Logo

Affiliation

Computational Biology and Bioinformatics Laboratory, iBRIC–Institute of Life Sciences (ILS), Bhubaneswar, Odisha, India.

Contact: bibhu.prasad@ils.res.in and anshuman@ils.res.in.

Source: github.com/Bibhuprasadbehera/StableProt

Use

This server is for academic and non-commercial research. Commercial use needs written permission from the authors. Predictions are provided as-is, with no warranty. They are not a clinical, diagnostic, or wet-lab result. The Design tab is a sequence-side hypothesis aid.

The served model is a 5-seed ensemble. The half-width shown is ±1σ (about 68% coverage in distribution), not a 95% confidence interval.

Privacy

Sequences are sent to this server only to run inference. The application does not write submitted sequences to disk. The host may keep ordinary web-server access logs (time, IP, path) that do not include the sequence body.

The page loads fonts from Google Fonts, which is a third-party request from your browser.

How to cite

No journal citation yet. Until then, cite the working manuscript and the repository:

Behera, B. P. & Daxit, A. StableProt: structure-aware deep learning for protein thermostability (Tm) and environmental adaptation (OGT) prediction. Working manuscript. https://github.com/Bibhuprasadbehera/StableProt