PPTstab

Two proteins sharing most of their sequence usually melt at similar temperatures, so a melting-temperature predictor trained on a redundant set can post a strong held-out score having learned little beyond "find the nearest homolog." Most published Tm predictors were built that way. PPTstab starts from the opposite constraint: its 17,312 training proteins were clustered with CD-HIT at 40% identity, so a validation score has to come from something other than memorized neighbors.

The second gap it closes is direction. A predictor tells you what a sequence's Tm is; it does not tell you which residue to change to raise it. PPTstab pairs its regressor with an exhaustive scan: given a sequence, the Design job enumerates every single-point substitution at every position, scores each with the same fixed model, and ranks them, so a user can pick the mutation that lands nearest a desired Tm. This is scoring rather than generation — the model proposes no sequence it was not handed — but it turns a passive predictor into something an engineer can act on.

PPTstab was developed by Purva Tijare, Nishant Kumar and Gajendra P. S. Raghava at Indraprastha Institute of Information Technology Delhi, posted to bioRxiv in September 2024 and published in Scientific Reports in May 2025. Its default feature encoder is ProtBert, used frozen.

#Key Features

• Frozen language model as a static encoder: ProtBert is used as a fixed feature extractor with no fine-tuning; the released extractor loads the public Rostlab/prot_bert checkpoint and mean-pools its final hidden layer into a 1024-dimensional vector per sequence.
• Assay context is an input, not an assumption: PPTstab takes a required binary flag for whether the target Tm is a cell or a lysate measurement and appends it to the feature vector. The same sequence returns two different Tm values depending on which one you set.
• Three interchangeable feature routes: embeddings (the default), Shannon entropy per residue, and amino-acid composition each ship with their own trained regression head, so the tool runs without a GPU when compositional features suffice.
• Exhaustive single-point mutant scanning: the Design job scores every single-residue analog of an input sequence, reporting each mutant's predicted Tm rather than emitting one suggestion.
• Proteome-scale screening: the standalone package runs over whole genomes and metagenomes to filter for thermostable proteins.

#Technical Details

The dataset descends from DeepSTABp's 35,114 sequences, drawn largely from the Meltome Atlas. After removing sequences over 2,500 residues or containing non-standard amino acids and applying CD-HIT at 40%, 17,312 proteins remained, split 13,849 for training and testing against 3,463 held out for validation. The deployed model concatenates the 1024-dimensional ProtBert embedding with the condition flag for a 1026-feature input, then averages two regressors: a TensorFlow network with four dense hidden layers of 256, 128, 64 and 32 ReLU units trained under MSE loss with Adam, and a scikit-learn MLP regressor. Predictions are denormalized against a fixed 30–90 °C range.

On the validation set that ensemble reaches a Pearson correlation of 0.89 and an R² of 0.80, with an RMSE of 4.11 °C and a mean absolute error of 3.00 °C. The best hand-engineered feature, Shannon entropy across all residues, reaches 0.80 and 0.63 on the same split; amino-acid composition reaches 0.77. Embeddings from protT5MLM matched ProtBert at 0.89, and ProtBert was chosen for its lower compute cost; Ankh reached 0.83 and ProtGPT2 0.82. A hybrid of entropy features and ProtBert embeddings did not improve on embeddings alone and was dropped.

#Applications

The direct use is triage: score a candidate enzyme before committing to expression and a thermal shift assay, or rank a mutant library so wet-lab effort goes to the substitutions most likely to raise Tm. Thermostability is the practical constraint on enzymes used in biofuel production, food processing and pharmaceutical manufacturing, and on heat-stable vaccines and diagnostics. The authors demonstrate the screening mode on the reviewed proteomes of Psychrobacter frigidicola (2,298 proteins), Shewanella oneidensis (4,068) and Thermus thermophilus (2,227), recovering the expected shift toward higher predicted Tm in the thermophile.

#Impact

PPTstab's contribution is a non-redundant benchmark for sequence-only Tm regression plus a usable design front end, delivered as a web server, a Python package and a repository carrying the trained heads and both dataset splits. Its limits follow from its construction. The training labels come from mass-spectrometry thermal proteome profiling, so the model inherits that assay's species coverage and its cell-versus-lysate distinction rather than measuring thermodynamic stability directly. A 3 °C mean absolute error is useful for ranking but coarse for absolute claims, and the hard-coded 30–90 °C output range clamps the extremes. Because the design module enumerates only single substitutions and scores them independently, it cannot see epistasis, and its rankings are in-silico throughout — the paper reports no experimental validation of any designed variant.