NanoMelt

A nanobody that unfolds at 55 °C can be a perfectly good binder and still useless as a reagent or a therapeutic, and the only way to find out is to express it and run a melting curve. That makes thermostability an obvious target for sequence-based prediction — except that the public record holds only a few hundred nanobody melting temperatures, gathered by different laboratories using nanoDSF, circular dichroism and dye-based DSF, whose numbers do not agree with one another. Fine-tuning a protein language model on labels that scarce and that inconsistent overfits, and scoring it on a held-out slice of the same small set says more about the split than about the model.

NanoMelt takes the opposite route: fine-tune nothing. Each sequence is encoded by several frozen representations, a small classical regressor is fit on each one, and a ridge regression meta-model stacks their predictions into a single apparent melting temperature. The deployed ensemble uses four views of the sequence — embeddings from ESM-1b and ESM-2, a one-hot encoding over 149 AHo-aligned positions, and VHSE physicochemical descriptors — so no single representation carries the prediction alone. The two generic protein language models beat every antibody- or nanobody-specific alternative tested, including AbLang and AntiBERTy.

It comes from Pietro Sormanni's group at the Centre for Misfolding Diseases, University of Cambridge, with collaborators at the German Cancer Research Center and Chimeris UK, and was published in mAbs in January 2025 after a September 2024 preprint. A single operational model, fit on the whole labelled set, ships as a command-line tool and a webserver.

#Key Features

• Ridge stacking over four sequence views: Gaussian process and support vector regressors trained on individual embeddings feed a ridge meta-model, which resists regression toward the dataset mean better than simple averaging.
• Evaluation built against small-data bias: performance comes from repeated nested cross-validation stratified on both experimental technique and sequence-similarity cluster, so no method or family dominates a fold.
• A released measurement set: the 640-nanobody dataset — assembled from NbThermo and the literature plus 129 new in-house nanoDSF measurements — ships with the code as a CSV.
• Prospective wet-lab validation: of six nanobodies mined from NGS repertoires, the three predicted stable expressed well and measured within 1.7 °C of prediction, while the three predicted unstable gave no detectable soluble product.
• Retrainable on your own data: a training command refits the whole pipeline on a user-supplied set of sequences and melting temperatures.

#Technical Details

Eight sequence representations were crossed with seven regression algorithms, and the best regressor per representation ranked by Spearman correlation on held-out folds: ESM-1b with an SVR led at ρ = 0.816, then ESM-2 (0.799), VHSE (0.797) and one-hot (0.784), with every antibody- or nanobody-specific embedding trailing below 0.76. Stacking the top four reaches a Pearson correlation of 0.853, Spearman 0.832, mean absolute error 4.1 °C and a standard deviation ratio of 0.86; adding more base models helped only marginally at a large inference cost. The 640 training sequences span 26.6–98.2 °C, measured by nanoDSF (251), circular dichroism (205) and sypro-orange DSF (165), at 64% mean pairwise identity. The 4.1 °C error should be read against the data's own noise floor: the 82 sequences with repeat measurements differ by 2.4 °C on average. Inference takes roughly 250 seconds per 1,000 sequences on an RTX 8000 GPU.

#Applications

The intended use is triage during nanobody discovery. Camelid immunization or naive-library panning followed by next-generation sequencing yields far more VHH candidates than any laboratory can express, and NanoMelt scores them in bulk so expression capacity goes to the ones likely to fold. On 83 external nanobody sequences it reached r = 0.751 and a mean absolute error of 4.8 °C, and it ranked point mutants of a single parent as reliably as it ranked diverse families — the regimes that matter for library screening and for humanization campaigns. It runs as a pip-installable command-line tool or through a free-registration webserver.

#Impact

NanoMelt's clearest contribution is showing how far general-purpose protein language models carry a specialized biophysical task when labels are scarce. Broad protein-stability tools fail on this data: the force-field method FoldX showed no significant correlation with measured melting temperatures on the 46 nanobodies with crystal structures, where NanoMelt reached r = 0.702, and the transformer predictor DeepSTABp managed r = 0.267 across the full set. The limits are worth stating: the target is an apparent melting temperature aggregated across three experimental techniques rather than a thermodynamic quantity, training and testing covered only VHH domains, headline accuracy comes from cross-validation rather than a large independent test set, and the paper itself cautions that evaluating on a dozen sequences will produce a correlation almost anywhere between 0 and 1. Code, model and data are released under a non-commercial license.