henhaobio
Sign in

mAbsBioengineeringPeer-reviewed journal article

Prediction of protein biophysical traits from limited data: a case study on nanobody thermostability through NanoMelt

Aubin RamoniD, Mingyang Ni, Olga PredeinaiD, Rebecca Gaffey, Patrick Kunz, Shimobi C. Onuoha, Pietro SormanniiD

PDFDOIPreprintCodeData
In-depth full-text digest available · Jump to digest ↓

Abstract

1In-silico prediction of protein biophysical traits is often hindered by the limited availability of experimental data and their heterogeneity. Training on limited data can lead to overfitting and poor generalisability to sequences distant from those in the training set. Additionally, inadequate use of scarce and disparate data can introduce biases during evaluation, leading to unreliable model performances being reported. Here, we present a comprehensive study exploring various approaches for protein fitness prediction from limited data, leveraging pre-trained embeddings, repeated stratified nested cross-validation, and ensemble learning to ensure an unbiased assessment of the performances. We applied our framework to introduce NanoMelt, a predictor of nanobody thermostability trained with a dataset of 640 measurements of apparent melting temperature, obtained by integrating data from the literature with 129 new measurements from this study. We find that an ensemble model stacking multiple regression using diverse sequence embeddings achieves state-of-the-art accuracy in predicting nanobody thermostability. We further demonstrate NanoMelts potential to streamline nanobody development by guiding the selection of highly stable nanobodies. We make the curated dataset of nanobody thermostability freely available and NanoMelt accessible as a downloadable software and webserver. 2 Significance StatementRapidly predicting protein biophysical traits with accuracy is a key goal in protein engineering, yet efforts to develop reliable predictors are often hindered by limited and disparate experimental measurements. We introduce a framework to predict biophysical traits using few training data, leveraging diverse machine learning approaches via a semi-supervised framework combined with ensemble learning. We applied this framework to develop NanoMelt, a tool to predict nanobody thermostability trained on a new dataset of apparent melting temperatures. Nanobodies are increasingly important in research and therapeutics due to their ease of production and small size, which allows deeper tissue penetration and seamless combination into multi-specific compounds. NanoMelt outperforms available methods for protein thermostability prediction and can streamline nanobody development by guiding the design and selection of highly stable nanobodies during discovery and optimization campaigns.

In-depth summary

Claude Sonnet 4.6 · xhigh effort · 8 Nov 2026 · from the full text of version 2 (19 Sept 2024) · CC BY 4.0

NanoMelt, a ridge-stacked ensemble of regressors on four diverse sequence embeddings, predicts nanobody melting temperature from 640 curated sequences, reaching Spearman 0.832 under nested cross-validation.

Predictors of protein biophysical traits need large, consistent datasets, but public nanobody thermostability data are scarce and heterogeneous (NbThermo holds 548 datapoints). This work curates 640 melting temperatures and pairs them with nested cross-validation and ensemble stacking to use limited data without leakage.

Key results

  1. SuggestedOn artificial solubility labels, pre-trained protein language model embeddings generalised from far fewer sequences than one-hot or VHSE encodings. ESM-1b and ESM-2 began to plateau beyond 600 training sequences, which informed the target dataset size.

    At 640 training sequences (ridge, intrinsic CamSol score): Spearman 0.871 (ESM-1b) and 0.888 (ESM-2) vs 0.605 (one-hot); AntiBERTy 0.751, AbLang 0.793, nanoBERT 0.802; NanobodyBuilder2 plateaued at 0.352 with 8,000 sequences. · 4.1 Preliminary search of minimal fitness…

  2. ShownOn real Tm data, kernel regressors (SVR, GPR) on generic ESM-1b embeddings were best, beating antibody- and nanobody-specific embeddings, but they overfit strongly.

    SVR on ESM-1b: test Spearman 0.816 ± 0.027 vs 0.993 ± 0.004 on training; AntiBERTy (SVR) 0.745 ± 0.036; nanoBERT (GPR) 0.723 ± 0.035; VHSE (SVR) 0.797 ± 0.027; one-hot (GPR) 0.784 ± 0.029. · 4.3 Selection of regression models

  3. ShownRidge stacking of models trained on four different embeddings gave the final NanoMelt ensemble. Stacking was preferred over averaging because it regressed less toward the mean, and its performance was consistent across sequence clusters.

    Nested CV over 640 datapoints: Pearson 0.853, Spearman 0.832, MAE 4.1 °C, SDR 0.86. Eight-model stacking vs averaging: Spearman 0.835 vs 0.840, SDR 0.88 vs 0.81. Cluster 4 Pearson 0.741, cluster 9 0.843. · 4.4 Ensemble learning

  4. ShownNanoMelt clearly outperformed an empirical force field, a general Tm predictor and zero-shot language-model likelihoods for ranking nanobody thermostability.

    On 46 PDB-backed nanobodies NanoMelt r = 0.702 vs FoldX r = 0.033. On the full dataset NanoMelt r = 0.853 vs DeepSTABp r = 0.267 and ESM-2 sequence likelihood r = 0.337 (p ∼ 10^-18); AntiBERTy likelihood was also weakly correlated. · 4.4 Ensemble learning; 4.5 Benchmarking

  5. SuggestedIn a small prospective test on six NGS-derived nanobodies at least 30% dissimilar to the training data, predicted-stable candidates expressed with accurately predicted Tm, and predicted-unstable ones did not express.

    Nb4, Nb5, Nb6 (predicted Tm > 73 °C) expressed, with absolute errors of 1.7, 0.6 and 0.6 °C. Nb1, Nb2, Nb3 (predicted Tm < 61 °C) showed no expression; their Tm was not measured. · 4.6 Selection of highly stable nanobodies with…

Methods, data and reproducibility

Eight sequence embeddings (one-hot, VHSE, ESM-1b, ESM-2, AbLang, AntiBERTy, nanoBERT, NanobodyBuilder2) feed seven regressors under repeated stratified nested three-fold cross-validation; predictions from four embedding-specific models are combined by ridge stacking.

  • Curated nanobody apparent melting temperature dataset (Tm 26.6 to 98.2 °C), from in-house measurements, NbThermo, Kunz et al. and mutational studies · camelid nanobodies · n = 640 unique sequences
  • New nanoDSF measurements of nanobodies, under the same conditions as Kunz et al. · n = 129 new datapoints
  • NbThermo database, filtered for missing, unalignable and duplicate entries · n = 405 retained of 548
  • Re-fitted raw nanoDSF curves from Kunz et al. under uniform conditions · n = 64 nanobodies
  • Two mutational studies (Rosace et al.; unpublished destabilising study) · n = 42 sequences
  • Artificial dataset: CamSol solubility scores for nanobody sequences sampled from the AbNatiV training set; 80:20 train/test split · camelid nanobodies · n = 10,000 sequences (2,000 held-out test)
  • Benchmark subset of nanobodies with PDB crystal structures, used for the FoldX comparison · n = 46 nanobodies
  • Prospective validation: nanobodies mined from NGS camelid repertoires, expressed in mammalian cells · camelid nanobodies · n = 6 candidates (3 expressed)
Methods & tools
CamSol, ANARCI, ESM-1b and ESM-2, AntiBERTy, AbLang, nanoBERT, NanobodyBuilder2, FoldX, DeepSTABp, AbNatiV, scikit-learn, LightGBM
Assay conditions & compute
NanoMelt runs about 250 sec for 1,000 sequences on a single GPU (NVIDIA RTX 8000).
Reproducibility gaps
Webserver requires registering a free account. The 19-point destabilising dataset comes from unpublished work. Software versions are not stated in the text.
Limitations (5)
  • The dataset is noisy: most points were aggregated from the literature with varied protocols and analyses, and buffer, heating rate and concentration are unknown for most entries, so adding the method as an input did not help. · stated by the authors
  • Some nanobodies had errors near 20 °C. For example, the W37A mutant (measured Tm 26.6 °C) was predicted at 47.5 °C, likely because the mutation hits a hyper-conserved site. · stated by the authors
  • SVR and GPR showed a stronger tendency to overfit (SVR on ESM-1b: 0.993 training vs 0.816 test Spearman), which the authors expect from kernel fitting. Correlations on test sets of about a dozen sequences vary widely. · stated by the authors
  • Prospective validation covers only three measured Tm values; the three predicted-unstable nanobodies did not express, so their Tm predictions were not tested. No confidence intervals are given for the 46-nanobody comparison. · noted in this summary
  • Headline metrics differ between sections: r 0.853, ρ 0.832, MAE 4.1 °C, SDR 0.86 in the text vs r 0.862, ρ 0.845, MAE 3.9 °C, SDR 0.84 in the Fig. 5 legend, which complicates exact comparison. · noted in this summary

Read this if you build sequence-based regressors for protein properties from a few hundred noisy labels and want guidance on nested cross-validation, stacking and the SDR metric.

Citations
12
FWCI
2.98
Field percentile
top 8%

Reproducibility ledger

Nobody has reported trying to reproduce this paper yet. Did you replicate an experiment or run the code? Your report saves the next lab a week.

Sign in to file a reproducibility report.

Related papers

Similarity Map
SelectedSame fieldCross-field
THISMurakami '26Pailozian '26Tijare '25Murakami '26Rodella '24Ramos '25
75% semantic matchBiophysics
Taihei Murakami, Kentaro Sasaki, Soichiro Oda+2 more·bioRxiv·Aug 4, 2026