Nanobodies have a promising future in biotherapeutics because, at a mere 12 to 15 kDal, they excel at penetrating tissue, neutralizing pathogen, and blocking select proteins. Delivering on that promise, however, requires scientists to accurately predict nanobody melting temperatures.
A computational approach developed by research biologists Jerome A. E. Alvarez, PhD, and Scott N. Dean, PhD, both at the U.S. Naval Research Laboratory, rather accurately estimates those melting temperatures based upon their sequence rather than through lab experiments, according to a recent paper.
Called TEMPRO (temperature estimation model using protein embeddings), “This method could provide an inexpensive alternative to estimating an antibody’s thermostability from its sequence alone,” Dean tells GEN. A researcher could input an antibody’s sequence into this pre-trained nanobody melting model, “and receive an estimated melting temperature with high confidence. This could reduce the number of candidate antibodies that need to be produced.”
Designed for nanobodies
In the lab, TEMPRO predicted the melting temperature of single domain antibodies with a mean absolute error of 4.03°C and a root mean squared error of 5.66°C. That’s significant, he says, particularly in light of a nanobody melting temperature range of 43°C to 98°C.
“Some other predictors we evaluated had mean absolute errors of more than double that (greater than 10°C), which may not be sufficient to predict whether your candidate antibody has the desired stability,” Dean points out.
To achieve such accuracy, the scientists combined the NbThermo database of 548 nanobody sequences and their source organisms with their 166-sequence in-house, curated nanobody dataset. After filtering duplicates, 567 unique sequences remained. TEMPRO prediction accuracy was validated using the actual melting temperatures of antibodies and their variants from a database not used previously in this study.
What resulted is a detailed correlation among biophysical and physicochemical properties. It showed that protein embeddings (numerical representations of protein properties and functions) are the element most closely correlated to nanobody thermostability. Embeddings also can identify the amino acids responsible for changes in melting temperatures.
Somewhat surprisingly, they note, hydrophobicity, solvent accessibility, and instability indices—which typically factor into nanobodies’ desirability as therapeutics—did not predict thermostability.
These results show that protein embeddings are reliable, highly-accurate predictors of nanobody melting temperatures. As such, they can help manufacturers reduce nanobody optimization and production costs.
The public data, pre-trained models, and code with dependency documentations are freely available on Alvarez’s and Dean’s GitHub page. “They may be incorporated into other’s development pipelines,” Dean adds.
