Thermostability Prediction Powered by Synergistic Deep Learning at Experimental and Theoretical Levels for Nanobodies

Nanobodies have emerged as powerful biorecognition elements
in
various fields, while thermostability is a key factor in practical
application. Experimental screening remains costly and low throughput,
while the scarcity of thermostability data poses a significant challenge
for machine learning application. In this work, we innovatively propose
a dual-scale synergistic deep learning strategy to improve prediction
reliability, including a NBsTem_Tm model trained on 514 experimental
melting temperature (Tm) data and a NBsTem_Q model
with a theoretical indicator inferred from conformation changes of
extensive MD simulations on 704 nanobody structures. Their synergy
can alleviate the scarcity of experimental data and the risk of low
generalization inherent in small-data NBsTem_Tm models to unseen samples.
The two models are constructed by integrating the antibody language
model into a joint deep learning architecture to sufficiently learn
the feature embedding at different levels. Consequently, NBsTem_Tm
achieves a Pearson value of 0.83 on the external test, significantly
outperforming three reported competitive models. NBsTem_Q obtains
accuracy of 0.84, also exhibiting applicable potential. In addition,
the two models can be applied for nanobodies with missing residues,
thus being robust to wide application. Benefiting from the two synergistic
models, a more reliable screening criterion (Tm >
65 °C and Qclass IV) is proposed for determining
highly thermostable nanobodies. The dual-scale framework coupled with
a proposed screening strategy is further applied to explore the INDI
database with tens of millions of unexplored nanobodies to fill the
absence of its thermostability property, discovering approximately
12% thermostable nanobodies. Finally, a user-friendly web server NBsTem
is developed to serve as a high-throughput screening platform for
nanobody design and development.