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

Area of Science:

• Biotechnology
• Computational Biology
• Structural Biology

Background:

• Nanobodies are valuable biorecognition tools, but their practical use is limited by thermostability.
• Experimental determination of nanobody thermostability is costly and low-throughput.
• Limited experimental data hinders machine learning applications for predicting thermostability.

Purpose of the Study:

• To develop a reliable and high-throughput method for predicting nanobody thermostability.
• To address the challenge of limited experimental data in machine learning models.
• To create a synergistic deep learning strategy for enhanced prediction accuracy.

Main Methods:

• A dual-scale synergistic deep learning strategy integrating two models: NBsTem_Tm (trained on experimental melting temperature data) and NBsTem_Q (using theoretical indicators from molecular dynamics simulations).
• Utilized an antibody language model within a joint deep learning architecture to learn feature embeddings at multiple levels.
• Developed a robust screening criterion (Tm > 65 °C and Qclass IV) for identifying highly thermostable nanobodies.

Main Results:

• The NBsTem_Tm model achieved a Pearson value of 0.83 on an external test set, outperforming existing models.
• The NBsTem_Q model demonstrated an accuracy of 0.84, showing applicable potential.
• The models successfully predicted thermostability for nanobodies with missing residues and identified ~12% thermostable nanobodies in the INDI database.

Conclusions:

• The proposed dual-scale deep learning framework effectively predicts nanobody thermostability, mitigating data scarcity issues.
• The developed NBsTem web server provides a user-friendly platform for high-throughput nanobody screening.
• This strategy significantly advances nanobody design and development by enabling efficient identification of thermostable candidates.