NanoBinder: a machine learning assisted nanobody binding prediction tool using Rosetta energy scores

Area of Science:

• Computational biology
• Protein engineering
• Machine learning in drug discovery

Background:

• Nanobodies are promising therapeutics due to their unique properties.
• Current computational tools for nanobody design, like Rosetta, suffer from high false-negative rates.
• This necessitates extensive, costly high-throughput screening and experimental validation.

Purpose of the Study:

• To develop an interpretable machine learning model, NanoBinder, for predicting nanobody-antigen binding.
• To reduce the false-negative rate in nanobody design and minimize experimental screening.
• To accelerate the development of novel nanobody therapeutics.

Main Methods:

• Developed NanoBinder, a Random Forest model trained on experimentally validated nanobody-antigen complexes.
• Integrated NanoBinder with Rosetta software, utilizing Rosetta energy scores as features.
• Employed SHAP summary plots for model interpretability to identify key binding features.

Main Results:

• NanoBinder accurately predicts non-binding nanobody-antigen pairs.
• The model demonstrates reasonable performance in identifying binders across diverse nanobody sets.
• Experimental validation on 49 nanobodies confirmed the model's predictive capabilities.

Conclusions:

• NanoBinder significantly enhances the accuracy of nanobody-antigen binding prediction.
• The model reduces the need for extensive experimental assays, saving time and resources.
• This approach offers a powerful tool to accelerate nanobody development and mitigate screening costs.