Analysing open-source protein folding models for nanobody binding prediction

Area of Science:

• Computational biology
• Structural biology
• Immunology

Background:

• Antibody-based therapeutics are a rapidly growing class of treatments.
• Computational pre-filtering using protein structure prediction models can reduce screening costs.
• The relationship between model confidence and functional binding is not fully understood.

Purpose of the Study:

• Evaluate open-source protein structure prediction models for in silico screening of nanobody-antigen interactions.
• Assess the suitability of model confidence measures for predicting binding properties.

Main Methods:

• Benchmarked Boltz-2, Chai-1, IntFold, and AlphaFold3 on nanobody-antigen binding prediction.
• Used alanine substitution to test detection of out-of-distribution sequences.
• Assessed confidence measure sensitivity via mutagenesis and evaluated generalization.

Main Results:

• Boltz-2 confidence measures showed the highest performance in identifying true binders.
• Local confidence measures (pLDDT, interface/CDR-focused) were best for detecting mutations and out-of-distribution sequences.
• No single measure excelled at both tasks; models had limited generalization to new antigens.

Conclusions:

• Robust in silico nanobody selection requires combining complementary global and local confidence measures.
• Practical guidance is provided for integrating protein structure prediction into AI-driven nanobody discovery.
• Improved generalization across antigens is needed for these models.