Machine learning enables efficient and effective affinity maturation of nanobodies

Area of Science:

• Protein engineering
• Computational biology
• Immunology

Background:

• Antibodies bind targets with high affinity due to large interaction surfaces.
• In vitro antibody selection often yields modest affinities, requiring affinity maturation.
• Affinity maturation is laborious, involving multiple rounds and strategies.

Purpose of the Study:

• To accelerate the discovery of optimized antibody binders.
• To investigate machine learning applications on yeast-display data for antibody engineering.
• To predict high-affinity antibody sequences from single-round selection data.

Main Methods:

• Yeast-display affinity maturation campaigns.
• Machine learning analysis of sequencing data from single sorting rounds.
• Comparison of linear models, deep neural networks, and semi-supervised approaches.
• Design and selection of nanobody binders using predictive models.

Main Results:

• Sparse sequencing data from one sort can predict sequences enriched after multiple rounds.
• Linear models outperform deep neural networks and semi-supervised methods in ranking affinity-enhancing substitutions.
• Linear models provide interpretable insights into residue preferences for engineering.
• Developed optimized nanobodies against relaxin family peptide receptor 1 (RXFP1), including sub-nanomolar binders.

Conclusions:

• Machine learning, particularly linear models, can significantly accelerate antibody affinity maturation.
• Predictive modeling based on limited data offers an efficient alternative to traditional multi-round optimization.
• This approach enables rapid engineering of high-affinity binders for therapeutic targets.