Predicting antibody-antigen affinity with a dual-level representation model.

Area of Science:

• Computational biology
• Structural bioinformatics
• Machine learning in protein science

Background:

• Accurate antibody-antigen interaction modeling is crucial but challenging, especially with limited structural data.
• Existing sequence-based methods often fail to fully leverage protein sequence information.
• This limitation impacts diverse antibody formats, including single-domain antibodies (sdAbs).

Purpose of the Study:

• To develop a novel deep learning framework for accurate sequence-based antibody affinity prediction.
• To address the limitations of current methods in exploiting sequence information for diverse antibody formats.
• To provide a robust computational tool for predicting antibody-antigen binding affinity.

Main Methods:

• Proposed DLP-Affinity, a dual-level deep learning framework.
• Integrated two modules: Residue-to-Residue (R2R) for local interface contacts and Global Stochastic Projection Embedding (GSPE) for global protein properties.
• Utilized a fine-tuned protein language model for enhanced feature representation.

Main Results:

• Achieved state-of-the-art performance on the AB-Bind dataset, reducing mean absolute error by up to 20.9%.
• Demonstrated highly competitive results on the sdAb-DB dataset.
• Showcased the framework's effectiveness in sequence-based affinity prediction.

Conclusions:

• DLP-Affinity offers a robust and accurate solution for sequence-based antibody affinity prediction.
• The dual-level deep learning approach effectively captures both local and global protein features.
• The framework shows promise for applications involving diverse antibody formats and limited structural data.