Bio-Inspired Mamba for Antibody-Antigen Interaction Prediction.

Area of Science:

• Biochemistry
• Computational Biology
• Immunology

Background:

• Antibody lead discovery is vital for immunotherapy development, necessitating identification of high-affinity binding candidates.
• Predicting antibody-antigen interactions (AAIs) and pinpointing binding sites (epitopes and paratopes) are critical but challenging tasks in computational biology.

Purpose of the Study:

• To introduce MambaAAI, a novel bio-inspired model utilizing the Mamba architecture for predicting AAIs and identifying binding sites.
• To leverage selective attention mechanisms for enhanced accuracy in AAI prediction and binding site localization.

Main Methods:

• Employed ESM-2, a pre-trained protein language model, to extract evolutionary representations from antibody and antigen sequences.
• Developed a dual-view Mamba encoder to learn embeddings of residue-level interaction matrices from both antibody and antigen perspectives.
• Utilized a multilayer perceptron for decoding learned embeddings to predict interaction probabilities.

Main Results:

• MambaAAI demonstrated marginally superior prediction accuracy compared to state-of-the-art baselines on large-scale neutralization datasets.
• The model exhibited robust generalization capabilities on unseen antibodies and antigens.
• Selective attention mechanism successfully identified critical epitope and paratope regions in SARS-CoV-2 antibody examples.

Conclusions:

• MambaAAI offers a significant advancement in predicting AAIs and identifying binding sites through dynamic selection of key residue sites.
• The model holds substantial potential for accelerating the discovery of lead immunotherapy candidates with reduced computational burden.
• MambaAAI's ability to uncover critical binding regions paves the way for more targeted drug development.