ParaAntiProt provides paratope prediction using antibody and protein language models

Area of Science:

• Immunology
• Computational Biology
• Bioinformatics

Background:

• Accurate paratope prediction is crucial for antibody design, cancer treatment, and personalized medicine.
• Traditional and existing machine learning methods often require 3D structures, are labor-intensive, or involve complex feature engineering.

Purpose of the Study:

• To develop a deep learning-assisted method for paratope identification using only amino acid sequences.
• To create an antigen-agnostic model that overcomes limitations of existing prediction techniques.

Main Methods:

• Utilized the ProtTrans architecture with pre-trained protein and antibody language models to extract sequence embeddings.
• Incorporated positional encoding for Complementarity Determining Regions (CDRs) to enhance structural understanding.
• Developed a deep learning model for paratope prediction based solely on amino acid sequences.

Main Results:

• Achieved high performance on benchmark datasets with 0.904 ROC AUC, 0.701 F1-score, and 0.585 MCC.
• Demonstrated strong performance in nanobody paratope prediction (0.912 ROC AUC, 0.665 PR AUC).
• Outperformed structure-based prediction methods with a 0.731 PR AUC.

Conclusions:

• The developed sequence-based deep learning method offers an efficient and accurate alternative for paratope prediction.
• This approach has significant potential for advancing antibody design, diagnostics, and therapeutic development.
• Ablation studies confirmed the impact of CDR positional encoding and language models on prediction accuracy.