Related Experiment Video
Updated: Aug 6, 2026
12:28
Bacterial Inner-membrane Display for Screening a Library of Antibody Fragments
Published on: October 15, 2016
Benchmarking AlphaFold and related deep learning approaches for modeling antibody and TCR antigen recognition
Rui Yin1,2, Shayana Saravanakumar1,2, Shu Yuan Shi1,2
1Department of Cell Biology and Molecular Genetics, University of Maryland, College Park, MD 20742, USA.
Biorxiv : the Preprint Server for Biology
|July 17, 2026
Summary
New computational methods improve modeling of antibody and T cell receptor (TCR) recognition. AlphaFold3 and increased sampling show promise for predicting immune complexes, aiding biotherapeutic and vaccine design.
Area of Science:
- Structural biology
- Immunoinformatics
- Computational biophysics
Background:
- Understanding antigen recognition by antibodies and T cell receptors (TCRs) is crucial for developing effective immune-based therapies and vaccines.
- Accurate computational modeling of these interactions is challenging, with existing methods like AlphaFold showing limitations in predicting immune recognition complexes.
Purpose of the Study:
- To evaluate the performance of AlphaFold2, AlphaFold3, and other deep learning methods for modeling antibody-protein, antibody-peptide, and TCR-peptide-major histocompatibility complex (pMHC) recognition.
- To assess the impact of increased sampling protocols on predictive accuracy for immune recognition.
Main Methods:
- Comparative analysis of AlphaFold2, AlphaFold3, and deep learning models using increased sampling protocols.
- Assessment of prediction accuracy across different immune complex classes: antibody-protein, antibody-peptide, and TCR-pMHC.
- Evaluation of AlphaFold confidence scores and modeling of noncanonical complexes.
Main Results:
- AlphaFold3 and increased sampling protocols generally outperform AlphaFold2 and default sampling for immune complex modeling.
- Predictive accuracy varies significantly across interface types, with antibody-peptide complexes remaining a challenge.
- Model pooling approaches demonstrate potential to enhance success rates, e.g., increasing antibody-peptide near-native success from 41% to 59%.
Conclusions:
- Enhanced computational strategies, including AlphaFold3 and optimized sampling, offer improved capabilities for modeling antibody and TCR recognition.
- Key distinctions exist among protocols, scoring metrics, and immune complex classes, necessitating tailored approaches for predictive modeling.
- Further development is needed to address challenges in modeling antibody-peptide interactions for robust immune complex prediction.
