Confidence scoring for deep learning-predicted antibody-antigen complexes: AntiConf as a precision-driven metric.

Area of Science:

• Computational Biology
• Structural Biology
• Drug Discovery

Background:

• Accurate antibody-antigen (Ab-Ag) complex structure determination is vital for developing new therapeutics.
• Deep learning models, like AlphaFold2 (AF2), have advanced protein structure prediction, but their application to Ab-Ag complexes requires further investigation.
• Optimal modeling strategies and the reliability of confidence scores for Ab-Ag complex prediction are active research areas.

Purpose of the Study:

• To evaluate and compare the performance of nine state-of-the-art deep learning methods for Ab-Ag complex structure prediction.
• To analyze the impact of recycling iterations on prediction accuracy across different models.
• To assess the utility of existing model confidence scores and introduce a novel metric, AntiConf, for improved Ab-Ag complex assessment.

Main Methods:

• A curated dataset of 200 antibody-antigen complexes was used for performance evaluation.
• Nine deep learning methods were tested: AF2, Boltz-1, Boltz-1x, Boltz-2, Chai-1, Protenix, Protenix-1, OpenFold3, and ESMFold.
• Model confidence scores, including pDockQ2 and predicted Template-Modeling (pTM), were analyzed, and a new metric, AntiConf, was developed by integrating them.

Main Results:

• Protenix-1 demonstrated the highest performance, followed closely by Chai-1 and AF2, across multiple success metrics.
• Recycling iterations positively impacted AF2, Chai-1, and Protenix variants but not Boltz variants.
• The novel AntiConf metric significantly improved precision and recall for all tested methods, outperforming individual confidence scores.

Conclusions:

• Protenix-1 and Chai-1 are highly effective deep learning tools for Ab-Ag complex modeling.
• The developed AntiConf metric offers a robust post-prediction scoring system, enhancing the reliability of computational Ab-Ag complex predictions.
• AntiConf has the potential to guide experimental validation and improve future deep learning architectures like AF3 for Ab-Ag complex prediction.