AbPair

1 Introduction

AbPair is an artificial intelligence tool trained on advanced AI algorithms and a vast dataset of natural antibody pairing sequences, designed to predict the natural pairing scores of heavy and light chains. It efficiently evaluates the pairing potential between heavy and light chains, providing a scientific basis for antibody design and optimization.

It was trained using a non-redundant dataset of 1.67 million real paired human sequences from the Observed Antibody Space¹(OAS)database, along with 5 million human sequences for incorrect pairings. The test set metrics achieved an AUC > 0.992 and an ACC > 0.945. The input requirement is complete human VH and VL sequences; otherwise, it may affect the accuracy of the predictions.

2 Performance

Table 1. Metrics of AbPair on test set.

Figure 1. Confusion matrix of AbPair on test set.

Figure 2. Reliability curve of the model on test set.

Figure 3. ROC Curve and Precision-Recall Curve of AbPair on test set.

3 Parameters

• Heavy Chain Sequence：Complete Fv sequence of the human heavy chain.
• Light Chain Sequence：Complete Fv sequence of the human light chain.

4 Results Explanation

• Score: Natural pairing score of heavy and light chains.

5 Reference

[1] Kovaltsuk, A. et al. Observed Antibody Space: A Resource for Data Mining Next-Generation Sequencing of Antibody Repertoires. The Journal of Immunology 201, 2502–2509. https://doi.org/10.4049/jimmunol.1800708