AVIDa-hIL6: A Large-Scale VHH Dataset Produced from an Immunized Alpaca for Predicting Antigen-Antibody Interactions

Hirofumi Tsuruta, Hiroyuki Yamazaki, Ryota Maeda, Ryotaro Tamura, Jennifer N. Wei, Zelda E. Mariet, Poomarin Phloyphisut, Hidetoshi Shimokawa, Joseph R. Ledsam, Lucy J. Colwell, Akihiro ImuraNeurIPS 2023

TL;DR

Researchers have released AVIDa-hIL6, the largest public dataset of its kind, containing over 573,000 antigen-antibody interaction pairs. By leveraging the unique biology of alpacas and high-throughput sequencing, the team provides a high-resolution map of how specific mutations in the human IL-6 protein affect the binding of "nanobodies" (VHHs). This massive repository offers a vital benchmark for machine learning models in therapeutic drug discovery.

What Problem Does This Paper Address?

Traditional antibody discovery is a slow, expensive "wet lab" process. While machine learning (ML) promises to accelerate this by predicting which antibodies will bind to a target, it faces a massive data bottleneck.

Existing datasets are often too small (a few thousand samples), lack "non-binder" examples (negative data), or omit the exact amino acid sequences of the antigens. Furthermore, viruses and bacteria constantly mutate to evade the immune system. We need models that can predict not just if an antibody binds to a "wild-type" protein, but how a single-point mutation might break that interaction. Until now, there wasn't enough high-quality data to train or even properly test such models.

Methods

The authors combined biological engineering with statistical rigor to create a massive, labeled dataset:

Overview of the AVIDa-hIL6 dataset generation pipeline.
  1. Alpaca Immunization: They immunized a single alpaca with a "cocktail" of human IL-6 (the wild-type protein plus 30 different artificial mutants). Alpacas produce VHHs (nanobodies), which are much simpler to sequence than human antibodies because they consist of only a single protein chain.
  2. Phage Display & NGS: They used phage display to "fish out" VHHs that bound to the targets and used Next-Generation Sequencing (NGS) to identify the exact amino acid sequences of hundreds of thousands of candidates.
  3. Statistical Labeling: Instead of a simple "yes/no" experimental result, they used a statistical Z-test to compare the frequency of a VHH before and after the binding process. If a VHH's proportion increased significantly, it was labeled a "binder."
  4. Noise Reduction: They developed a novel algorithm to filter out "noise"—VHHs that appeared to bind but were actually sticking to the experimental equipment or contaminating proteins.

Datasets

The primary contribution is the AVIDa-hIL6 dataset, which provides an unprecedented look at antibody-antigen interactions at scale.

Dataset statistics including sample counts, binding patterns, and sequence diversity.
  • Size: 573,891 antigen-VHH pairs (20,980 binders and 552,911 non-binders).
  • Diversity: Includes the wild-type human IL-6 and 30 mutants (generated via alanine scanning).
  • Comparison: It is roughly 10 times larger than previous public datasets like SAbDab or CoV-AbDab and is the first to provide exhaustive non-binder data alongside exact sequences for both sides of the interaction.

Key Insights and Ideas

  • VHHs as the "ML-Friendly" Antibody: Because VHHs are encoded by a single gene, researchers can perform exhaustive analysis on the order of $10^6$ sequences without the complex "cell cloning" required for conventional antibodies.
  • The Power of Negative Data: Most biological databases only record what works. By providing over 550,000 non-binding pairs, this dataset allows ML models to learn the "boundary" of what makes a successful interaction.
  • Mutation Sensitivity: The dataset captures "sensitive cases" where a single amino acid change in the antigen completely flips the binding status. This is the "gold mine" for training models to handle emerging viral variants.
  • Immune System as a Generator: The alpaca’s immune system acts as a natural "generative model," refining antibodies through a process called affinity maturation, which creates clusters of highly similar sequences that the authors used to verify label reliability.

Key Results

The authors benchmarked the dataset using several models, including AbAgIntPre (a state-of-the-art CNN-based model) and PIPR (a residual RCNN).

  • Generalization Gap: When models were trained only on the wild-type protein, they struggled to predict binding for unknown mutants (recalls were only ~57–68%).
  • Scaling Laws: As more mutant types were added to the training set, performance improved significantly. With 16 antigen types in the training set, precision reached ~95% and recall climbed to ~85%.
  • Model Comparison: Interestingly, complex specialized models like AbAgIntPre only slightly outperformed a simple Multi-Layer Perceptron (MLP), suggesting that current architectures aren't yet fully capturing the nuances of protein mutations and that AVIDa-hIL6 provides a challenging new frontier for model development.

Limitations and Future Work

  • Antigen Diversity: The dataset is currently limited to a single protein (IL-6). While the method can be applied to any protein, the current data doesn't help a model generalize to entirely different classes of antigens (e.g., a viral spike protein).
  • Artificial vs. Natural Mutations: The 30 mutants were created via alanine scanning (replacing an amino acid with alanine). Natural mutations are often more complex, involving substitutions between various amino acids.
  • Single Source: The data comes from one alpaca. Individual immune responses vary, and "self-tolerance" (the body avoiding making antibodies that attack its own tissues) might bias the available antibody pool.

The authors plan to expand the dataset to include other antigens, such as SARS-CoV-2 variants, to build a more universal foundation for AI-driven antibody discovery.

Summary updated

BibTeX
@inproceedings{tsuruta2023avida,
  author = {Hirofumi Tsuruta and Hiroyuki Yamazaki and Ryota Maeda and Ryotaro Tamura and Jennifer N. Wei and Zelda E. Mariet and Poomarin Phloyphisut and Hidetoshi Shimokawa and Joseph R. Ledsam and Lucy J. Colwell and Akihiro Imura},
  title = {{AVIDa-hIL6: A Large-Scale VHH Dataset Produced from an Immunized Alpaca for Predicting Antigen-Antibody Interactions}},
  booktitle = {Advances in Neural Information Processing Systems},
  year = {2023},
}
Figure