KyDab - a comprehensive database of antibody discovery selection campaigns.

KyDab is a comprehensive, publicly accessible database that curates over 120,000 paired antibody sequences and full-funnel selection data from standardized Kymouse immunization studies to support the development and evaluation of artificial intelligence models for antibody discovery.

The Gist

Imagine you are trying to find the perfect key to unlock a specific door (a disease). In the world of medicine, these "keys" are antibodies, and finding the right one usually involves a massive, expensive, and time-consuming treasure hunt.

For a long time, the "maps" we had for this treasure hunt were incomplete. They only showed us the keys that already worked and made it to the final lock. They didn't show us the thousands of keys that were too big, too small, or just didn't fit at all. Without seeing the failures, it's very hard for computers (Artificial Intelligence) to learn how to design the perfect key from scratch.

Enter KyDab: The "Full-Story" Map.

This paper introduces KyDab (Kymouse Antibody Database), a new, open-source library that changes the game. Here is how it works, broken down simply:

1. The Lab Mouse "Factory"

The researchers used a special type of mouse called a Kymouse. Think of these mice as tiny, biological 3D printers. When you give them a piece of a virus or bacteria (an immunogen), their immune systems automatically start printing millions of different antibody "keys" to fight it. These mice are engineered so that the keys they print are already human-friendly, making them perfect for making medicines.

2. The "Full-Funnel" Collection

Usually, when scientists do this, they test thousands of keys, throw away the bad ones, and only publish the few winners. It's like showing a movie but only releasing the final 5 minutes where the hero wins.

KyDab is different. It releases the entire movie, from the opening scene to the credits.

• The Good: It includes the winning keys (antibodies that bind well).
• The Bad: It includes the losing keys (antibodies that didn't work).
• The Data: It has over 120,000 pairs of antibody sequences and details on how they were tested.

3. Why This Matters for AI

Imagine you are teaching a robot to bake the perfect cake.

• Old Way: You only show the robot pictures of cakes that turned out perfectly. The robot tries to copy them but doesn't understand why a cake might fail (e.g., "Oh, I forgot the eggs, that's why it's flat").
• KyDab Way: You show the robot pictures of perfect cakes and burnt cakes, flat cakes, and cakes that fell apart. You tell the robot exactly what ingredients were used for each.

Because KyDab includes the "failures" (negative data) and the "successes" (positive data) in a consistent, organized way, AI models can finally learn the real rules of antibody design. They can learn to predict which keys will fit before a human even picks up a test tube.

4. The "Library" Analogy

Think of previous databases as a library that only has the Bestsellers section. Everyone knows what's in there, but it doesn't help you write a new story.

KyDab is like a library that includes:

• The Bestsellers.
• The manuscripts that were rejected by editors.
• The drafts with typos.
• The notes on why the editor rejected them.

This gives writers (scientists and AI) a much deeper understanding of what makes a story (or an antibody) successful.

The Bottom Line

The authors of this paper are essentially saying: "We built a massive, organized database of our entire antibody discovery process, including the mistakes. We are giving it away for free so that Artificial Intelligence can learn faster, cheaper, and better."

This could lead to new medicines for cancer, viruses, and other diseases being discovered much faster than ever before, because the computers will finally have the "full story" to learn from.

Technical Summary

1. Problem Statement

The field of therapeutic antibody discovery is increasingly relying on Artificial Intelligence (AI) for virtual screening, affinity maturation, and generative design. However, the development and benchmarking of these AI models are hindered by significant data gaps in existing public resources:

• Lack of "Full-Funnel" Data: Existing databases (e.g., Observed Antibody Space, SAbDab, Thera-SAbDab) primarily curate mature, clinically advanced antibodies or broad sequence collections. They fail to capture the industrial discovery workflow, which involves large pools of clones, multiple down-selection stages, and both positive and negative experimental outcomes.
• Scarcity of Affinity-Resolved Data: There is a critical shortage of experimentally validated binding data (affinity measurements). Most public resources lack negative screening data (non-binders), which is essential for training robust machine learning models to distinguish true binders from false positives.
• Inconsistency: Existing datasets often suffer from heterogeneous curation methods, making it difficult to compare results across different studies or platforms.

2. Methodology

To address these gaps, the authors introduced KyDab (Kymouse Antibody Database), a curated resource generated from the Kymouse humanized mouse platform. The methodology involved:

• Data Source: 11 distinct immunization studies using Kymouse transgenic mice (engineered to produce fully human variable regions).
• Immunogens: 51 unique antigens targeting viral, bacterial, and parasitic pathogens (e.g., SARS-CoV-2, Malaria, Pertussis, Influenza, Acinetobacter baumannii).
• Workflow Standardization:
  • Immunization & Screening: Mice were immunized, and sera were screened (e.g., via ELISA) to identify positive responders.
  • Single-Cell Sorting: Antigen-specific B cells were isolated from spleen, lymph nodes, or bone marrow using flow cytometry.
  • Sequencing & Processing: Paired heavy-light chain (VH-VL) sequences were generated. Raw data was processed using an in-house bioinformatics pipeline and re-annotated using RIOT (a sequence annotation tool) to ensure consistent IMGT scheme alignment and to strip commercially sensitive information.
  • Down-Selection: Clones were prioritized based on lineage expansion, somatic hypermutation, sequence convergence, and developability.
  • Affinity Characterization: A subset of selected clones was recombinantly expressed and tested for binding affinity using assays such as Surface Plasmon Resonance (SPR), HTRF, and ELISA.
• Data Curation: The database includes both positive (binders) and negative (non-binders) data, along with comprehensive metadata (mouse ID, tissue source, immunogen details).

3. Key Contributions

• First Large-Scale Full-Funnel Database: KyDab is the first public resource to provide end-to-end antibody discovery data, spanning from initial immune repertoire generation to final affinity characterization.
• Scale and Diversity: The current release contains 123,527 productive paired heavy-light chain sequences derived from 51 immunogens across 11 studies.
• Inclusion of Negative Data: Unlike most resources, KyDab explicitly includes negative screening outcomes and clones that failed to bind, which is crucial for training AI models to reduce false-positive rates.
• Standardized Metadata: All data is processed through a unified pipeline, ensuring high comparability across different campaigns and antigens.
• Accessibility: The database is publicly accessible via a dedicated portal (https://kydab.naturalantibody.com) featuring interactive visualizations and flexible download options.

4. Results

• Dataset Composition:
  • Total Sequences: >120,000 paired sequences.
  • Binding Data: 1,657 clones with experimentally measured binding data (using SPR, HTRF, or ELISA).
  • Coverage: Includes diverse targets such as Pertussis toxin, Malaria CSP, SARS-CoV-2 variants, RSV, Influenza, and bacterial toxins.
• Repertoire Diversity Analysis:
  • The authors performed sequence identity clustering (at 70%, 80%, and 90% thresholds) on CDR and Framework regions.
  • Key Finding: The Heavy Chain CDR3 region consistently exhibited the highest diversity across all datasets (often >50% unique clusters at 90% identity), confirming its role as the primary driver of antigen specificity.
  • Framework Regions: Showed the lowest diversity, as expected due to their conserved structural nature.
  • Light Chain: Displayed lower diversity compared to the heavy chain, contributing less to overall repertoire variability.
• Data Utility: The inclusion of both high-affinity binders and non-binders allows for the creation of balanced training sets for AI models, addressing the "selection bias" inherent in typical discovery pipelines.

5. Significance

• AI Model Development: KyDab provides the "ground truth" data necessary to train and validate next-generation AI models for antibody discovery. It enables the development of algorithms that can predict binding affinity and selectivity more accurately by learning from real-world industrial workflows.
• Benchmarking: It serves as a standardized benchmark for evaluating the performance of generative design and virtual screening tools against a realistic, full-funnel dataset.
• Community Impact: By releasing this data, the authors aim to foster collaboration between academia and industry, encouraging the sharing of structured discovery data to accelerate the development of robust, generalizable AI tools for therapeutic antibody discovery.
• Future Outlook: The database is designed to be continuously updated as new datasets become available, ensuring it remains a dynamic resource for the scientific community.

In summary, KyDab bridges the critical gap between theoretical AI capabilities and practical industrial antibody discovery by providing a high-quality, standardized, and comprehensive dataset that reflects the complexities and biases of real-world therapeutic development.