Resource: Introducing SNAC-DB: A New Resource for Antibody & NANOBODY® VHH-Antigen Modeling

Abhinav Gupta • July 25, 2025
Resource: Introducing SNAC-DB: A New Resource for Antibody & NANOBODY® VHH-Antigen Modeling
Predicting antibody and NANOBODY® VHH--antigen complexes remains a notable gap in current AI models, limiting their utility in drug discovery. We present SNAC-DB, an open-source, machine-learning-ready database and pipeline developed by structural biologists and ML researchers to address this challenge.

Key features of SNAC-DB include:

Expanded Coverage: 32% more structural diversity than SAbDab, capturing overlooked assemblies such as antibodies/nanobodies as antigens, complete multi-chain epitopes, and weak CDR crystal contacts.

ML-Friendly Data: Cleaned PDB/mmCIF files, atom37 NumPy arrays, and unified CSV metadata to eliminate preprocessing hurdles.

Transparent Redundancy Control: Multi-threshold Foldseek clustering for principled sample weighting, ensuring every experimental structure contributes.

Rigorous Benchmark: An out-of-sample test set comprising public PDB entries post--May 30, 2024 (disclosed) and confidential therapeutic complexes.

Using this benchmark, we evaluated six leading models (AlphaFold2.3?multimer, Boltz-2, Boltz-1x, Chai-1, DiffDock-PP, GeoDock) and found that success rates rarely exceed 25 %, built-in confidence metrics and ranking often misprioritize predictions, and all struggle with novel targets and binding poses.

I had the opportunity to present this work last week at the [ICML] Int'l Conference on Machine Learning 2025 Workshop on DataWorld: Unifying Data Curation Frameworks Across Domains (https://dataworldicml2025.github.io) in Vancouver.

Paper: https://www.researchgate.net/publication/393900649_SNAC-DB_The_Hitchhiker's_Guide_to_Building_Better_Predictive_Models_of_Antibody_NANOBODY_R_VHH-Antigen_Complexes

Dataset: https://zenodo.org/records/16226208

Code: https://github.com/Sanofi-Public/SNAC-DB

I hope SNAC-DB will accelerate the development and evaluation of more accurate models for antibody/NANOBODY® VHH complex prediction.

We welcome any comments on how to make this resource more user-friendly or of any deficiencies.

Best,
Abhinav Gupta
Senior AI/ML Scientist
Large Molecule Research, Sanofi
Email: abhinav.gupta[at]sanofi.com

Discussion forums: Resource: Introducing SNAC-DB: A New Resource for Antibody & NANOBODY® VHH-Antigen Modeling

Expanded view | Monitor forum | Save place

Start a new thread:

You have to be to post a reply.