Streamline Your Antibody Discovery: DeepRank-Ab Is Now Available on Vecura

What is DeepRank-Ab?

DeepRank-Ab is a specialized geometric deep-learning scoring function designed to evaluate the structural quality of antibody-antigen docking models. By leveraging atom-level graphs enriched with advanced ESM2 sequence embeddings and ANARCI-derived CDR annotations, it predicts DockQ scores to assess how closely a docked pose matches a true biological complex. It serves as a vital bridge in the computational antibody discovery pipeline, helping researchers quickly filter through large ensembles of candidate structures.

It helps users streamline the post-docking analysis phase by ranking thousands of candidate models in a fraction of the time required for traditional physics-based methods. It is especially useful for antibody engineers and structural biologists working with HADDOCK or other docking engines who need to prioritize high-confidence poses for downstream MD simulations or binding affinity studies.

What can users do with DeepRank-Ab on Vecura?

With DeepRank-Ab on Vecura, users can:

• Automatically compute DockQ scores for ensembles of docked antibody-antigen poses.
• Seamlessly process PDB files containing complex structural data without manual configuration.
• Identify the most structurally plausible antibody-antigen models using a ranked, descending output list.
• Assess the reliability of heavy/light-chain interfaces through automated structural quality flags.

What the output means

The output provides a ranked list of candidate models, each assigned a predicted_dockq score ranging from 0 to 1, where higher values indicate better predicted quality. Additionally, it offers a quality_flag (ok, low_HL_contacts, or not_applicable) to alert users to potential structural issues in the antibody framework.

This output should be used to support scientific decision making by narrowing down candidates for labor-intensive follow-up studies. It does not replace experimental validation and should be interpreted as a computational estimate of structural fidelity.

Why this matters

The development of high-affinity antibodies is a cornerstone of modern biotherapeutics, yet generating and validating accurate 3D models of antibody-antigen interactions remains a major computational hurdle. Traditional docking workflows often produce vast quantities of candidate models, most of which are structurally inaccurate, necessitating reliable and rapid scoring metrics to prioritize the candidates that are most likely to be biologically meaningful.

DeepRank-Ab addresses this bottleneck by integrating cutting-edge graph neural networks with deep protein language models (ESM2). By focusing on the unique geometry and sequence features of antibody CDR loops, it provides a tailored solution that outperforms general-purpose protein-protein docking scorers, ultimately enabling faster and more accurate structural insights in the early stages of drug discovery.

• Developed by: The Bonvin Lab (HADDOCK group), Utrecht University
• Source: DeepRank-Ab GitHub Repository
• Reference: Original Paper (bioRxiv 2025)