Use ABB4-STEROIDS Online

ABB4-STEROIDS: Sampling Antibody Conformational Ensembles with State-of-the-Art Accuracy

ABB4-STEROIDS (AntibodyBuilder4 Structure predictor Tuned on Ensembles of complementary determining Regions Observed In molecular Dynamics Simulations).

ABB4-STEROIDS is an open-source, generative structure prediction model developed by researchers at the University of Oxford and the University of Copenhagen. It is specifically designed to sample the complex conformational ensembles of antibodies with unprecedented speed and accuracy, bypassing the massive computational constraints of traditional molecular dynamics (MD) simulations.

What is ABB4-STEROIDS?

Traditional protein structure predictors focus on providing single, static models. However, antibody function—specifically binding affinity, specificity, and developability—is heavily dictated by structural flexibility, which is predominantly concentrated in the complementarity-determining regions (CDRs).

ABB4-STEROIDS solves this challenge by operating as an SE(3) flow matching model trained specifically on antibody dynamics. Built upon the Invariant Point Attention (IPA) architecture, it maps an input sequence directly into a diverse, physically valid structural ensemble.

A Novel Four-Stage Training Strategy

To eliminate force-field biases and maximize physical realism, the model was trained using a groundbreaking multi-stage protocol:

1. Stage 1 (ABB4-base): Trained on 8,205 experimental crystal structures to master basic single structure prediction capabilities.

2. Stage 2 & 3 (ABB4-STEROIDS-CG): Pre-trained on a massive corpus of 4.2 million representative structural frames derived from ~136,000 coarse-grained antibody MD simulations from FlAbDab. This stage encourages wide sample diversity.

3. Stage 4 (ABB4-STEROIDS Final): Fine-tuned on a brand-new set of 83 all-atom MD simulations (totaling 1.2 µs of simulation time per antibody). This critical refinement phase mitigates coarse-grained biases, improves equilibrium density weighting, and significantly minimizes atomic clashes.

State-of-the-Art Benchmarks & Performance

ABB4-STEROIDS achieves state-of-the-art accuracy across both computational and experimental benchmarks, outperforming top alternative methods such as Boltz-1, AlphaFlow, BioEmu, aSAM, and AlphaFold2 MSA subsampling.

• Accurate Magnitude of Flexibility: In benchmarks measuring ensemble RMSD and RMSF errors, ABB4-STEROIDS consistently achieves top performance in reproducing true computational landscapes.

• Capturing Dynamic Antibodies: While static baseline models struggle to simulate highly flexible loops, ABB4-STEROIDS excels at capturing the conformational states of dynamic CDRH3 loops.

• Experimental Consistency: When tested against experimentally resolved antibody ensembles from SAbDab, ABB4-STEROIDS showed the highest point-biserial correlation to binary flexibility labels across all CDR loops.

• Optimal Pareto Front: The model defines the absolute Pareto front for conformation coverage, providing the best scientific balance between structural accuracy and spatial diversity.

What is Tamarind Bio?

Tamarind Bio is an advanced computational biology platform dedicated to democratizing access to cutting-edge deep learning tools for molecular design and structure prediction. By providing a clean, accessible interface and scalable cloud-based infrastructure, Tamarind Bio empowers researchers, structural biologists, and therapeutics developers to run complex machine learning models instantly from their browsers—without needing specialized supercomputers, local installations, or advanced command-line expertise.

How to Use ABB4-STEROIDS on Tamarind Bio

Running ensemble generation with ABB4-STEROIDS on Tamarind Bio takes only a few seconds:

1. Input Your Sequence: Provide your antibody's Heavy Chain (HC) and Light Chain (LC) amino acid sequences into the text fields.

2. Configure Inference Parameters: * Sampling Depth: Set your desired number of model rollout evaluations (diffusion/flow matching steps). The paper indicates that while the default is 100 steps, it can be seamlessly halved to 50 steps for optimal speed with negligible performance impact.

   • Ensemble Size: Define how many unique structures you want to sample (e.g., 100 structures for a comprehensive ensemble analysis).

3. Run Inference: Click "Generate Ensembles." Tamarind Bio's infrastructure handles the underlying vector field integrations and structural module rollouts automatically.

4. Download & Analyze: Download your structural ensemble as an all-atom coordinates file format. You will receive a diverse array of snapshots charting out the entire thermodynamic apo-ensemble landscape of your antibody's CDR loops.

Source