Use ABB4-STEROIDS Online
Commercially Available ABB4-STEROIDS No-Code Web Server
ABB4-STEROIDS: A Generative Model for Antibody Flexibility
ABB4-STEROIDS (AntibodyBuilder4 Structure predictor Tuned on Ensembles of complementary determining Regions Observed In molecular Dynamics Simulations) is a generative structure prediction model that samples conformational ensembles of antibodies rather than a single static structure. It was developed by researchers at the University of Oxford and the University of Copenhagen, and it is openly available.
Conformational flexibility is fundamental to protein function. In antibodies, it can affect key properties such as affinity and specificity. Predicting a single static structure is now highly accurate for many proteins, but predicting conformational ensembles remains a challenge. Molecular dynamics (MD) simulations are powerful but computationally expensive, and earlier deep learning methods have not matched their accuracy. ABB4-STEROIDS aims to close that gap.
Training data | Scale |
|---|---|
Structural frames from coarse-grained MD | 4.2 million |
Coarse-grained antibody simulations | ~136,000 |
New all-atom antibody MD simulations | 83 (released with the model) |
Model size | 19.6 million parameters |
Why Antibody Flexibility Matters
Most structural diversity in antibodies sits in the complementarity-determining regions (CDRs), while the frameworks stay comparatively static.
Flexibility can let one antibody bind several antigens in different conformational states, and it affects the thermodynamics of binding.
Several studies of individual antibodies link maturation to CDR rigidification, although repertoire-scale studies have not seen the same trend.
Because affinity and specificity are central to antibody drug development, mapping CDR conformational landscapes is valuable.
Why Existing Tools Fall Short
AlphaFold3 and Boltz-1 are trained on crystal structures only and have limited ability to sample diverse conformations.
MSA subsampling with AlphaFold2 increases diversity but is an inference-time workaround rather than a model built for ensembles.
General MD-trained models such as AlphaFlow, BioEmu and aSAM improve on this, but the authors note that no earlier method was trained and evaluated on antibody data alone.
All-atom MD is limited by cost, even with specialized supercomputers.
How ABB4-STEROIDS Works
ABB4-STEROIDS maps an antibody sequence to an ensemble of structures through flow matching over backbone frames, following the FrameFlow approach.
Input embedding. The sequence is turned into per-residue and residue-pair representations. A 14-category antibody region embedding marks each residue as one of the six CDRs or one of the eight framework regions.
Structure module. Eight layers built on AlphaFold2-style Invariant Point Attention update the representations and backbone frames.
Flow matching inference. Sampling starts from pure noise. The model repeatedly predicts a clean structure, builds a vector field and denoises the backbone frames.
All-atom reconstruction. Side-chain torsion angles from the final step are used to build the all-atom structure.
Four-Stage Training
Stage | Data | Purpose |
|---|---|---|
1. ABB4-base | 8,205 experimental antibody structures | Basic single-structure prediction |
2. Coarse-grained pre-training | All ~136,000 simulations, plus high-quality ones, 4.2M frames | Encourage diverse sampling |
3. High-quality refinement | High-quality coarse-grained subset only | Sharpen the sampled ensembles. Stages 1 to 3 give ABB4-STEROIDS-CG |
4. All-atom fine-tuning | 83 all-atom MD simulations | Reduce coarse-grained bias and clashes. Gives the final ABB4-STEROIDS |
The all-atom simulations ran three independent 300 ns replicas per antibody, giving 1.2 µs of simulation each. They used the AMBER ff19SB force field with the TIP3P-FB water model in OpenMM.
How ABB4-STEROIDS Performs
The authors benchmarked against Boltz-1, AlphaFold2 with MSA subsampling, aSAM, AlphaFlow and BioEmu, generating 100 structures per antibody for every method.
Reproducing MD Flexibility
Magnitude of flexibility: On both coarse-grained (100 antibodies) and all-atom (14 antibodies) test sets, the ABB4 models outperformed the alternatives at matching the flexibility seen in MD, measured by ensemble RMSD and RMSF.
Underestimation elsewhere: Other methods generally underestimated the flexibility of canonical CDRs, and Boltz-1 ensembles showed very low diversity.
Flexible CDRH3 loops: On the seven all-atom antibodies with flexible CDRH3 loops, accuracy for Boltz-1, MSA subsampling and aSAM dropped sharply, while ABB4-STEROIDS was the most accurate predictor.
Capturing Conformational States
Multiple states: For the CDRH3 of antibody 8F2T, ABB4-STEROIDS was the only model to capture two major equilibrium states. For the CDRH1 of 8VVB, it uniquely captured a second state that all other methods missed.
Population weighting: All-atom fine-tuning shifted more predictions into regions the MD simulations actually occupy. It also reduced the coarse-grained model's tendency to over-represent rare states.
Consistency with Experiment
Flexibility labels: Using antibodies with multiple crystal structures from SAbDab, labeled "rigid" or "flexible," the ABB4 models had the highest correlation between predicted diversity and experimental flexibility across all CDRs.
Conformation coverage: The ABB4 models covered the most experimental conformations for canonical loops at 2 Å and for CDRH3 at 3 Å, defining the best balance between accuracy and diversity.
Physical Validity
All-atom fine-tuning reduced atomic clashes compared with the coarse-grained-only model. ABB4-STEROIDS had a clash profile comparable to BioEmu. Static predictors such as Boltz-1 and AlphaFold2 produced the fewest, as expected.
Speed Versus Accuracy
The default is 100 model evaluations. Halving to 50 had a negligible impact on performance. Accuracy dipped slightly at 20 and 10 steps, with a larger loss at 5.
What is Tamarind Bio?
Tamarind Bio is a no-code bioinformatics platform built to give life scientists and researchers access to powerful computational tools. Many cutting-edge machine learning models are hard to deploy and use. Tamarind provides an intuitive, web-based environment that removes the complexity of high-performance computing, software dependencies and command-line interfaces.
The platform is designed for biologists, chemists and other researchers who may not have a background in programming or cloud infrastructure but want to run models on their own data. Key features include:
A user-friendly graphical interface for setting up and launching experiments
A robust API for integration into existing research pipelines
An automated system for managing and scaling computational resources
Tamarind treats information and data security as a top priority, as detailed in its Trust Center and Terms of Service.
Accelerating Discovery with ABB4-STEROIDS on Tamarind Bio
The authors describe how ensemble-level antibody modeling can support drug discovery:
Large-scale flexibility studies: Investigate antibody flexibility across many candidates without running MD for each one.
Rational design of flexibility-modulated properties: Properties such as affinity, specificity and developability depend on the unbound conformational ensemble.
Finding flexible versus rigid loops: Identify which CDRs, especially CDRH3, move most before choosing engineering strategies.
Capturing bound conformations: Bound conformations are commonly contained in the unbound ensemble, and the model covered many experimental antigen-bound conformations.
Selecting model checkpoints: The open-source release includes the full model, a coarse-grained-only version, and a base model for single-structure prediction.
How to Use ABB4-STEROIDS on Tamarind Bio
Access the platform. Log in to tamarind.bio and open the ABB4-STEROIDS tool.
Enter your antibody sequences. Provide the paired heavy chain (VH) and light chain (VL) amino acid sequences.
Set sampling depth. The default in the paper is 100 model evaluations. Reducing to 50 saves time with negligible performance impact. Going lower, to 20 or 10, brings slight accuracy losses.
Set ensemble size. Choose how many structures to sample. The paper generated 100 structures per antibody for its benchmarks, which is a sensible starting point.
Run the job. The model starts from noise and denoises backbone frames over the chosen number of steps.
Download the ensemble. The output is a set of all-atom structures sampled from the antibody's unbound conformational landscape.
Analyze flexibility. Compare structures across the ensemble, focusing on CDRs. The paper measured diversity with pairwise ensemble RMSD, per-residue RMSF, and tICA on backbone torsion angles.
Interpret with care. Treat the ensemble as an unbound (apo) prediction, and see the limitations below.
Things to Keep in Mind
Apo ensembles only. The model was trained on unbound antibody simulations. Antigen binding tends to rigidify CDRs, so the predicted flexibility matches the unbound state, not the bound one.
CDRH3 is hard. CDRH3 flexibility is on a similar scale to current prediction error, and recall for CDRH3 was lower for all methods.
Specific states can be missed. Relative population densities may not be accurate.
CDRH2 bias. The coarse-grained data is known to overestimate CDRH2 flexibility, and fine-tuning only partly corrects this.
Single structures. ABB4-base is competitive for single-structure prediction but is not state-of-the-art. Use other tools if you only need one best structure.
Source: Spoendlin, Cagiada, Ifashe, Vavourakis & Deane, Sampling antibody conformational ensembles with ABodyBuilder4-STEROIDS, bioRxiv (2026). DOI: 10.64898/2026.04.14.718378. Code: github.com/oxpig/ABB4. All-atom MD dataset: DOI 10.5281/zenodo.19471889