Use ITsFlexible Online
Commercially Available ITsFlexible No-Code Web Server
ITsFlexible: Predict Antibody and TCR CDR3 Loop Flexibility with Graph Neural Networks
ITsFlexible is a graph neural network that classifies antibody and T cell receptor (TCR) CDR3 loops as rigid or flexible from a single structure. It was developed by researchers including Fabian C. Spoendlin and Charlotte M. Deane (Department of Statistics, University of Oxford), with collaborators, and published in Nature Machine Intelligence in 2025. The code is available on GitHub (oxpig/ITsFlexible).
Tools such as AlphaFold predict a single static structure well, but predicting how much a protein region moves remains an open problem, largely because training data on conformational diversity is scarce. This matters for antibodies and TCRs, where the CDR3 loops strongly influence binding affinity and specificity.
Crystal structures of the same loop are often deposited in several conformations, yet confidence scores and diversity measures from structure predictors turned out to be poor indicators of flexibility. ITsFlexible was built to fill that gap with a model trained directly on observed loop conformations.
Property | Detail |
|---|---|
Task | Binary classification of CDR3 loops as rigid or flexible |
Model | Graph neural network with three E(n)-equivariant graph convolutional layers |
Input | The loop plus all residues within 10 Å, from the antibody or TCR structure |
Training data | ALL-conformations dataset: more than 1.2 million loop structures and more than 100,000 unique sequences from the PDB |
Loop types covered | Antibody CDRH3 and CDRL3, TCR CDRB3 and CDRA3 |
Publication | Nature Machine Intelligence, 2025 (DOI 10.1038/s42256-025-01131-6) |
Why CDR Loop Flexibility Prediction Matters for Antibody and TCR Engineering
Loop flexibility affects how an antibody or TCR engages its target. A rigid CDR3 can be modeled with more confidence, while a flexible one may adopt several shapes that change docking, affinity and specificity. Knowing which case you are in helps prioritize candidates and interpret predicted structures.
How ITsFlexible Works
The ALL-conformations dataset
The authors mined PDB structures deposited before 22 November 2023 for loops between two adjacent antiparallel beta-strands, adding antibody structures from SAbDab and TCR structures from STCRDab. Structures were filtered to X-ray entries at better than 3.5 Å resolution with no unresolved loop residues. Identical loop sequences were clustered by C-alpha RMSD, with a 1.25 Å threshold defining a distinct conformation.
Labels and model
Flexible: a loop sequence seen in more than one conformation.
Rigid: a loop seen in a single conformation across at least five separate PDB files.
Unknown: everything else, excluded from labeled training.
Architecture: nodes carry amino acid type and loop or context flags; edges connect residues within 10 Å; the network pools to a sigmoid output giving a flexibility score.
Roughly 20,000 unique sequences were labeled, with more than 16,000 rigid and about 4,000 flexible. The final model was the best of ten trained, chosen by validation PR AUC.
Benchmark Results: Accuracy on Crystal Structures, Molecular Dynamics and Cryo-EM
Test | Outcome |
|---|---|
PDB loop test set (2,845 loop motifs) | PR AUC 0.62, ROC AUC 0.84, ahead of baselines and zero-shot methods |
Antibody and TCR CDR3 test sets | Best in nearly all sets; MSA subsampling was slightly better on the small CDRA3 set |
Molecular dynamics (19 antibodies) | Near-perfect separation for CDRH3 loops, slightly weaker for CDRL3 |
Cryo-EM on three antibodies | Two of three predictions were consistent with the experiments |
The compared alternatives included logistic regression on loop length or solvent exposure, AlphaFold2 pLDDT, diversity of AlphaFold2 models from MSA subsampling, and ABodyBuilder2 error estimates. In the cryo-EM test, a CDRH3 scored 0.02 (rigid) and showed one homogeneous class, one scored 0.76 (flexible) and showed heterogeneity at the binding interface, and one scored 0.60 but showed no heterogeneity.
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 ITsFlexible on Tamarind Bio
Triage antibody candidates: flag CDRH3 and CDRL3 loops likely to be flexible before committing to expression or structure work.
Add context to predicted structures: pair a single predicted model with a flexibility score instead of relying on pLDDT alone.
Study TCRs: score CDRB3 and CDRA3 loops, keeping in mind the small CDRA3 benchmark set.
Guide experiments: use scores to decide where heterogeneity, such as in cryo-EM, may be expected.
How to Use ITsFlexible on Tamarind Bio
Open the tool: log in to tamarind.bio and select ITsFlexible.
Provide a structure: supply an antibody or TCR structure, as the model reads the loop and surrounding residues within 10 Å. Experimental or predicted structures can be used, though the authors note accuracy is lower with AlphaFold2 models for antibodies.
Select the loop: choose the CDR3 of interest (CDRH3, CDRL3, CDRB3 or CDRA3), if exposed.
Run the job: submit and wait for the prediction.
Download results: retrieve the flexibility score for each loop.
Interpret the score: higher scores indicate a loop more likely to be flexible; treat values near the middle with caution, as in the authors' cryo-EM example.
Use downstream: combine scores with docking, design or experimental planning.
Things to Keep in Mind
Crystal structures do not directly measure flexibility, so unseen conformations may exist and rigid labels are less certain than flexible ones.
The cryo-EM experiments used antigen-bound antibodies, which can rigidify CDRs; predictions may better match unbound antibodies.
Flexibility conditional on antigen binding is not addressed.
The CDRA3 test set is small, so its performance estimates are less reliable.
Accuracy is lower when antibody inputs are AlphaFold2 models than with crystal or ImmuneBuilder structures.
Source: Spoendlin et al., "Predicting the conformational flexibility of antibody and T cell receptor complementarity-determining regions", Nature Machine Intelligence 7, 1755-1767 (2025), DOI 10.1038/s42256-025-01131-6. Code: github.com/oxpig/ITsFlexible.