Ibex

Antibodies and T-cell receptors recognize their targets through hypervariable loops that carry almost no evolutionary signal, so the multiple sequence alignments powering general structure predictors add little for them. MSA-free specialized models grew up around that observation, but they inherited a subtler problem from the structural databases they train on: the same variable domain is often deposited in both an antigen-bound (holo) and an unbound (apo) form, and a model trained on undifferentiated data either averages the two or collapses onto whichever state is more common.

Ibex, from Prescient Design at Genentech with Large Molecule Research at Roche, makes that ambiguity explicit. It is a pan-immunoglobulin structure predictor covering the variable domains of antibodies, nanobodies, and T-cell receptors, and it takes a conformation token as an input feature alongside the sequence. During training, that token labels each structure apo or holo; at inference, setting it selects which state to predict, so a single sequence yields two distinct, state-specific structures. Released as a preprint in July 2025 and published in mAbs in December 2025, Ibex ships as a pip-installable inference package (prescient-ibex) with a pretrained checkpoint.

The practical argument for the model is cost. On a benchmark of previously unseen in-house antibody structures, Ibex matches the accuracy of co-folding systems such as Boltz-1 and Boltz-2 while predicting a variable domain in 0.7 seconds on a single NVIDIA A10G, against 63.9 seconds for Boltz-2 — before the MSA search those models normally require.

#Key Features

• Conformation-conditioned output: A residue-level apo/holo token, reinforced by a residual connection from the input embedding into every structure module block, lets one sequence produce separate bound and unbound predictions rather than an averaged compromise.
• Pan-immunoglobulin coverage: A single checkpoint handles paired antibody Fv domains, single-domain nanobodies, and TCR alpha/beta variable domains, learning the distinguishing motifs implicitly rather than through separate specialized models.
• MSA-free, sub-second inference: Sequence embeddings replace alignment search entirely, giving throughput suitable for screening large designed libraries.
• Recapitulated biophysics: Predicted apo and holo states reproduce the expected CDR-CDR hydrogen bond networks scored with the Rosetta energy function, and hold their accuracy on antibodies whose CDR H3 loops are far from anything public.

#Technical Details

Ibex builds on the AlphaFold 2 structure module, reusing code from OpenFold, with 16 consecutive invariant point attention blocks that iteratively refine coordinates. Inputs combine ESM-C 300M embeddings passed through a two-layer MLP, one-hot sequence encoding, and residue-level chain and conformation features; a pLDDT head predicts per-residue confidence over 50 bins. Training draws on three sources: 14,000 experimental antibody, nanobody, and TCR structures from SAbDab and STCRDab (760 of them matched apo/holo pairs, clustered into 4,200 groups), 22,000 immunoglobulin-like domains from the PDB, and a 60,000-structure distillation set built by folding diverse Observed Antibody Space sequences with ESMFold and Boltz-1. A three-stage curriculum shifts the sampling mix progressively toward experimental data and adds structural violation losses at the end. The released model is an ensemble of eight independently trained networks returning the prediction closest to the mean; training all eight took roughly 10 days on 64 NVIDIA H100 GPUs.

On the ImmuneBuilder test set, Ibex reaches 2.72 Å mean CDR H3 RMSD for antibodies and the lowest error of any compared method on the TCR CDR beta-3 (1.84 Å) and alpha-3 (1.93 Å) loops. On a private set of 286 antibodies with novel CDR H3 loops, it records 2.28 Å mean CDR H3 RMSD, ahead of Boltz-1 (2.30 Å), Boltz-2 (2.42 Å), Chai-1 (2.55 Å), ABodyBuilder3 (2.78 Å), and ESMFold (3.27 Å).

#Applications

The model targets therapeutic antibody and TCR engineering, where thousands of designed sequences need structures before any of them reach a bench. Sub-second inference makes structure a routine annotation in library-scale screening rather than a per-candidate expense, and the apo/holo switch supplies a physically appropriate starting conformation for docking or for property prediction on antibodies that display induced-fit recognition. The command-line tool accepts CSV or Parquet batches and optionally refines outputs with OpenMM.

#Impact

Ibex sharpens a distinction that structure predictors usually leave implicit — which conformational state a prediction represents — and shows the conditioning can be added cheaply to an established architecture. Its benchmarking also produced a notable negative result: sampling up to 1,000 seeds from Boltz-1 or Chai-1 barely improves CDR H3 accuracy, and both models rarely propose alternative loop backbones, so confidence-ranked oversampling does not substitute for explicit state conditioning. Real limits remain. The binary apo/holo split simplifies a continuous conformational landscape and cannot represent multiple distinct bound forms; no antigen context is used, so binding-induced changes are not target-specific; gamma and delta TCR chains are outside training. The code is Apache 2.0, but the weights carry a Genentech Apache 2.0 Non-Commercial license limiting use to non-commercial academic research, and the input embeddings inherit the EvolutionaryScale Cambrian Open License.