Antibody ensemble generation: a diffusion model sampling Fv conformations from sequence in seconds, matching molecular-dynamics distributions.
No providers recorded yet. Browse all providers
A crystal structure captures the conformation that packs best into a lattice, which need not be the one that binds antigen. Blackler and colleagues solved 20 structures of one antibody with no antigen present: the bound-state geometry was already there among them, sampled by the antibody's own dynamics. Molecular dynamics exposes flexibility in 84% of CDR-H3 loops where crystal structures alone suggest 64%. A predictor returning one conformation per sequence answers a different question than antibody engineers are asking.
Cadense — Conformational Antibody Diffusion Ensembles — answers the ensemble question directly. Built at Prescient Design, the machine learning group inside Genentech, it is an equivariant diffusion model that generates diverse Fv structures from sequence, trained on all-atom molecular dynamics so that the spread of its samples reproduces a simulation's rather than a generative model's noise. What takes days of simulation takes seconds of sampling, and nothing is fit per antibody: one trained checkpoint serves sequences it has never seen.
Two earlier generative ensemble models, AlphaFlow and BioEmu-1, do not transfer cleanly here. Both operate on single chains, so a paired Fv has to be assembled by diffusing one chain and grafting loops onto a static partner, which clashes at the interface and discards correlated motion. Both also condition on multiple sequence alignments, which carry little signal for antibodies, whose diversity comes from V(D)J recombination rather than evolutionary divergence. Cadense models the paired Fv as one object and replaces the alignment with protein language model embeddings.
The score model is an Equiformer-style SE(3)-equivariant graph attention network restricted to steerable features of maximum order one, with hierarchical message passing over backbone, side-chain and global nodes. The input conditioner uses Pairformer-style blocks whose semi-local triangle attention cuts the usual cubic cost to O(kn²). Training data combine 4,895 experimental structures, MD trajectories for 13,192 unique Fv sequences from SAbDab, TheraSAbDab and pOAS, and 87,669 predicted structures from paired OAS; simulation starting structures came from ESMFold, refined with Rosetta, and ran 200 ns under Amber FF14SB. Evaluation sets were held out by MMseqs2 clustering at 80% identity.
On 22 held-out pOAS antibodies, per-residue Cα RMSF from Cadense-dist ensembles correlates with MD at a median Pearson of 0.84, against 0.57 for Boltz-2, which flattens the flexibility profile. On a held-out set of 33 apo/holo pairs, Cadense was the only method to place samples near both experimental references on the three pairs shown in detail and across seven medium-hard examples in aggregate; for one newly crystallized in-house Fab, unseen by every method, its best sample reached 0.89 Å CDR-H3 RMSD. On 72 antibodies scored for developability, ensemble-averaged surface properties gave four green-to-red misclassifications against 14 for a single MD frame. In HADDOCK3 docking of 44 internal antibody–antigen complexes, Cadense ensembles outscored single-structure inputs from ABodyBuilder3 and Ibex, holding their DockQ scores on CDR-H3 loops seven or more edits from any public sequence, where those baselines degrade.
Cadense targets the stages of antibody discovery where molecular dynamics is the right tool and the wrong budget. Ensemble-based docking supplies a set of paratope conformations rather than one guess, which matters most for novel CDR-H3 loops. Developability triage — hydrophobic patch scores, surface charge, and the risk flags derived from them — becomes an ensemble average computed in under a minute per antibody rather than a property read off one snapshot. Flexibility itself becomes screenable across a designed library, since CDR rigidity tracks binding affinity.
Ensemble generation is where antibody structure prediction is heading, and Cadense does it for paired Fv regions without alignments or chain-grafting workarounds. That its ensembles track MD flexibility profiles and shift developability calls the way MD would argues the samples carry physical information, not decorative variance. The work is a preprint awaiting peer review, and the limits are stated plainly: the model cannot extrapolate beyond the 200 ns timescales it saw, yields no kinetic information, covers Fv regions rather than full-length antibodies, and models the antibody with no antigen dynamics. Neither code nor weights accompany the preprint, so the results stand on the authors' internal benchmarks; the two apo Fab structures they solved carry Protein Data Bank accessions 38LM and 38LO, both still held for publication and not yet downloadable.
Much of this page is generated or calculated automatically. Flag anything that looks off and we will re-run it.