Loading…
Antibody CDR design model post-trained by on-policy distillation, cutting RAbD CDR-H3 backbone RMSD from 2.37 Å to 1.95 Å.
No providers recorded yet. Browse all providers
Antibodies recognize their targets through complementarity-determining regions (CDRs), six variable loops that form the antigen-binding interface. Diffusion models can now generate CDR sequences and backbones jointly against a fixed antigen, but they inherit a mismatch built into denoising training: the loss is computed on noisy states produced by perturbing a native structure, while sampling walks through states the model itself generated. For the long, flexible CDR-H3 loop, small backbone errors compound along that reverse trajectory, and the resulting geometry drifts away from anything the training distribution ever covered.
ABOPD (Antibody CDR Design via On-Policy Distillation) is a post-training framework that closes this gap. It first pretrains a hybrid base checkpoint, called H-DiffAb, that mixes single-CDR, multi-CDR, and all-six-CDR masking regimes in one model. It then builds a backbone-aware teacher from that base by granting it privileged access to native backbone geometry, and distills the teacher's per-step coordinate transitions into the student on states the student actually visits during its own reverse-denoising rollouts. The supervision is therefore dense, structural, and on-policy rather than a single sequence-level reward or a fixed offline dataset of teacher outputs.
The framework comes from a multi-institution Chinese academic collaboration led by the Shanghai Innovation Institute. It extends the DiffAb line of antigen-conditioned CDR diffusion models. Where antibody design systems such as the all-atom immune receptor model IgGM2, the retrieval-augmented binder generator RADiAnce, and the de novo nanobody pipeline Germinal each introduce a new generative architecture, ABOPD instead targets the post-training stage of an existing one.
H-DiffAb pretraining samples a target residue set at each update from a hybrid distribution weighted 0.25 single-CDR, 0.40 multi-CDR, and 0.35 all-six-CDR, noising and reconstructing only residues in that set, at roughly eight times the compute budget of standard DiffAb training on SAbDab-derived antibody-antigen complexes. The teacher's privileged descriptor comprises Cα-centered Cartesian coordinates and validity masks for the N, Cα, C, O, and observed Cβ atoms; explicit target residue labels and atoms beyond Cβ are withheld to limit information leakage. Training then combines an offline denoising anchor on forward-noised reference states with an on-policy coordinate-matching term computed on student rollouts, weighted uniformly across timesteps.
On RAbD CDR-H3 generation, ABOPD lowers backbone RMSD from 2.37 Å to 1.95 Å over the H-DiffAb base while raising amino-acid recovery from 36.35% to 37.31%, with marginal gains in lDDT and TM-score, and it outperforms both supervised fine-tuning and offline-distillation controls trained from the same base. In simultaneous six-CDR redesign, CDR-H3 RMSD drops from 2.94 Å to 2.48 Å, intra-CDR steric clashes fall from 0.93% to 0.42%, and CDR-antigen clashes fall from 1.58% to 0.67%. Timestep-resolved analysis shows the teacher's advantage emerging around the midpoint of denoising and peaking in the late refinement steps, consistent with the claim that trajectory drift is a geometry-refinement problem.
ABOPD is aimed at antibody engineers who need higher-fidelity backbone geometry from a generative CDR designer. Because the released checkpoint conditions on a fixed antigen structure, it can redesign CDR-H3 against a chosen epitope or co-design all six loops in one pass, serving as the proposal stage of affinity maturation, humanization, or epitope-focused de novo campaigns. Its broader appeal is methodological: the recipe is a general post-training procedure for diffusion-based structure generators, applicable wherever recursive sampling drift degrades designed geometry. Inference runs from the released code and checkpoint; there is no hosted API.
ABOPD imports on-policy distillation — a technique developed for post-training language models — into structural biomolecular diffusion, showing that dense per-step geometric supervision beats both supervised fine-tuning and offline teacher imitation on the same base model. The gains are concentrated in backbone accuracy: side-chain quality does not improve, with Jensen-Shannon divergence on side-chain dihedrals rising from 0.2451 to 0.2566 and the interface-improvement rate sitting marginally below baseline. The authors also note that dense trajectory supervision can concentrate generation on structurally reliable modes at the cost of exploration, a real tension for a task where novelty matters. As a preprint, the results are in-silico RAbD benchmarks awaiting peer review and experimental characterization of the designed antibodies.
Much of this page is generated or calculated automatically. Flag anything that looks off and we will re-run it.