PaperlayerView exact revision on arXiv

Source-linked AI summary

AbGaze: Attentive Geometric Representation Learning for End-to-End Antibody Design

Jiashuo Wang, Siqi Fan, Yizhen Luo, Zaiqing Nie

arXiv:2609.35296v1cs.AI

TL;DR

Antigen–antibody modeling needs representations that capture cross-molecular geometry, which existing distance-based or surface-intrinsic approaches characterize only partly. AbGaze encodes distance, direction, and surface-normal information in antibody-residue frames and adaptively aggregates interactions. It outperforms the second-best method across its reported design, structure, and affinity tasks.

  • Problem

    Existing interface representations provide limited explicit characterization of cross-molecular spatial relationships relative to antibody conformations.

    Source: §1 · p. 2

  • Method

    AbGaze encodes antigen-surface distance, direction, and surface normals in antibody-residue frames, using adaptive attention and local-frame geometric supervision in a shared masked-prediction framework.

    Source: §1 · p. 2

  • Results

    Relative to the second-best method, AbGaze improves amino-acid recovery by 7.1%, reduces per-CDR Cα RMSD by 14.9% on average over six CDRs, improves DockQ by 6.6%, and raises IMP by 32.5%.

    Source: §6 · p. 10

  • Takeaways & Limitations

    The results support explicitly representing paratope–epitope spatial relationships in antibody-residue local frames rather than leaving them implicit in scalar distances or surface-intrinsic features.

    Source: §6 · p. 10

Abstract

from arXiv · show

Computational antibody design requires representations that capture the geometric patterns underlying antigen--antibody interactions, yet existing approaches often rely on scalar distances or surface-intrinsic features, leaving cross-molecular geometry largely implicit. We present AbGaze, an end-to-end antibody design framework based on attentive geometric representation learning, which encodes distance, spatial direction, and surface-normal orientation of antigen surfaces relative to antibody-residue local frames, and adaptively aggregates these geometric interactions according to their interfacial context. The learned interaction representation is shared across multi-CDR co-design, complex structure prediction, and affinity optimization, with local-frame geometric supervision further constraining the representation. AbGaze outperforms prior methods across all three tasks: relative to the second-best method, it improves amino-acid recovery by 7.1% and reduces structural error by 14.9% on average over the six CDRs, improves interface docking quality (DockQ) by 6.6%, and raises the affinity improvement rate (IMP) by 32.5%.

1 INTRODUCTION

Antigen-conditioned antibody modeling depends on accurately representing local antigen–antibody geometry, but existing representations leave cross-molecular relationships partly implicit. AbGaze encodes these relationships in antibody-residue frames and uses adaptive attention and geometric supervision across antibody-design tasks.

1 INTRODUCTION

  • Antibodies bind antigenic epitopes through CDRs, whose sequence and conformational diversity supports diverse binding modes.Modeling antibodies conditioned on antigen structures therefore requires accurate characterization of interface relationships.Source: §1 · p. 1
  • End-to-end approaches jointly model antibody sequences and structures conditioned on antigen information, while interface studies show local geometric and chemical patterns provide interaction cues.Earlier antibody pipelines separated complex prediction, CDR design, and side-chain assembly, allowing intermediate errors to propagate.Source: §1 · p. 1
  • Existing representations provide limited explicit characterization of cross-molecular spatial relationships, although distance, approach direction, surface-normal angle, and interface context all matter.Scalar distances cannot distinguish contacts with different directions, while surface-intrinsic features are not expressed relative to antibody conformations.Source: §1 · p. 2
  • AbGaze represents antigen-surface distance, spatial direction, and surface normals relative to antibody-residue coordinate frames.This local interface representation is paired with adaptive atom–surface attention conditioned on interface geometry, with FAPE and dihedral supervision guiding spatial updates.Source: §1 · p. 2
  • AbGaze uses one binding prior and masked-prediction framework for multi-CDR co-design, complex structure prediction, and affinity optimization.Stochastic masked decoding also yields multiple plausible sequence–structure realizations.Source: §1 · p. 2

2 RELATED WORK

Antigen-conditioned antibody modeling has progressed from localized CDR design to end-to-end sequence–structure generation. Molecular-interface approaches characterize surface patterns, while AbGaze relates fine-grained surface geometry to residue-local frames.

2.1 ANTIGEN-CONDITIONED ANTIBODY MODELING

  • Antigen-conditioned antibody methods progressed from CDR generation to full-atom refinement, broader design regimes, and flow-matching models for antigen conformational transitions.Examples include HERN, MEAN, DiffAb, dyMEAN, AbDiffuser, IgGM, dyAb, and AbFlow.Source: §2.1 · pp. 2–3
  • AbGaze’s overview combines paratope-frame geometry and adaptive attention to iteratively refine antibody conformation.The attention mechanism learns contextual interactions and guides structural refinement.Source: Figure 1 · p. 3

2.2 MOLECULAR INTERFACE REPRESENTATION

  • Surface-interface methods use geometric fingerprints and intrinsic curvature, point-cloud matching, structural-environment filtering, and target-conditioned binder synthesis.These approaches include MaSIF, PInet, PeSTo, and ProBID-Net.Source: §2.2 · p. 3
  • AbGaze bridges fine-grained surface topography and frame-aligned local representations using a two-level surface-attention module.Prior approaches either coarse-grain atomic contact geometry at residue level or use atom-to-surface displacement vectors without residue-local decomposition.Source: §2.2 · p. 3

3 METHODOLOGY

AbGaze models antigen–antibody interfaces with residue-local geometric features and adaptive atom–surface attention. The same interface representation supports geometric supervision and generation across multiple modeling tasks.

3.1 OVERVIEW AND PROBLEM SETTING

  • AbGaze represents complexes as residue graphs with full-atom coordinates and jointly generates antibody sequences and structures from antigen context and masked target regions.Antigen surface meshes are constructed for epitope residues.Source: §3.1 · p. 3

3.2 ORIENTATION-AWARE LOCAL INTERFACE REPRESENTATION

  • Residue-local frames encode atom–surface distance, approach direction, and surface-normal incidence, distinguishing contacts that have equal distances but different orientations.The geometric quantities remain edge-local and are mixed across edges through invariant edge messages.Source: §3.2 · pp. 4–5

3.3 ADAPTIVE ATOM–SURFACE INTERACTION AND GATED COORDINATE UPDATE

  • Two-level attention aggregates atom–surface interactions and uses frame-aligned queries to attend over antigen surface vertices, producing residue-level interaction summaries and interface edge messages.Gated coordinate updates use invariant edge messages, while geometric descriptors, queries, and attended centroids are recomputed across encoder layers.Source: §3.3 · p. 5

3.4 GEOMETRIC SUPERVISION

  • FAPE constrains structural consistency across residue frames, while torsion and angle losses regularize local backbone geometry within the geometric training prior.The total objective also combines sequence, coordinate, bond, and docking losses, plus an auxiliary per-residue RMSD prediction loss.Source: §3.4 · p. 6

3.5 GENERATION ACROSS MODELING TASKS

  • Masking CDRs configures joint design, observing the full sequence enables structure prediction, and tuning initialization noise configures affinity optimization within the shared generator.Across tasks, the shared representation, attention, and equivariant refinement remain the same; configurations differ by masking, initialization, and sampling or optimization.Source: §3.5 · p. 6

4 EXPERIMENTS

AbGaze is evaluated on four RAbD settings using a shared protocol, with baselines re-run under that protocol. It leads on design, structure prediction, and affinity measures across the reported settings.

4.1 SETUP

  • Evaluation covers all-CDR design, CDR-H3 design, complex structure prediction, and affinity optimization on 60 RAbD complexes with overlapping training clusters removed.All baselines are re-run under the protocol; design uses five samples per target, structure prediction ten draws, and affinity optimization thirty candidates.Source: §4.1 · p. 7, §4 · p. 7

4.2 ALL-CDR DESIGN

  • AbGaze improves AAR by 7.1% and reduces RMSD by 14.9% on average across six CDRs, with DockQ 0.422 vs. 0.396 (+6.6%).Pooled AAR is 66.2%; L3 AAR reaches 0.69 (+19%), and H3 Cα RMSD falls to 1.65 Å (−10%).Source: §4.2 · p. 7, Table 1 · p. 8

4.3 CDR-H3 DESIGN

  • AbGaze leads CDR-H3 design with AAR 45.60% (+4.5%), CAAR 32.20% (+11.8%), DockQ 0.443 (+3.5%), and the lowest Cα RMSD (8.06).The reported gains concentrate at the interface, where encoding approach distance, direction, and surface-normal orientation helps align paratope geometry with epitope constraints.Source: §4.3 · pp. 7–8, Table 2 · p. 8

4.4 COMPLEX STRUCTURE PREDICTION

  • For complex structure prediction, AbGaze attains the best DockQ (0.435) and RMSD (8.03), while end-to-end methods have TM-score ≥0.97.The reported differences are at the interface, in the paratope’s spatial arrangement relative to the antigen surface.Source: Table 3 · p. 8, §4.4 · p. 8

4.5 AFFINITY OPTIMIZATION

  • IMP reaches 70.6%, with the best ∆∆G of −11.10; the improvement rate is 17.3 points above dyMEAN and 17.8 above AbFlow.The same frozen generator and ∆∆G regressor are used for all methods; AbGaze requests more mutations (∆L 8.07 vs. 4.25 for dyMEAN).Source: Table 4 · p. 9, §4.5 · p. 9

5 ABLATION

Ablations on RAbD all-CDR design show distinct effects from removing orientation-aware interactions and local-frame geometric supervision.

Ablation study

  • Without residue-local FAPE and dihedral terms, antibody Cα RMSD rises by 0.147 Å, lDDT falls by 0.043, and TM-score falls by 0.009, while DockQ is essentially unchanged.The local-frame terms are described as geometric regularizers that constrain adjacent-residue orientation and stabilize the overall fold.Source: §5 · p. 9, Table 5 · p. 9
  • Removing orientation-aware interactions reduces DockQ by 0.040 and AAR by 4.1%, while increasing antibody Cα RMSD by 0.061 Å.The authors attribute the losses to the removal of explicit spatial and surface-normal directionality, which carries information beyond scalar distances.Source: Table 5 · p. 9, §5 · p. 9

6 CONCLUSION

AbGaze combines an orientation-aware local interface representation with adaptive atom–surface attention. Across its reported tasks, it improves sequence recovery, structural error, docking quality, and affinity improvement rate relative to the second-best method.

Conclusion

  • A single model improves amino-acid recovery by 7.1%, reduces per-CDR Cα RMSD by 14.9% on average over six CDRs, improves DockQ by 6.6%, and raises IMP by 32.5%.The authors identify flexible-antigen extension, experimental validation, and developability optimization as next steps.Source: §6 · pp. 9–10

A EXTENDED METHOD DETAILS

The extended method details specify how AbGaze builds and attends to local atom–surface geometry, updates transient interface coordinates, and generates sequence–structure designs. They also establish that the interface update is SE(3)-equivariant and delimit the sequence-diffusion correspondence.

A.1 ATOM-SLOT SCHEMA AND INTERFACE EDGES

  • AbGaze uses a 14-slot atom vocabulary, padding absent atoms at Cα and masking them; context and interface edges use k-NN with k=9 over residues.An interface edge pairs an epitope residue with an antibody residue and carries the epitope surface patch.Source: §A.1 · p. 13, Table 6 · p. 14

A.2 SURFACE PATCH CONSTRUCTION

  • MSMS antigen-surface vertices within 10 Å of epitope atoms are assigned to nearest epitope residues, then sampled or padded to 50 vertices per patch.The precomputed mesh remains fixed across decoding steps, and padded vertices are masked from attention.Source: §A.2 · pp. 13–14

A.3 DESCRIPTOR COMPOSITION AND ATTENTION AXES

  • The 40-channel descriptor combines frame-aligned geometry, surface-normal orientation, and atom and vertex features; separate attention normalizations aggregate over vertices and atom slots.Masked entries are excluded from normalization, and the resulting readouts and edge messages are SE(3)-invariant.Source: §A.3 · p. 14

A.4 TRANSIENT COORDINATES AND LAYER-WISE RECURRENCE

  • Within each refinement round, transient epitope coordinates evolve across encoder layers, are reinitialized in the next round, and never replace the fixed antigen context.Inter-chain updates read prior epitope states through pair-distance features; complex-structure prediction uses an empty mask, while affinity optimization adds Gaussian noise to template initialization.Source: §A.4 · pp. 14–15, Table 7 · p. 15

A.5 SE(3)-EQUIVARIANCE OF THE INTERFACE UPDATE

  • The interface update is SE(3)-equivariant because its attended centroid and displacement basis co-transform, and equivariance composes across encoder layers.The proof combines the surface-attention update with an EGNN inter-chain operator built from invariant distances and covariant point differences.Source: §A.5 · p. 15

A.6 GENERATION PROCEDURE

  • Training and decoding share masked initialization and refinement rounds, with recurrent node-state memory carried across reveal rounds and randomized token commitment.Softmax sampling temperature controls sequence–structure design concentration, with greedy decoding as the temperature approaches zero; the diffusion correspondence is limited to sequence unmasking, not coordinate diffusion or a joint-process ELBO identity.Source: §A.6 · pp. 15–16

B TRAINING AND TASK UNIFICATION

Training uses one jointly trained model and objective across task configurations, with task behavior determined by masking, initialization, and sampling or optimization. Affinity optimization alone uses an externally fitted, frozen prediction head.

B TRAINING AND TASK UNIFICATION

  • The surface construction uses MSMS per-residue patches subsampled to 50 vertices.Training uses Adam with an exponentially decayed learning rate.Source: B TRAINING AND TASK UNIFICATION · p. 16
  • Training anneals the fraction of masked residues revealed as conditioning, so the model learns decoding from intermediate generative states.All tasks are trained jointly in one run, and no schedules or loss weights are tuned per task.Source: B TRAINING AND TASK UNIFICATION · p. 16
  • All tasks share one architecture, objective, and checkpoint, with task behavior determined by masking configuration rather than task-specific adaptation.Complex-structure prediction observes the full antibody sequence, while affinity optimization uses the unchanged frozen generator and a separate frozen ΔΔG regressor.Source: B TRAINING AND TASK UNIFICATION · pp. 16–17
  • The shared objective includes masked-token cross-entropy and coordinate, bond, shadow-paratope, inter-edge, FAPE, torsion, and auxiliary per-residue RMSD losses.Table 8 lists the loss terms and weights; the structure, docking, and auxiliary terms use smooth-ℓ1 penalties.Source: Table 8 · p. 17, B TRAINING AND TASK UNIFICATION · p. 17

C EVALUATION DETAILS

The evaluation uses a leakage-controlled RAbD benchmark and a unified protocol for baseline comparison. It reports docking, structural, and binding-residue recovery metrics under specified alignment and interface definitions.

C EVALUATION DETAILS

  • Training and validation use 5,370 cleaned SAbDab complexes, while the benchmark is the 60-complex RAbD set with CDR-H3 clusters held out from training and validation.The split removes training and validation complexes sharing a 40%-identity CDR-H3 cluster with any hold-out benchmark complex.Source: C EVALUATION DETAILS · p. 17
  • Baselines use the same benchmark, evaluator implementations, inputs, initialization, and per-task selection rules under the unified protocol.Some structure-prediction and affinity-optimization baselines are locally retrained or reproduced on RAbD.Source: C EVALUATION DETAILS · p. 17
  • The affinity proxy is fitted on designed variants of training complexes only, frozen during optimization, and used to score every compared method.No test complex is used to fit the proxy.Source: C EVALUATION DETAILS · p. 17
  • DockQ is measured on the CDR-H3–antigen interface, RMSD uses Kabsch-aligned Cα coordinates, and CAAR restricts amino-acid recovery to binding residues within 6.6 Å of the epitope.Full-antibody RMSD includes heavy and light chains, with per-CDR RMSDs aligned per loop.Source: C EVALUATION DETAILS · p. 17

D COMPARISON WITH IGGM

On IgGM’s own SAbDab split, AbGaze exceeds IgGM on five of six CDRs, while IgGM leads on L1. The comparison is qualified by differences in test sets, training-set size, and modeled atoms.

D COMPARISON WITH IGGM

  • The comparison is not on a shared test split: IgGM uses post-2023 SAbDab entries, whereas AbGaze’s unified protocol uses RAbD.IgGM also trains on more complexes and models backbone atoms only, precluding full-atom metric comparisons.Source: D COMPARISON WITH IGGM · p. 18
  • AbGaze outperforms IgGM on five of six CDRs, with margins of 5.0 points on H2 and 4.2 points on L3; IgGM leads on L1 by 2.2 points.The comparison uses IgGM’s published results on its own SAbDab split.Source: D COMPARISON WITH IGGM · p. 18, Table 9 · p. 18

E DIVERSITY ANALYSIS

With five samples per target at τ = 0.5, all-CDR design diversity differs substantially by loop: H3 is nearly saturated, while framework-proximal L2 is most conservative.

E DIVERSITY ANALYSIS

  • Table 10 reports, per CDR, distinct-sequence counts and mean pairwise normalized Hamming distance across five all-CDR samples per target at τ = 0.5.The decoding distribution's concentration is controlled by τ.Source: Table 10 · p. 18, E DIVERSITY ANALYSIS · p. 18
  • H3 has 4.92/5 distinct sequences and 0.458 pairwise Hamming distance, with only 1.7% of targets identical; L2 has 1.42 distinct sequences and 70% identical.These figures describe five samples per target in the all-CDR design setting at τ = 0.5.Source: E DIVERSITY ANALYSIS · p. 18
  • The contrast tracks loop context: H3 dominates binding specificity and tolerates the most sequence variation, whereas framework-proximal L2 is most conservative.Source: E DIVERSITY ANALYSIS · p. 18

F GAINS WITHOUT DIVERSE SAMPLING

A single deterministic design retains strong benchmark performance without diverse sampling, while the main protocol keeps diverse decoding for parity and practical candidate-pool needs. The hyperparameters are not tuned per task.

F GAINS WITHOUT DIVERSE SAMPLING

  • The single-draw design predicts all masked target-region positions in one forward pass, without iterative commit-and-resample or temperature τ.It is evaluated on all-CDR RAbD under otherwise identical conditions, against full generative AbFlow and dyMEAN results.Source: F GAINS WITHOUT DIVERSE SAMPLING · pp. 18–19
  • 64.0% AAR, down from 66.2%, while deterministic design beats full generative AbFlow and dyMEAN on every metric; mean RMSD improves from 1.052 to 1.021 Å.DockQ is 0.407, above both baselines (0.396/0.379).Source: Table 11 · p. 19, F GAINS WITHOUT DIVERSE SAMPLING · p. 19
  • The main protocol retains diverse absorbing-state decoding for like-for-like sampling and candidate pools suited to screening, filtering, and best-of-N retrieval.The paper says diversity concentrates on loops with the most design freedom.Source: F GAINS WITHOUT DIVERSE SAMPLING · p. 19

G HYPERPARAMETERS

Related papers

Loading 2609.35296v1…