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

Jiashuo Wang; Siqi Fan; Yizhen Luo; Zaiqing Nie

cs.AISep 28, 2026

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

Authors: Jiashuo Wang, Siqi Fan, Yizhen Luo, Zaiqing Nie

Organizations: Institute for AI Industry Research (AIR), Tsinghua University · Department of Computer Science and Technology, Tsinghua University · PharMolix Inc.

Abstract

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%.

Figures & tables

Figure 1: Overview of AbGaze. We propose AbGaze, a unified framework for antigen–antibody interaction modeling based on orientation-aware geometry and adaptive attention. The framework iteratively updates the antibody conformation by explicitly encoding the relative spatial relationships between the paratope local frames and the antigen surface, while an adaptive attention mechanism learns contextual interactions and guides structural refinement.

Figure 2: Overview of local interface representation and adaptive attention. (a) Global interaction view between the antibody (CDR) residue and the antigen surface mesh. (b) Orientation-aware geometric encoding expressed in the antibody-residue local frame, capturing distance ( $r$ ), spatial direction ( $\Theta,\Phi$ ), and surface normal ( $\Theta^{n},\vec{n}$ ). (c) Adaptive atom–surface attention weighting used to dynamically locate the interaction center on the antigen surface.

Metric	Item	AbGaze	AbFlow	dyMEAN
AAR $\uparrow$	L1	0.77	0.69	0.76
	L2	0.85	0.82	0.83
	L3	0.69	0.58	0.52
	H1	0.79	0.74	0.76
	H2	0.71	0.65	0.69
	H3	0.43	0.38	0.38

Table 1: Performance comparison on all-CDR antibody design.

\uparrow

indicates higher is better, while

\downarrow

indicates lower is better. Bold denotes the best performance.

Method	AAR $\uparrow$	TMscore $\uparrow$	lDDT $\uparrow$	CAAR $\uparrow$	RMSD $\downarrow$	DockQ $\uparrow$
RosettaAb	32.31%	0.9717	0.8272	14.58%	17.70	0.137
DiffAb	35.31%	0.9695	0.8281	22.17%	23.24	0.158
MEAN	37.38%	0.9688	0.8252	24.11%	17.30	0.162
HERN	32.65%	–	–	19.27%	9.15	0.294
dyMEAN	43.65%	0.9726	0.8454	28.11%	8.11	0.409
AbFlow	42.10%	0.9735	0.8518	28.80%	8.45	0.428

Table 2: Performance comparison on CDR-H3 antibody design on the RAbD benchmark.

Model	TMscore $\uparrow$	lDDT $\uparrow$	RMSD $\downarrow$	DockQ $\uparrow$
HDock	0.9723	0.8503	18.46	0.170
HERN	0.9722	0.8441	10.19	0.424
dyMEAN	0.9730	0.8568	9.04	0.409
AbFlow	0.9720	0.8526	8.66	0.419
AbGaze	0.9735	0.8572	8.03	0.435

Table 3: Performance comparison on antigen–antibody complex structure prediction.

Method	Best $\Delta\Delta G$ $\downarrow$	IMP (%) $\uparrow$	$\Delta L$ $\downarrow$
DiffAb	-3.29	38.8	5.62
dyMEAN	-4.47	53.3	4.25
AbFlow	-9.31	52.8	6.98
AbGaze	-11.10	70.6	8.07

Table 4: Comparison of affinity optimization performance.

Variant	AAR $\uparrow$	RMSD $\downarrow$	DockQ $\uparrow$	lDDT $\uparrow$	TMscore $\uparrow$
AbGaze (Full)	66.2%	1.052	0.422	0.831	0.973
(A) w/o Orient.-Aware Attn.	$-4.1\%$	$+0.061$	$-0.040$	$-0.007$	$-0.003$
(B) w/o Local-Frame Supv.	$-1.2\%$	$+0.147$	$+0.012$	$-0.043$	$-0.009$

Table 5: Ablation study on all-CDR antibody design.

Appendix figures & tables7 assets

Supplementary material from the paper’s appendix.

Appendix

Res.	Slots 5–14	Res.	Slots 5–14
Gly	—	Ser	CB, OG
Ala	CB	Thr	CB, OG1, CG2
Val	CB, CG1, CG2	Cys	CB, SG
Leu	CB, CG, CD1, CD2	Pro	CB, CG, CD
Ile	CB, CG1, CG2, CD1	Phe	CB, CG, CD1, CD2, CE1, CE2, CZ
Asp	CB, CG, OD1, OD2	Tyr	CB, CG, CD1, CD2, CE1, CE2, CZ, OH

Appendix

Table 6: Side-chain heavy atoms occupying slots 5–14, in slot order; slots 1–4 hold the backbone atoms.

State	Initialization	Evolution	Training loss
antigen atoms	observed context	fixed	—
mesh $\{\mathbf{v}_{j}\}$ , normals	precomputed (MSMS)	fixed	—
sequence (framework)	observed	fixed	—
sequence (CDRs)	$[\mathrm{MASK}]$	committed per round	$\mathcal{L}_{\mathrm{seq}}$
coordinates (all)	framework template	refined per round	structure terms
transient states $\tilde{\mathbf{x}}_{i,p}$	from observed antigen	Eq. 6 only	none

Appendix

Table 7: Provenance of the coordinate and sequence states consumed by the interaction encoder.

Term	Definition (one line)	Weight	Source
$\mathcal{L}_{\mathrm{seq}}$	per-round CE, masked residues	1	standard
$\mathcal{L}_{x}$	Kabsch-aligned coordinates	1	Kong et al. (2023)
$\mathcal{L}_{\mathrm{bond}}$	backbone/side-chain bond lengths	1	Kong et al. (2023)
$\mathcal{L}_{\mathrm{SP}}$	shadow-paratope coordinates	1	Kong et al. (2023)
$\mathcal{L}_{\mathrm{ed}}$	predicted inter-edge distances	1	Kong et al. (2023)
$\mathcal{L}_{\mathrm{FAPE}}$	frame-aligned point error, $d_{\max}$ clamp	0.5	Jumper et al. (2021)

Appendix

Table 8: Loss terms of Eq. 9 ; structure, docking, and auxiliary terms are smooth-

\ell_{1}

penalties.

CDR	AbGaze	IgGM	AbGaze wins
H1	0.752	0.740	✓
H2	0.694	0.644	✓
H3	0.397	0.360	✓
L1	0.728	0.750	$\times$
L2	0.753	0.743	✓
L3	0.677	0.635	✓

Appendix

Table 9: CDR-level AAR comparison with IgGM on its own SAbDab split (post-2023); IgGM numbers are from its published Table 2.

CDR	Distinct ( $\uparrow$ )	Diversity ( $\uparrow$ )	% all-same
H1	2.60	0.092	21.7%
H2	2.93	0.127	16.7%
H3	4.92	0.458	1.7%
L1	2.00	0.070	50.0%
L2	1.42	0.062	70.0%
L3	3.93	0.181	3.3%

Appendix

Table 10: Design diversity on all-CDR design (five samples per target,

\tau=0.5

Method	AAR $\uparrow$	CAAR $\uparrow$	RMSD $\downarrow$	DockQ $\uparrow$	lDDT $\uparrow$	TM-Score $\uparrow$
dyMEAN	60.1%	50.3%	1.357	0.396	0.803	0.965
AbFlow	59.7%	49.8%	1.104	0.379	0.815	0.971
AbGaze (Reg.)	64.0%	54.9%	1.021	0.407	0.834	0.974
AbGaze (Full)	66.2%	57.1%	1.052	0.422	0.831	0.973

Appendix

Table 11: All-CDR design on RAbD without diverse sampling; notation as in Table 1 .

Hyperparameter	Value	Description
embed dim	64	residue/atom embedding
hidden size	128	message-passing hidden size
encoder layers	3	equivariant encoder layers
refinement rounds $R$	3	rounds per decoding step
$k$ (neighbors)	9	KNN neighbors per residue
atom channels $C$	14	4 backbone + $\leq$ 10 side-chain atoms

Appendix

Table 12: Hyperparameters.

Explore similar work

AgForce Enables Antigen-conditioned Generative Antibody Design
May 20, 2026Mansoor Ahmed, Murray PattersonAntibody
1Georgia State University, Atlanta, GA, USA · 2Georgia Institute of Technology, Atlanta, GA, USA
ConTact: Contact-First Antibody CDR Design via Explicit Interface Reasoning
May 20, 2026Mansoor Ahmed, Spencer VonBank, Nadeem Taj +3Antibody Contact
Georgia State University, Atlanta, USA · Georgia Institute of Technology, Atlanta, USA · DePauw University, Indiana, USA +2
Antigen-specific Antibody Multi-modal Foundation Model for Functional Antibody Design
Jul 22, 2026Xiaoliang Shi, Zichen Wang, Runze Ma +2Antibody Epitope Prediction
Shanghai Jiao Tong University

May 20, 2026, cs.LGJ/K move · Enter open · S save

AgForce Enables Antigen-conditioned Generative Antibody Design

Mansoor Ahmed, Murray Patterson

1Georgia State University, Atlanta, GA, USA · 2Georgia Institute of Technology, Atlanta, GA, USA

Antibody

Open paper AskPDF arXiv

Antibody design methods condition on antigen structure to generate complementarity-determining regions (CDR), yet a systematic evaluation of baseline methods reveals that they largely ignore the antigen input. We identify three failure modes that explain this behavior. Antigen blindness arises because models derive predictions from antibody framework context rather than antigen information, producing nearly identical CDRs regardless of the target. Vocabulary collapse reduces predicted amino acids to three to five per position, far below the ground truth distribution in native sequences. Moreover, any model trained with standard per-position cross-entropy converges to the positional marginal distribution, making it provably unable to produce antigen-specific sequence predictions. We propose a novel encoder-decoder architecture called AgForce, that uses a graph neural network (GNN) as the encoder and specialized decoders for sequence-structure co-design. Specifically, we apply framework dropout, gated bottlenecks, and hyperbolic cross attention that prevent the antibody shortcut path. In the decoder, a Mixture Density Network (MDN) sequence head with Potts-like pairwise coupling and annealed Multiple Choice Learning (aMCL) replaces the cross-entropy objective with a multi-component distribution whose optimal solution differs from the positional marginal. An antigen cycle consistency head routes gradients through the sequence decoder, forcing predicted distributions to encode antigen identity. AgForce achieves the best binding quality and sequence recovery simultaneously on the CHIMERA-Bench dataset, improving amino acid recovery by 8% over the strongest sequence baseline while surpassing the baselines across all interface metrics, and nearly doubling the effective vocabulary of GNN methods. The source code is available at: https://github.com/mansoor181/ag-force.git

Abstract

Explore similar work

AgForce Enables Antigen-conditioned Generative Antibody Design

ConTact: Contact-First Antibody CDR Design via Explicit Interface Reasoning

Antigen-specific Antibody Multi-modal Foundation Model for Functional Antibody Design