AbFlowNet: Optimizing Antibody-Antigen Binding Energy via Diffusion-GFlowNet Fusion
Authors: Abrar Rahman Abir, Haz Sameen Shahgir, Md Rownok Zahan Ratul, Md Toki Tahmid, Greg Ver Steeg, Yue Dong
Organizations: Bangladesh University of Engineering and Technology · University of California Riverside · University of Southern California · Princeton University
Complementarity Determining Regions (CDRs) are critical segments of an antibody that facilitate binding to specific antigens. Current computational methods for CDR design utilize reconstruction losses and do not jointly optimize binding energy, a crucial metric for antibody efficacy. Rather, binding energy optimization is done through computationally expensive Online Reinforcement Learning (RL) pipelines rely heavily on unreliable binding energy estimators. In this paper, we propose AbFlowNet, a novel generative framework that integrates GFlowNet with Diffusion models. By framing each diffusion step as a state in the GFlowNet framework, AbFlowNet jointly optimizes standard diffusion losses and binding energy by directly incorporating energy signals into the training process, thereby unifying diffusion and reward optimization in a single procedure. Experimental results show that AbFlowNet outperforms the base diffusion model by 3.06% in amino acid recovery, 20.40% in geometric reconstruction (RMSD), and 3.60% in binding energy improvement ratio. ABFlowNet also decreases Top-1 total energy and binding energy errors by 24.8% and 38.1% without pseudo-labeling the test dataset or using computationally expensive online RL regimes.
Figures & tables
Figure 1: AbFlowNet reframes the diffusion process as a GFlowNet where each partially denoised CDR is a state and the transition probabilities are flows through edges . The initial state’s flow is learned, and the final state’s flow is the binding energy of the reference CDR. To train, we simply enforce forward and backward flow parity, in addition to the diffusion losses.
Figure 2: Overview of antibody-antigen complex structure. Left: Y-shaped antibody composed of heavy (V_H) and light (V_L) chains. Center: the variable domain contains six CDR loops, which determine antigen recognition. Right: structural view of the complex with the CDR-H3 loop highlighted.
Methods
Sampling
CDR Etotal (↓)
CDR-Ag ΔG (↓)
Test Set Used
# Param. (M)
Budget
(with Δ )
(with Δ )
in Sampling
Reference
4.52
-13.72
—
—
GNN Baselines
HERN [ 24 ]
2,528
7594.94
1159.34
—
5
MEAN [ 29 ]
2,528
3113.70
114.98
—
-
dyMEAN 1 [ 30 ]
2,528
15025.67
2391.00
—
-
Table 1: Top-1 CDR Etotal and CDR-Ag ΔG (kcal/mol) for de novo CDR-H3 design. (↓) indicates lower is better. We also show the percentage reduction over DiffAb with the same sampling budget. AbX and AbNovo report Mean CDR Etotal and CDR-Ag ΔG instead of Top-1.
Figure 3: Top-1 CDR-Ag ΔG on the RAbD test set as a function of sampling budget N . AbDPO’s reported value ( −4.85 at N=2,528 ) is shown for reference. Exact values are reported in Appendix D
CDR
DiffAb
AbFlowNet
H1
-8.000
-9.067
H2
-8.399
-9.605
H3
-2.112
-3.246
L1
-8.463
-9.295
L2
-8.505
-9.542
L3
-7.058
-7.934
Table 2: FoldX ΔG on RAbD ( ↓ , kcal/mol).
Figure 4: De novo Generated and Reference CDR-H3s for 4dvr complex. For DiffAb and AbFlowNet, we generated 100 CDRs and selected the one with the lowest ΔG .
Appendix figures & tables17 assets
Supplementary material from the paper’s appendix.
Appendix
Table 3: Evaluation of the generated antibody CDRs (sequence-structure co-design) on the RAbD test dataset (60 sequences) using AAR (%), RMSD (Å), and IMP (%) metrics.
Figure 5: Hyperparameter search for TB loss weight w in Eqn. 17 on the RAbD [ 3 ] dataset. The RMSD of DiffAb on the L3 CDR region is significantly worse than that of AbFlowNet. We retrained and evaluated DiffAb and AbFlowNet using a different seed to confirm this discrepancy (retrained DiffAb RMSD 4.06 Åand retrained AbFlowNet RSMD 1.362 ).
Figure 6: Training Diffusion+GFlowNet models with different training steps on the RAbD dataset [ 3 ] . Separating the reconstruction and flow matching steps do not meaningfully improve performance over AbFlowNet.
Table 4: Evaluation of state-of-the-art CDR design methods on the RAbD test dataset. The DiffAb baseline has been independently benchmarked using different parameters by AbDPO [ 60 ] , AbNovo [ 43 ] , and IgGM [ 51 ] , which we denote in parentheses.
N
DiffAb
AbFlowNet
Improvement
5
81.16
45.95
35.21
10
39.38
25.25
14.13
50
18.71
3.37
15.34
100
11.20
1.71
9.49
500
6.51
-0.10
6.61
1000
5.59
-1.15
6.74
Appendix
Table 5: Top-1 CDR-Ag ΔG (kcal/mol) on the RAbD test set as a function of sampling budget N . AbFlowNet remains uniformly better than DiffAb across all budgets.
CDR
DiffAb
AbFlowNet
DiffAb-MPNN
AbFlowNet-MPNN
H_CDR1
-8.00
-9.07
-7.67
-8.81
H_CDR2
-8.40
-9.60
-7.68
-8.86
H_CDR3
-2.11
-3.25
-2.39
-3.84
L_CDR1
-8.46
-9.30
-8.43
-9.31
L_CDR2
-8.50
-9.54
-7.82
-8.83
L_CDR3
-7.06
-7.93
-6.41
-7.50
Appendix
Table 6: FoldX ΔG (kcal/mol) on the RAbD test set under co-design and structure → sequence variants. Bold indicates best per row.
Metric
DiffAb
AbFlowNet
Mean CDR Etotal (kcal/mol)
503.39
396.19
Mean CDR-Ag ΔG (kcal/mol)
25.47
10.03
Appendix
Table 7: Mean Top-1 metrics on the novel-antigen subset of RAbD (CDR-H3, N=100 ). Lower is better.
CDR
DiffAb
AbFlowNet
H_CDR1
70.24
70.99
H_CDR2
44.09
44.81
H_CDR3
29.51
31.65
L_CDR1
58.71
61.89
L_CDR2
60.04
62.92
L_CDR3
53.20
54.53
Appendix
Table 8: Mean pairwise sequence identity (%, lower is more diverse).
Table 9: Evaluation of the generated antibody CDRs (sequence-structure co-design) on the DiffAb test dataset (19 sequences).
Methods
# Samples
CDR
+Etotal
CDR-Ag
+ΔG
Test Set
Etotal ( ↓ )
(%)
ΔG ( ↓ )
(%)
Used
Reference
1.63
-4.80
DiffAb
100
26.33
11.50
AbFlowNet (Ours)
4.23
89.5
1.47
149.7
No
Appendix
Table 10: Summary of Top-1 CDR Etotal and CDR-Ag ΔG (kcal/mol) of CDR-H3’s designed by DiffAb and AbFlowNet on the DiffAb Test Dataset. ( ↓ ) denotes that a smaller number is better.
Figure 7: tSNE clustering of ESM-2 embeddings of CDRs generated by AbFlowNet and the reference CDR of the RAbD dataset. The reference CDR is generally in the same cluster as the generated CDRs, showing that these are evolutionarily plausible and possess natural-like sequence properties according to ESM-2.
Figure 8: Distribution of sample CDRs generated by AbFlowNet on the RAbD test dataset.
Complex
Energy of Complex Ecomplex (↓)
Binding Energy CDR-Ag ΔG (↓)
Total CDR Energy Etotal (↓)
Min
Mean ± Std
Ref
Min
Mean ± Std
Ref
Min
Mean ± Std
Ref
1a14_H_L_N
465.24
739.27 ± 349.05
1829.81
-23.46
5.76 ± 154.80
185.48
18.46
151.25 ± 168.06
56.52
1a2y_B_A_C
-695.23
-643.06 ± 86.15
-412.00
-20.92
-15.31 ± 2.99
17.32
-9.88
10.63 ± 12.45
-13.09
1fe8_H_L_A †
66.58
2244.82 ± 1279.23
-300.06
151.96
1998.45 ± 1199.38
62.53
212.53
1444.82 ± 731.12
29.13
1ic7_H_L_Y †
94.59
920.13 ± 702.52
29.49
54.26
250.78 ± 308.75
75.85
54.36
541.44 ± 397.55
17.72
1iqd_B_A_C
-591.29
-464.97 ± 75.50
-60.88
-37.13
16.03 ± 35.73
206.79
-1.97
33.88 ± 21.23
41.34
Appendix
Table 11: Mean, standard deviation, and minimum of Energy Metrics of CDRs generated by AbFlowNet on the RAbD test dataset. † denotes complexes that could not be processed with PyRosetta FastRelax.
Figure 9: De novo Generated and Reference CDR-H3s for 4dvr complex. We selected the one with the highest ΔG out of 100 generations for DiffAb and AbFlownet.
Figure 10: De novo Generated and Reference CDR-H3s for 3rkd complex.
Figure 11: De novo Generated and Reference CDR-H3s for 5hi4 complex.
CDR
Method
ΔG
CDR
Method
ΔG
H-CDR1
DiffAb
-58.18
L-CDR1
DiffAb
-56.23
RW
-57.97
RW
-56.12
AbFlowNet
-58.98
AbFlowNet
-56.32
RW + AbFlowNet
-57.40
RW + AbFlowNet
-55.84
H-CDR2
DiffAb
-56.03
L-CDR2
DiffAb
-56.29
RW
-54.50
RW
-56.35
Appendix
Table 12: Mean Top-1 CDR ΔG over RAbD test complexes per CDR. RW denotes Reward-weighted.
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
Mansoor Ahmed, Murray Patterson
1Georgia State University, Atlanta, GA, USA · 2Georgia Institute of Technology, Atlanta, GA, USA
Antibodies are essential therapeutic molecules, and their complementarity-determining regions (CDRs) form the primary antigen-recognition interface. Recent protein generative models have demonstrated broad capabilities in biomolecular design, yet post-training strategies for downstream objectives remain limited. Standard denoising training operates on noisy states obtained by perturbing native structures, whereas recursive generation proceeds through model-generated intermediate states. For flexible antibody CDR loops such as CDR-H3, this mismatch can allow backbone deviations to accumulate along the denoising trajectory and compromise antigen-facing loop geometry. We introduce ABOPD, an antibody design framework based on on-policy distillation that leverages privileged native geometry during training to supervise states visited along the model's own denoising trajectories. With this fine-grained structural supervision, ABOPD substantially improves structural recovery on RAbD CDR-H3 generation, reducing RMSD by 0.42 Å (from 2.37 Å to 1.95 Å) and outperforming supervised fine-tuning and offline distillation controls, offering a path to higher-fidelity protein design.
Zhuo Yang, Jiaying He, Jiaqing Xie +5
Shanghai Innovation Institute · Southeast University · Zhejiang University +4
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%.
Jiashuo Wang, Siqi Fan, Yizhen Luo +1
Institute for AI Industry Research (AIR), Tsinghua University · Department of Computer Science and Technology, Tsinghua University · PharMolix Inc.