AbFlowNet

Complementarity-determining regions are the loops that give an antibody its specificity, and designing them computationally means searching a space of 20^L sequences before structural variation is even considered. Diffusion models learned to co-design CDR sequence and backbone geometry conditioned on an antigen, but they are trained purely on reconstruction: they learn to recover the reference CDR, not to produce one that binds tightly. Binding energy has therefore been bolted on afterward, through online reinforcement learning that samples thousands of candidates per target and scores them with a physics-based estimator — expensive, dependent on that estimator's reliability, and, for preference-optimization approaches, often reliant on the test complexes themselves to build the preference data.

AbFlowNet removes the second stage. It reframes each step of the reverse diffusion process as a state in a GFlowNet and trains with the trajectory balance objective, so that a binding energy reward propagates back through the whole denoising trajectory alongside the ordinary diffusion losses. The energies come from the training set, precomputed once, rather than from sampling during optimization. The model was developed by researchers at Bangladesh University of Engineering and Technology, UC Riverside, USC, and Princeton, and posted as a preprint in May 2025.

It sits alongside other structure-conditioned antibody generators such as DiffAbOpt, which instead steers diffusion with oracles fitted to campaign assay data, and general-purpose designers like RFdiffusion.

#Key Features

• Diffusion steps as GFlowNet states: Each denoising transition is treated as a state transition in a flow network, letting the trajectory balance objective distribute a terminal energy reward across every step of generation.
• Single-stage joint optimization: Reconstruction losses and the binding energy signal are optimized in one training procedure, so no separate reward-model or post-training phase is required.
• Precomputed rather than online rewards: Binding energies for the 9,410 training complexes are calculated once with Rosetta, eliminating per-target sampling and reward evaluation during optimization.
• No test-set leakage: Unlike preference-optimization baselines that construct preference data from the evaluation complexes, AbFlowNet never touches test structures during training.
• Structural accuracy preserved: Jointly optimizing reconstruction and energy avoids the degradation in amino-acid recovery and RMSD that reinforcement-learning post-training typically causes.

#Technical Details

The generator uses the transformer parametrization introduced by DiffAb, jointly diffusing CDR residue types, Cα coordinates, and residue orientations conditioned on the surrounding antibody framework and antigen. AbFlowNet adds a single learnable global partition parameter for the trajectory balance objective. Training data comes from SAbDab, filtered to structures better than 4 Å with protein antigens and clustered at 50% CDR-H3 sequence identity, yielding 9,410 antibody-antigen complexes; each is relaxed with constrained Rosetta FastRelax and scored with InterfaceAnalyzer to obtain total energy and interface ΔG labels. Both AbFlowNet and its DiffAb baseline are trained for 200,000 Adam steps at a learning rate of 1e-6, with the trajectory balance term enabled only for the final 5,000 steps at weight 5e-6, since sampling a full trajectory costs roughly 20 seconds per step.

Evaluation covers the 60-complex RAbD benchmark and the 19-complex DiffAb test set, which includes antigens from SARS-CoV-2, MERS, and influenza. Averaged across all six CDR regions, AbFlowNet improves amino-acid recovery by 3.06% over its diffusion baseline, RMSD by 20.40%, and the binding energy improvement ratio by 3.60%. On de novo CDR-H3 design at a sampling budget of 100, Top-1 CDR total energy falls from 480.25 to 362.03 kcal/mol (down 24.8%) and interface ΔG from 11.20 to 1.71 kcal/mol (down 38.1%), reaching gains comparable to a reinforcement-learning method that used 10,122 samples per test complex.

#Applications

AbFlowNet is aimed at in-silico antibody engineering — generating CDR-H3 candidates against a known antigen structure, or redesigning all six loops of an existing antibody — for groups that want energy-aware generation without standing up a reinforcement-learning pipeline and a per-target reward oracle. Because the reward is precomputed over the training set, adapting the recipe to a different objective means recomputing a dataset-level energy label rather than building an online scoring loop, which makes it a practical template for other structure-conditioned generative design tasks.

#Impact

The methodological point is that GFlowNet trajectory balance offers a cheaper alternative to online reinforcement learning for injecting a physical objective into a diffusion model, and that doing so jointly avoids the reconstruction degradation that post-hoc reward optimization introduces. The caveats are equally clear. All results are computational, with Rosetta InterfaceAnalyzer energies serving as the proxy for binding — a proxy the paper itself notes is imperfect — and no experimental validation is reported. The improvement comes from just 5,000 trajectory-balance steps at the end of a long standard training run, so the effect of extended joint training is unresolved. Training code, inference code, the preprocessing chain and both the AbFlowNet checkpoint and its retrained DiffAb baseline are committed directly to the repository under Apache-2.0.