FlowAB

Give a modern antibody structure predictor a heavy and light chain and five of the six complementarity-determining regions come back within about an ångström of the crystal structure. The sixth, CDR-H3 — the loop that makes most of the antigen contact — comes back roughly three ångströms out. The conventional repair is a physics relaxation: Rosetta or OpenMM, tens to hundreds of seconds per structure, minimizing an energy that knows nothing about which parts of the prediction were already right.

FlowAB, from Shenzhen's International Digital Economy Academy, reframes that repair as a transport problem. Rather than sampling a backbone from noise, it learns an SE(3) vector field whose source distribution is an existing CDR predictor's output and whose target is the native structures in SAbDab. Because the two are already close — residue i in the prediction corresponds to residue i in the crystal — the flow needs no optimal-transport coupling and converges in two integration steps (Δt = 0.5), against the thousands of iterations a comparable diffusion model needs. Refining one antibody takes 0.25 seconds.

A flow defined on residue frames has a failure mode, though: each frame moves under its own vector field, with nothing tying neighbouring residues together, so the peptide chain can be pulled apart. FlowAB's second component answers this. Bond, angle and improper-torsion potentials around the peptide bond enter the sampling ODE as a gradient added to the velocity, pushing every step toward a physically plausible amide plane. The vector field is trained once; the guidance is applied purely at sampling time.

#Key Features

• Prediction as the prior: The flow's starting distribution is another model's output — the paper uses ABlooper and DiffAB — so this is a post-hoc pass that composes with whatever CDR predictor a lab already runs, not a replacement for one.
• Energy guidance at sampling time: A peptide-bond length potential (equilibrium 1.32 Å), two angle potentials (114° and 123°) and an improper torsion holding four backbone atoms on the amide plane at 180° are differentiated into the velocity, with an inverse temperature controlling their strength.
• Frame-level SE(3) modelling: An ℝ³ flow moves Cα coordinates and an SO(3) flow adjusts residue orientations — the residue-frame representation also used for de novo backbone generation by models such as FrameFlow.
• Scoped to the CDRs: Only the six hypervariable loops move; framework and antigen are held fixed and fed to the network as conditioning context.

#Technical Details

MLPs embed each residue and residue pair — amino acid type, torsion angles, heavy-atom coordinates, pairwise distances and dihedrals — Invariant Point Attention modules borrowed from AlphaFold2 mix them, and two heads read out the ℝ³ and SO(3) vector fields. Training regresses both on SAbDab, with an auxiliary loss on pairwise heavy-atom distances under 0.6 nm in the second half of the time interval.

Evaluation uses two established splits. On the IgFold split (197 test antibodies), ABlooper + FlowAB reaches backbone RMSDs of 0.82, 0.70 and 2.98 Å on CDR-H1/H2/H3 and 0.70, 0.38 and 1.03 Å on L1/L2/L3 — the best figure on all six loops in the paper's comparison, ahead of IgFold at 3.27 Å on H3 and AlphaFold-Multimer at 3.56 Å, in 1.48 seconds against 23 seconds for IgFold with Rosetta and 174 for ABlooper with OpenMM. On the smaller DiffAB split, refinement cuts ABlooper's CDR-H3 error from 2.45 to 2.20 Å and DiffAB's from 3.25 to 2.70 Å. An ablation shows flow matching and energy guidance each help alone and compound when combined.

#Applications

FlowAB fits antibody engineering workflows that generate structures in bulk: screening a designed library, preparing models for docking or interface analysis, or cleaning up predictions before an affinity estimate that depends on CDR-H3 geometry. Taking a predicted complex and returning refined loops in a fraction of a second, it is cheap enough to run across a whole repertoire rather than a handful of leads — the regime where a Rosetta or molecular dynamics relaxation becomes prohibitive.

#Impact

Presented as a regular paper at IEEE BIBM 2024, FlowAB is the first flow-matching model built specifically for antibody structure refinement, and its recipe — train an unconditional vector field once, then steer it with a differentiable physical potential at sampling time — is a general pattern for adding physics to a generative structure model without retraining it. The authors state the limitations plainly: the potential covers bond, angle and torsion terms rather than a full force field, and the state-of-the-art claim is conditional on an appropriate prior model, since FlowAB refines an existing backbone and cannot predict one from sequence. It is a different tool from FlowDesign, which also applies flow matching to CDRs but co-designs sequence and structure rather than correcting geometry. Neither code nor trained weights have been released.