AbMEGD

An antibody's binding specificity is set almost entirely by its complementarity-determining regions (CDRs), and the heavy-chain CDR-H3 loop in particular is the hardest part of an antibody to design computationally. Diffusion-based generative models have become the dominant approach to this problem, but most of them operate on a residue-level abstraction — one type, one Cα coordinate, and one orientation frame per amino acid. That representation discards the atom-level geometry that actually mediates contact with an antigen: bond lengths, bond angles, dihedral torsions, and side-chain orientation.

AbMEGD, released as a preprint in June 2025 by researchers at Wuhan University and Macquarie University, addresses that gap with a multi-scale equivariant graph diffusion framework for antibody sequence-and-structure co-design. It keeps the residue-level diffusion formulation established by DiffAb but fuses it with an atom-level geometric encoder adapted from ViSNet, so the denoising network sees both global sequence-structure context and local atomic detail. The whole pipeline is E(3)-equivariant, meaning predictions are consistent under rotation and translation of the complex without data augmentation.

The model sits alongside other catalog entries that generate antibody loops — ABGNN, which pairs an antibody language model with a hierarchical graph network, and ConformAb, which conditions discrete diffusion on canonical CDR backbone conformations — and inherits its broader design philosophy from general-purpose structure diffusion models such as RFdiffusion.

#Key Features

• Multi-scale representation: Residue-level embeddings (amino acid type, local frame, backbone dihedrals, fragment type) are fused with atom-level features over the heavy atoms N, Cα, C, O, and Cβ, giving the model both global context and side-chain-resolution geometry.
• Vector-scalar equivariant message passing: ViS-MP blocks exchange information between scalar and vector node representations, encoding bond angles, dihedral torsions, and improper angles in linear time while preserving E(3)-equivariance.
• Three coupled diffusion processes: Amino acid identity is diffused multinomially over 20 types, Cα coordinates with a Gaussian process, and residue orientations with an isotropic Gaussian on SO(3), so sequence and structure are generated jointly rather than in sequence.
• De novo design and optimization modes: The same trained model supports single-CDR co-design, simultaneous multi-CDR co-design, and affinity optimization of an existing CDR by partial forward-then-reverse diffusion.
• Energy improvement without Rosetta supervision: Binding-energy gains are achieved by the generative model itself, not by an energy function in the training loop.

#Technical Details

Residue and atom features are combined through an invariant point attention module of the kind introduced in AlphaFold 2, then passed to the denoising network. Training data came from SAbDab as of September 2024, excluding complexes with resolution worse than 4 Å and those binding non-protein antigens; antibodies were clustered at 50% CDR-H3 sequence identity, and five clusters comprising 19 antibody-antigen complexes — including SARS-CoV-2, MERS, and influenza antigens — were held out as the test set. Each model generated 100 samples per CDR, with generated and native structures refined using the AMBER99 force field in OpenMM and scored with Rosetta.

Against DiffAb on CDR-H3, AbMEGD raises amino acid recovery from 28.70% to 38.93%, improvement percentage (the fraction of designs with better Rosetta interface energy than the native) from 20.42% to 23.74%, and lowers RMSD from 3.481 Å to 3.419 Å. It also reports the best CDR-H2 recovery (55.40%) and the lowest CDR-H1 RMSD (1.169 Å) among the four methods compared, which include RAbD and DIFFFORCE. In the optimization setting, perturbing CDR-H3 for four diffusion steps reaches an improvement percentage of 31.48% versus 22.18% for DiffAb while keeping RMSD to the original loop at 1.284 Å.

#Applications

The natural use case is therapeutic antibody engineering: redesigning one or more CDR loops against a known antigen epitope, or nudging an existing lead toward better predicted interface energy while preserving its backbone conformation. The optimization mode is the more immediately practical of the two, because it explores a neighborhood of a validated binder rather than sampling loops from scratch. Because AbMEGD conditions on a given antigen structure and antibody framework, it fits into structure-guided campaigns where a co-crystal or predicted complex is already in hand.

#Impact

AbMEGD is an incremental but well-controlled advance on the DiffAb formulation, and its main contribution — showing that atom-level equivariant features improve residue-level antibody diffusion — is a result other CDR generators can adopt directly. Its evaluation is entirely in silico, using Rosetta interface energy as a proxy for affinity, with no experimental binding data; the test set of 19 complexes is small, as is standard in this literature but still a limit on how much the reported gains generalize. The GitHub repository provides training, inference, and design scripts but carries no license file and distributes no pretrained checkpoints, so reproducing the reported numbers requires retraining from SAbDab.