AAMFM

An antibody binds through a handful of hypervariable loops, and which loop sequence works depends on the antigen surface it has to complement. Protein language models are strong at single-chain sequence and structure modeling but carry no explicit representation of that antibody-antigen pairing, least of all at the epitope level, so they underperform when the task is to design an antibody for a specified target. AAMFM (Antigen-specific Antibody Multi-modal Foundation Model), from Shanghai Jiao Tong University, learns unified representations of antibody sequence and structure conditioned on antigen context: geometric interface features and epitope annotations enter through a cross-modal adapter, so antibody and antigen are modeled jointly in a shared latent space.

Rather than train a new backbone, AAMFM adapts an existing one in three stages. Domain-adaptive continual pretraining specializes an ESM3 checkpoint on paired antibody sequences and structures; supervised fine-tuning then teaches antigen-conditioned CDR design, with the antigen encoded by an adapter derived from GearNet; finally, Calibrated Direct Preference Optimization (Cal-DPO) aligns the generator with binding-specific objectives using preference pairs scored by a structural prior.

That last stage is what separates AAMFM from most antibody generators: diffusion-based designers optimize a reconstruction objective and hope the recovered sequences fold and bind, while AAMFM ranks its own candidates with a cofolding oracle and trains on the resulting preferences. It is antibody-specific, conditioned on a known antigen structure, and complementary to structure-first design pipelines such as Germinal.

#Key Features

• Epitope-level antigen conditioning: Generation is steered by the specific surface being targeted rather than by antibody sequence statistics alone.
• Lightweight cross-modal adapter: The antigen pathway adds 5.1M parameters, about 0.36% of the base model, leaving the pretrained backbone largely intact.
• Preference alignment to a structural oracle: Cal-DPO trains on candidate pairs ranked by predicted foldability and binding rather than by sequence recovery.
• One checkpoint, many targets: The trained model is applied at inference across arbitrary antigens, with sampling controlled by targets and samples per target.
• Full-antibody and CDR-only design: Generation covers whole variable regions or individual loops such as CDR-H3, for single heavy chains or merged heavy-plus-light configurations.

#Technical Details

The backbone is ESM3-open (1.4B parameters). Stage one continues ESM3's masked-token pretraining on roughly 1.4 million paired antibody sequence-structure records from the Observed Antibody Space, with structures supplied by ABodyBuilder2 and IgFold. Stage two fine-tunes on antibody-antigen complexes from SAbDab, filtered to resolutions better than 4 Å and to protein antigens, with clustering at 50% CDR-H3 identity to prevent leakage into the RAbD evaluation set; the GearNet-derived encoder turns antigen geometry into node embeddings that are concatenated with interface features and fused by an MLP before being projected back into the backbone's dimension. Stage three builds 30,000 preference pairs from roughly 13,000 candidates, keeping a pair only when the preferred sequence exceeds the other by at least 0.2 in AF3-score and 0.1 in pseudo log-likelihood (PLL) — AF3-score coming from Protenix 0.4.0, an open reproduction of AlphaFold 3.

On full antibody design the Cal-DPO variant reports PLL of -0.87, AF3-score 0.892, pTM 0.908, and ipTM 0.888, against -0.96 / 0.862 / 0.892 / 0.855 for the strongest baseline, AbX. On CDR-H3 design it reaches 2.62 Å RMSD versus 3.06 Å for AbX, while its amino-acid recovery of 39.91% is lower than AbX's 43.14% — the model favors structurally and energetically plausible loops over literal recapitulation of the native sequence.

#Applications

The natural use is epitope-targeted antibody engineering: given an antigen structure and a chosen epitope, generate and rank CDR variants or complete variable regions for downstream expression and binding assays. Because conditioning is explicit, one checkpoint serves campaigns against unrelated targets — suiting in-silico triage ahead of library construction, loop redesign within an existing scaffold, and affinity optimization.

#Impact

AAMFM shows that a general protein foundation model can be turned into a competitive antigen-specific designer by two cheap moves — adapter-based conditioning and preference alignment against a cofolding scorer — rather than a bespoke generative architecture. Several caveats bound that claim: the work is a preprint and has not been peer reviewed; the designed antibodies have not been validated in vitro, which the authors state and are pursuing with experimental collaborators; and no pretrained weights have been released, so using the model means running the full pipeline against separately obtained data. The repository carries no license.