APM

APM, the All-Atom Protein Generative Model, is a flow-matching generative model built for multi-chain proteins. Most proteins do their work in complexes, yet the generative models that transformed single-chain design largely treat multimers as an afterthought: a common workaround chains subunits together with a poly-glycine linker and folds the result as one sequence, forcing an unnatural head-to-tail topology onto assemblies that occupy independent spatial positions. APM models multiple chains natively, and at atomic resolution — the inter-chain contacts that determine whether two subunits bind are made by side-chain atoms, not backbone frames. It was developed at ByteDance Seed with collaborators at Hunan University and the University of Chinese Academy of Sciences, and presented at ICML 2025.

The model sits in the lineage of sequence-structure co-design methods that generate an amino acid sequence and its backbone jointly under a flow-matching objective, and extends that family in two directions: side-chain conformations are generated and then used to refine the structure, and folding and inverse folding are learned alongside unconditional generation. The result is one checkpoint that generates complexes from scratch, folds and inverse-folds multimers, and serves as a base for functional design tasks such as antibody CDR-H3 co-design and receptor-targeted peptide design.

#Key Features

• Native multi-chain modeling: Chains are represented separately with inter-chain and intra-chain attention rather than concatenated through a pseudo-linker, so generated complexes have independent chain placements with complementary binding interfaces.
• Three-module all-atom pipeline: A Seq&BB module co-generates sequence and backbone under flow matching, a Sidechain module predicts torsions, and a Refine module corrects the complex using the resulting all-atom geometry.
• Decoupled sequence and structure noising: The two modalities are corrupted on independent time schedules, and half of all training steps are folding or inverse-folding tasks, preserving the sequence-structure dependency that joint noising otherwise destroys.
• Protein language model conditioning: Per-layer representations from the 650M-parameter ESM-2 are aggregated with learned weights to encode input sequences, compensating for the limited volume of structural training data.
• Zero-shot and fine-tuned use: The released checkpoint samples binders and peptides zero-shot, while supervised fine-tuning pushes accuracy further.
• Open code and weights: Code is Apache-2.0 on GitHub; checkpoints and processed training data are on Zenodo under CC BY 4.0.

#Technical Details

APM totals 127M parameters across its three modules — 52M in Seq&BB, 22M in Sidechain, and 54M in Refine — with residue and pair representations at 384 and 192 dimensions in the Seq&BB and Refine modules and 256 and 128 in the smaller Sidechain module. Side chains are parameterized by four torsion angles per residue rather than explicit atom coordinates, keeping the all-atom representation compact. Training draws on 187,494 single-chain structures — 18,684 from the PDB, 140,769 Swiss-Prot predictions filtered at pLDDT above 85, and 28,041 AlphaFold DB structures filtered at pLDDT above 95 — plus 11,620 multi-chain PDB biological assemblies, with SAbDab entries and peptide-containing assemblies removed to prevent leakage into downstream benchmarks. The Seq&BB and Sidechain modules are trained separately, then all three jointly in an iterative loop.

On unconditional complex generation against Chroma, APM produces tighter interfaces at every chain-length combination tested; for 100-100 pairs it reaches an average all-atom relaxed binding energy of -130.31 Rosetta energy units against Chroma's +102.33. Ablating the all-atom modules to backbone-only generation weakens binding and raises relaxation RMSD, isolating side-chain information as the source of the gain. On the RAbD antibody benchmark, fine-tuned APM leads all four antigen-specific baselines on amino acid recovery, Cα RMSD, total energy, and binding energy. On the LNR peptide test set, fine-tuned APM reaches -19.90 average binding energy with 69.34% of designs below zero, far ahead of the peptide-specific baselines PPFlow, DiffPP, and PepGLAD (12.50–38.17% below zero); the zero-shot checkpoint posts the best raw binding energy of any method at -23.71, while RFdiffusion leads on the proportion below zero and on the composite foldability success rate.

#Applications

APM targets the design of bioactive protein complexes: de novo binders against a chosen target, antibody CDR loops conditioned on an antigen, and receptor-targeted peptides. Because the same checkpoint also performs multimer folding and inverse folding, it can span generation and evaluation within a single design loop, and the chain-by-chain generation mode lets a designer fix one chain and sample a partner. The open code and weights make it usable as a practical design tool and as a base model for further fine-tuning.

#Impact

APM treats multi-chain, all-atom protein generation as a single unified problem rather than stitching together single-chain components, and its ablations give direct evidence that atomic side-chain detail — not just backbone geometry — drives interface quality. The authors name three limitations plainly: folding accuracy still trails specialized predictors, the Refine module's role is narrow, and downstream evaluation covers only antibody and peptide design. All benchmarking is computational, relying on pyRosetta binding energies and structure-prediction confidence scores that correlate with but do not substitute for experimental binding measurements.