Melo-1-preview

A co-folding model can drop a peptide into roughly the right pocket, score well on interface metrics, and still return coordinates with two heavy atoms 1.4 Å apart — a structure that no downstream affinity calculation or design loop can use. Melo-1-preview is built around treating that as one problem rather than two: it takes near-native pose recovery (DockQ ≥ 0.80) as its primary peptide endpoint, and audits hard atomic overlaps at the interface and inside the peptide chain as a first-class result reported alongside accuracy, not as an afterthought.

The model is an AF3-like all-atom diffusion predictor released on 6 September 2026 by Atomelody, a Shenzhen macromolecule drug discovery company founded earlier that year. Sequence and multiple-sequence-alignment inputs condition single-residue and pairwise representations, which in turn condition an all-atom coordinate denoiser and a confidence head; templates and RNA MSAs are used when available. Proteins, peptides, nucleic acids and small molecules are represented jointly in a single assembly, so peptide and ligand placement is part of the prediction rather than a docking step bolted on afterwards.

It arrives in a crowded lineage. The first wave of AlphaFold 3 reproductions — Boltz-1, HelixFold3, Chai-1 and Protenix — kept AF3's pair representation width of 128, while a second wave including Protenix-v2, OpenDDE and ESMFold2 scaled that width to 256–512. Melo-1-preview sits in the middle of that range and argues the point explicitly: its report credits architectural efficiency and representation quality rather than width, a line of work its authors trace to their own earlier HelixFold-Single and HelixFold-S1.

#Key Features

• Joint accuracy and geometry endpoint: hard clashes — non-bonded heavy-atom pairs closer than 1.6 Å, excluding bonded and two-hop neighbours within the peptide — are audited on every prediction and reported next to DockQ.
• Near-native peptide pose as the target: because current methods already place peptides near the pocket, the primary metric is the strict DockQ ≥ 0.80 bin rather than the acceptable or medium bins.
• Single joint assembly: protein chains, peptides, DNA, RNA and ligands, with ligand–polymer and ligand–ligand covalent bonds, are all denoised together.
• Held-out evaluation by construction: training stops at a 30 September 2021 PDB deposition cutoff, and the peptide and ligand benchmarks are drawn from later depositions and novel pockets.
• Mid-width pair representation: a pair width of 256 places it below the widest contemporary systems while remaining competitive with them.

#Technical Details

Training uses experimental structures from the RCSB PDB with a deposition cutoff of 30 September 2021 — the same training-date alignment AlphaFold 3 used — together with antibody–antigen complexes under that cutoff, plus long-monomer distillation structures and MGnify-derived MSAs from the OpenFold3 monomer distillation release. The parameter count is not disclosed. Evaluation runs on FoldBench, on PXM-22to25 (protein–peptide complexes deposited 2022–2025, released with Protenix-v2), and on the pocket-novelty-oriented Run-N-Pose set, scored with PXMeter and OpenStructure LDDT; every locally run method generates 25 candidates from five seeds on assemblies of up to 2,000 tokens and is ranked by its own confidence score. On FoldBench peptide interfaces (n = 48) Melo-1-preview reaches 47.9% high success against a published AlphaFold 3 reference of 41.2%, and 41.0% on PXM-22to25 (n = 99 clusters); 91.7% of its FoldBench peptide predictions carry neither interface nor peptide-internal clashes. Protein–ligand success is 66.7% on FoldBench (n = 541) and 50.6% on Run-N-Pose (n = 638). The wider profile is asymmetric: 45.6% on protein–protein and 49.4% on protein–DNA, 11.5% on protein–RNA where absolute success remains low for every method, and 20.0% on antibody–antigen, behind OpenDDE at 30.6% and Protenix-v2 at 30.0%.

#Applications

The model targets the structural front end of macromolecule drug discovery: cyclic peptide and nanobody campaigns, where conformational flexibility makes near-native interface modelling the bottleneck, and virtual screening or lead optimisation, where a sterically clean pocket pose is a precondition for affinity prediction rather than a nicety. Atomelody positions it as the folding base of a wider platform covering design, affinity and developability prediction, feeding a protein design system that ends in SPR, ELISA and BLI wet-lab validation.

#Impact

Melo-1-preview is a preview release with no published code, weights, hosted API or web server; the company's technical report is the only primary documentation, and frames the release as a first step with a fuller version to follow. That makes it a capability disclosure rather than a tool the field can pick up, and its claims rest on the developer's own evaluation — one that is at least careful to run baselines under identical inputs and sampling budgets. Its more portable contribution is methodological: pairing DockQ with an explicit clash audit is a reproducible way to check that interface gains are not bought with implausible coordinates, and the same audit shows where the model is weakest, with antibody–antigen near-native recovery the clear remaining deficit.