VenusFold

Training one model to co-fold proteins, nucleic acids and small molecules is a balancing problem before it is an architectural one. The molecule types differ in how often they appear and in how hard their interfaces are to place, and diffusion training adds a second imbalance: each noise level is a different denoising task with its own gradient scale. A minibatch ends up dominated by easy high-signal-to-noise samples, and the optimizer spends its budget where there is least left to learn.

VenusFold is an all-atom biomolecular structure predictor that attacks that imbalance and leaves the network alone. Released by Matwings Technology through the AI4Protein namespace of Liang Hong's lab at Shanghai Jiao Tong University, it follows the AlphaFold 3 design closely — input embedders, MSA and template modules, a Pairformer trunk, and an atom-level diffusion module that denoises coordinates for the whole complex, with distogram and confidence heads alongside. The one architectural change is ReLU giving way to SiLU.

The new part is the training recipe. Minimum signal-to-noise (Min-SNR) weighting caps the contribution of high-SNR diffusion samples over EDM preconditioning, redirecting each update toward under-optimized noise regimes; Momentum Orthogonalized by Newton–Schulz (Muon) rescales the singular values of matrix-shaped updates so no handful of high-gain directions dominates by magnitude alone. Both act only during training, which places VenusFold beside Boltz-2, Protenix-v2, Chai-1, OpenFold3, OpenDDE and ESMFold2 as a drop-in predictor rather than a new paradigm.

#Key Features

• Min-SNR diffusion-loss weighting: A per-sample cap (γ = 5) over EDM preconditioning limits how much high-SNR denoising tasks contribute to the gradient; the weights are deliberately not renormalized by batch mean.
• Muon on matrix-valued parameters: Pairformer and MSA-module matrices update through a Nesterov momentum step whose polar factor is approximated by five Newton–Schulz iterations of pure matrix multiplies; every other parameter falls back to AdamW.
• One checkpoint across every molecule type: Proteins, peptides, DNA, RNA, ligands, ions and modified residues share a single all-atom representation, applied to new targets with no per-system refitting.
• Interface-weighted coordinate loss: The aligned coordinate objective upweights DNA and RNA atoms fivefold and ligand atoms tenfold over protein atoms, so rarer interface chemistry is not drowned out by backbone.
• Confidence-ranked sampling: A head estimating pLDDT, predicted aligned error and predicted distance error lets callers rank candidate poses; the shipped configuration draws five samples over 200 diffusion steps with ten recycling cycles.

#Technical Details

The released checkpoint carries 464,442,431 parameters. Its corpus combines three sources: the PDB training split distributed with Protenix-v2, restricted to structures released on or before 30 September 2021; an AlphaFold Database quaternary-structure distillation set clustered down to 195,657 representative complexes; and Teddymer, 609,271 unique sequences paired with synthetic dimer structures from Encyclopedia of Domains annotations. Alignments came from JackHMMER against UniProt, UniRef90, MGnify and BFD. Training ran on ten nodes of eight 80 GB NVIDIA A800 GPUs, with an EMA decay of 0.999.

On FoldBench, a nine-category benchmark of monomers and complexes, the authors report a mean lDDT-PLI of 0.805 against 0.769 for the strongest competing entry, joint protein–ligand success of 69.59%, and protein-pocket lDDT-LP of 0.916. Docking success reaches 72.73% on protein–protein, 87.11% on protein–DNA and 84.62% on protein–RNA, with a protein-monomer TM-score of 0.940 and protein–peptide success of 90.70%. A 1,000-step ablation isolating the recipe reports rank lDDT of 0.4215 against 0.2685 for an AdamW–EDM baseline. Every figure is self-reported from a single training seed in an unrefereed technical report; the authors flag the ablation as early optimization behavior, not a converged estimate.

#Applications

The obvious use is structure-based drug design: predicted ligand poses together with pocket geometry give medicinal chemists something to triage hits and reason about lead optimization, and the protein–DNA and protein–RNA accuracy extends the same workflow to transcription-factor recognition, nucleic-acid-binding enzymes and gene-editing complexes. Weights and inference code are public, so the model runs on local hardware — it needs an NVIDIA GPU with a CUDA 12.x toolkit, and it consumes precomputed MSA and template files rather than generating them, so MSA-conditioned results need a separate alignment pipeline.

#Impact

VenusFold's contribution is an argument rather than a component: that all-atom co-folding accuracy is still limited by how these models are optimized, not only by how they are designed. Because Min-SNR weighting and Muon touch nothing in the AlphaFold 3 trunk, the recipe is portable to any model in the family. The caveats are real — the benchmarks are the authors' own, run at one seed and published in a self-released report with no peer review or third-party reproduction; antibody–antigen complexes improved least; and the repository ships inference only, with training code and data manifests described but not released. Code is MIT-licensed and the weights Apache-2.0, so the checkpoint is usable even where the recipe behind it is not reproducible.