Loading…
Protein sequence design model that identifies each residue from the voxelized atomic microenvironment around it, reaching 68.33% accuracy on TS500.
No providers recorded yet. Browse all providers
A 20 Å cube of protein interior sampled at 1 Å resolution is 8,000 voxels, and almost all of them are empty — a residue's neighbourhood holds only a few hundred atoms. The 3D convolutional networks built for this representation cope with that sparsity by smearing each atom into a Gaussian cloud, which fills the grid at the cost of blurring away the very contacts that distinguish one amino acid's environment from another's. EMOCPD keeps the atoms sharp and changes the network instead: multi-head self-attention inside the convolutional stack weights the handful of occupied voxels that matter rather than averaging over mostly empty space.
The task is inverse folding posed as classification. For every position in a backbone, the side chain is deleted, the surrounding atoms are voxelized in a local frame centred on the Cβ, and the network predicts which of the 20 amino acids belongs there. Over a whole structure that yields a per-residue probability distribution, readable as a designed sequence or as a ranking of candidate substitutions at chosen sites.
EMOCPD was developed at Jiangnan University in Wuxi, jointly by its School of Artificial Intelligence and Computer Science and its School of Biotechnology, released as a preprint in October 2024 and published in the Journal of Chemical Information and Modeling in 2025. It sits alongside graph-based inverse-folding models such as ProteinMPNN, which the paper cites at 58.1% recovery on TS500, but takes the voxel route rather than encoding the backbone as a graph.
Every residue is represented as a 7 × 20 × 20 × 20 voxel grid built from 1 Å cubes. The network is a Stem module, four iRMB blocks, four MHSA-iRMB blocks and three downsampling stages, followed by an MLP classifier with a 720-unit hidden layer and a 20-way output. Training used PyTorch with Adam, a learning rate of 10⁻⁵ and weight decay of 10⁻³, for 8 epochs of 10,700 steps at 150 samples per step. The corpus is self-constructed from the PDB: chains filtered to at most 50% sequence identity and better than 2.5 Å resolution, with the TS50 and TS500 benchmark proteins removed, leaving 20,118 training proteins and 1,605,000 residue microenvironments plus 298 validation proteins.
Accuracy exceeds 80% on training samples and reaches 62.25% on validation. On the held-out benchmarks EMOCPD recovers 68.33% of native residues on TS500 and 62.32% on TS50, against 55.53%/50.71% for DenseCPD, 42.20%/40.69% for ProDCoNN, 40.25%/39.16% for SPROF and 36.60%/33.60% for SPIN2. Top-K accuracy passes 95% at K = 5, where DenseCPD, ProDCoNN and SPROF need K = 9, 12 and 13. Per-structure accuracy on TS500 spans 44.68% to 89.58%.
The intended use is enzyme engineering by targeted mutagenesis, not de novo design. On wild-type PETase (PDB 5xjh, 263 residues) the model recovers 192 native residues, 73.00%, against 58.94% for DenseCPD. The authors then applied it to the PET-degrading enzyme TfM7 and characterised the proposed mutants experimentally: N213T retained 17.99% relative activity after 12 hours at 70 °C, 1.8 times the wild type, and 8.81% after 24 hours, 2.9 times the wild type, while I83M raised expression to 0.09 mg per OD·mL, also 1.8 times the wild type.
EMOCPD is an argument that the ceiling earlier 3D CNNs hit came from how they handled sparsity, not from the voxel representation itself. The paper is unusually explicit about where its model is weak: correlating per-structure accuracy against composition gives −0.46 for the residues it labels negative and +0.46 for those it labels positive, so designs on proteins rich in the former should be trusted less, and useful gains generally required combining several mutations, a strategy the authors leave open. Code and preprocessed benchmark data are on GitHub without a license, but the trained checkpoint the prediction script expects is not in the repository, so reproducing the reported numbers means retraining from the published recipe.
Much of this page is generated or calculated automatically. Flag anything that looks off and we will re-run it.