Neural network writes amino acid sequences to fit fixed protein backbones

The work

Technical · from the record

The authors trained a message passing graph neural network, ProteinMPNN, to write amino acid sequences for a given protein backbone, including backbones made of multiple chains and backbones with symmetry constraints that tie residue identities across positions. On a test set of 402 monomer backbones it recovered 52.4% of native sequence positions against 32.9% for Rosetta fixed-backbone design, and took 1.2 seconds rather than 4.3 minutes per 100 residues. Sequences were then generated for backbones whose earlier Rosetta or AlphaFold-derived sequences had failed experimentally: of 96 designs produced in E. coli, 73 expressed solubly and 50 had the target monomeric or oligomeric state by size exclusion chromatography, while 13 of 76 tetrahedral nanoparticle sequences formed assemblies of the expected mass. Crystal, cryoEM and negative stain EM structures of selected designs were compared against the design models.

How AI was used

Protein assemblies from the PDB solved by X-ray crystallography or cryoEM to better than 3.5 A and under 10,000 residues were clustered at 30% sequence identity with mmseqs2 into 25,361 clusters, split into training, validation and test groups. Backbones were encoded only as graph edge features: radial basis function distances between N, Ca, C, O and a virtual Cb atom, a relative positional encoding capped at plus or minus 32 residues, and a binary same-chain or different-chain indicator. A three-layer encoder with edge updates and a three-layer decoder, each with 128 hidden dimensions, were trained in PyTorch on a single A100 GPU with negative log likelihood and label smoothing, using a decoding order sampled at random from all permutations so that positions can be fixed or tied; Gaussian noise was added to backbone coordinates during training at several levels, and a Ca-only variant was trained separately. At inference, sequences were sampled at a chosen temperature for fixed backbones, with residue identities at symmetry-equivalent positions tied by averaging logits, and ranked by the model's averaged log probability of the sequence given the structure. Designed sequences were passed as single sequences to five AlphaFold ptm models with three recycles, taking the highest average pLDDT model, and compared with the target backbones by lDDT-Ca, pLDDT and inter-chain PAE before selected designs were encoded in synthetic genes, expressed in E. coli and characterised.