ProxelGen

Protein structure generative models have been explored along many axes — architectures, diffusion versus flow matching, which geometric priors to encode — but almost all share one unexamined choice: the protein is a point cloud. Whether the model operates on residue frames, alpha-carbon coordinates, or all atoms, its representation grows with the number of residues, forcing the designer to declare a protein length before generation begins and tying conditioning to sequence indices. ProxelGen, from Barzilay and Jaakkola's groups at MIT CSAIL, changes the representation instead of the model.

A protein is written as a multi-channel 3D density array — proxels, by analogy to pixels — sampled on a fixed grid. Generation then becomes a problem the image generation literature already knows how to solve: a convolutional autoencoder compresses the grid, and a latent flow model generates in that compressed space. Because the array has fixed dimensions regardless of protein size, ProxelGen needs no target length, and because conditioning inputs live in the same spatial grid, a designer can mask out a region of space or supply an arbitrary shape and have it respected directly.

#Key Features

• Density representation with chain ordering: Seven channels encode Gaussian densities around N, CA, and C atoms, a bond channel at backbone bond midpoints, and a three-component "chain flow" vector field running N-terminus to C-terminus, which is what lets an ordered chain be recovered from an unordered density.
• Length-free generation: Grid dimensions are independent of residue and chain count, so no protein length or chain composition is specified in advance — a constraint every atomistic generative model imposes.
• Spatial inpainting for motif scaffolding: A motif is supplied as occupied voxels and the rest of the space is masked. Unlike prespecified methods, ProxelGen does not require the designer to declare where in the designed sequence each motif segment appears, which is the binding constraint on multi-segment motifs.
• Arbitrary shape conditioning: A binary voxel mask derived from a protein surface, or any other geometry, serves as a target envelope that the model fills.
• A density-native quality metric: ProxCLR, a 3D ResNet trained with SimCLR on proxels, supplies embeddings for a Fréchet distance computed directly on densities rather than on structural descriptors.

#Technical Details

Proxels are 32 x 32 x 32 grids at 1.5 Å spacing with a 1 Å Gaussian kernel, so consecutive alpha-carbons rarely land in the same cell. A beta-VAE with 3D convolutions and self-attention compresses these by a factor of 512 in dimension, its encoder constrained to unit covariance to keep the latent geometry regular given how sparse proxels are. A 3D UNet flow model, trained as a linear stochastic interpolant with timesteps oversampled near t = 0.2, generates the latents; a transformer alternative performed substantially worse. To recover coordinates, a 60M-parameter Proteina model is fine-tuned to decode latent proxel tokens and a pretrained 200M model refines the result. Training uses the FoldSeek-clustered AlphaFold Database subset from Genie 2 restricted to structures fitting the grid — 98,584 structures split 93,654 / 4,930 — on single A6000 GPUs.

On 256 unconditional samples, ProxelGen reaches an FID of 6.05, novelty of 0.73, and 53.13% designability, close to the 50.39% of native structures round-tripped through proxels; low-temperature Proteina reaches 97.77% designability at an FID of 20.47 and much lower diversity. On the 24-task RFdiffusion motif scaffolding benchmark, scored with eight ProteinMPNN sequences per design folded by ESMFold, ProxelGen is strongest where motifs are fragmented: 12 unique successes on the four-segment 1BCF motif against one for every baseline, and 51 on the three-segment 1QJG against 18, 5, 3, and 1. On single-segment motifs it often trails. Shape-conditioned generation achieves an F1 of 88.6 against the input envelope at a TM-score of 0.60 to the original structure.

#Applications

The representation suits design problems stated geometrically rather than sequentially: scaffolding functional sites split across discontinuous segments, fitting a binder into a known surface pocket, or generating proteins that occupy a specified volume. The unconditional model also works as a distribution-matching generator, with secondary structure content and higher-order contact counts closer to natural proteins than designability-optimised baselines produce.

#Impact

ProxelGen is the first work to apply density-based generation to proteins, after the idea took hold in small-molecule design, and it shows that latent diffusion — repeatedly disappointing on atomistic protein representations — works once the representation has fixed dimensions. Its clearest limitation is stated by the authors: generated proxels frequently fail to form a single connected chain, with the chain flow splitting or merging, and the coordinate decoder threads a chain through anyway at the cost of unnatural kinks. Connectivity is a global property that convolutional models are poorly suited to enforce. The decoder is also motif-unaware, filtering out a substantial fraction of scaffolding designs before scoring. No code or pretrained weights have been released.