Diffusion Models for Proteins

The generative framework operates in two phases. The forward process systematically corrupts a native protein structure by adding Gaussian noise to its 3D atomic coordinates over a fixed number of timesteps, eventually destroying all structural information. The reverse process is a learned denoising neural network that iteratively removes this noise, starting from random coordinates and progressively refining them into a valid protein backbone. This is directly analogous to image diffusion models like DALL-E, but operates on SE(3) space—the continuous group of 3D rotations and translations—rather than pixel grids.

Rather than operating on raw atomic coordinates, diffusion models represent each residue as a rigid-body frame consisting of a 3D translation vector and a 3x3 rotation matrix. This parameterization captures the backbone geometry—specifically the Cα carbon position and the orientation of the local N-Cα-C plane. The denoising network predicts incremental updates to these frames at each timestep. This approach naturally enforces the constant bond lengths and angles of protein geometry, dramatically constraining the generative space to physically plausible conformations.

To generate functionally relevant proteins, the denoising process must be steerable. Common conditioning signals include:

• Target motif specification: Fixing the coordinates of a functional site (e.g., a catalytic triad) while generating the surrounding scaffold.
• Secondary structure constraints: Guiding the model to produce specific helix or sheet patterns.
• Symmetry operators: Enforcing cyclic or dihedral symmetry for oligomeric assemblies.
• Text-based prompts: Using protein language model embeddings from ESM-2 to condition generation on sequence-level functional descriptions. These signals are injected via cross-attention or feature-wise linear modulation layers within the denoising U-Net or equivariant backbone.

The model is trained to predict the score function—the gradient of the log probability density with respect to the atomic coordinates—rather than directly predicting the denoised structure. At each noise level, the network learns to estimate the direction and magnitude of the noise added to the clean structure. The loss function is typically a weighted mean squared error between the predicted and true noise, with weights that emphasize intermediate noise levels where structural features begin to emerge. This score-matching framework, formalized in denoising score matching, provides a stable training signal.

A mathematical property ensuring a neural network's predictions transform consistently with the rotation and translation of input 3D coordinates. In protein diffusion, the denoising network must be SE(3) equivariant so that rotating the noisy input frame produces an identically rotated clean prediction.

• Translation equivariance: Shifting all atom coordinates shifts predictions identically
• Rotation equivariance: Rotating the input rotates the output correspondingly
• Frame-based representations: Residues are parameterized as local coordinate frames that transform rigidly
• Contrast with SE(3) invariance, where outputs remain unchanged under input transformations

The computational creation of entirely novel protein sequences and structures that do not exist in nature, designed to perform a specific function. Diffusion models excel here by generating backbone structures unconditionally or conditioned on functional motifs.

• Unconditional generation: Produces diverse, physically plausible backbones without target constraints
• Motif-scaffolding: Fixes a functional site (e.g., an enzyme active site) and generates a supporting scaffold around it
• Inpainting: Replaces specific structural regions while preserving the rest
• Tools like RFdiffusion and FrameFlow have experimentally validated designs for binders, enzymes, and symmetric oligomers

A core mechanism from AlphaFold2 that operates on 3D point clouds, updating representations in a way that is invariant to global rotation and translation. IPA computes attention weights using both sequence features and pairwise 3D distances between residue frames.

• Projects local coordinate frames into a global invariant space
• Uses query, key, value projections on both sequence and geometric features
• Preserves the relative spatial relationships critical for structure prediction
• Serves as a foundational building block adapted by many protein diffusion architectures for their denoising networks

A deep learning-based tool for inverse protein folding that generates amino acid sequences predicted to fold into a given 3D backbone structure. It is the standard sequence design partner for diffusion-generated backbones.

• Message-passing neural network operating on k-nearest neighbor graphs of backbone coordinates
• Encodes spatial environments including backbone dihedral angles and pairwise distances
• Autoregressive decoding generates sequences residue by residue
• Achieves higher experimental expression and stability rates than physics-based design tools like Rosetta
• Typically used after a diffusion model generates a backbone to produce the corresponding sequence for experimental testing

A per-residue confidence metric output by AlphaFold2, estimating the local accuracy of a predicted structure on a scale from 0 to 100. While originally for structure prediction, the concept of per-residue confidence scoring is critical for evaluating diffusion-generated designs.

• pLDDT > 90: High confidence, suitable for detailed structural analysis
• pLDDT 70-90: Good backbone prediction, side-chain details less reliable
• pLDDT < 50: Low confidence, may indicate disorder
• Diffusion-specific analogs include self-consistency metrics where the same backbone is fed through inverse folding and re-folding to check structural reproducibility

The application of physics-based simulations using force fields like AMBER or CHARMM to optimize a predicted or generated protein structure by resolving atomic clashes and bond geometry violations. Diffusion-generated backbones often require MD relaxation.

• Energy minimization: Finds the nearest local energy minimum to relieve steric clashes
• Equilibration: Brings the system to a stable temperature and pressure
• Production runs: Samples conformational dynamics over nanosecond to microsecond timescales
• Validates that a designed structure is thermodynamically stable, not just geometrically plausible
• Can reveal if a diffusion-designed backbone collapses or unfolds, providing a computational filter before expensive experimental validation