Diffusion model for generative protein design
Inventors
Ingraham, John • Ismail, Ahmed • Obermeyer, Fritz Heinrich • Baranov, Maxim • Wang, Wujie • Costello, Zachary Kohl • GRIGORYAN, Gevorg
Assignees
Interested in licensing this patent?
MTEC can help explore whether this patent might be available for licensing for your application.
Publication Number
US-12580040-B2
Publication Date
2026-03-17
Expiration Date
Abstract
A system is disclosed for de novo protein generation. The system receives a set of design condition(s) that specify target characteristics of a synthetic protein. The system defines a modular energy function as a composition of a diffusion energy component and one or more conditioner energy components. The system applies a diffusion model to determine a denoised protein backbone. In applying the diffusion model, in each sampling step: the system transforms one prior sampled state of the synthetic protein from unconstrained space into constrained space based on the one or more design conditions, denoises the prior sampled state in the constrained space, and samples a subsequent sampled stated by applying a gradient of the modular energy function to the denoised prior sampled state in the constrained space. The final sampled state is a denoised protein backbone for the synthetic protein that satisfies the set of design condition(s).
Core Innovation
The invention generates a protein or a protein complex by using a trained diffusion model to transform an initial state representing a protein backbone into a final state representing a denoised protein backbone. The initial state specifies three-dimensional coordinates of heavy atoms in amino acid residues of the protein backbone, and the transforming is performed by sampling using a reverse-time stochastic differential equation, a Langevin dynamics stochastic differential equation, or a hybrid combining both.
The trained diffusion model comprises a graph neural network with nodes for amino acid residues of the protein backbone and has a sparse edge structure. Sampling is performed using a stochastic differential equation with a structured covariance enforcing protein chain and radius of gyration statistics, and the denoised protein backbone is input to a trained sequence generation model to generate an amino acid sequence.
The invention further supports sampling conditioned by one or more protein property conditions using energy components, including constraints and restraints such as a domain classifier constraint, secondary structure constraint, distance constraint, substructure RMSD constraint, substructure infilling restraint, shape constraint, symmetry constraint, and text caption restraint. It also describes symmetry-conditioned sampling for protein complexes using a symmetry group G, and shape conditioning defined from couplings via Optimal Transport, including Wasserstein distance and Gromov-Wasserstein distance.
Claims Coverage
The independent claims are clm-00001, clm-00006, and clm-00010. Across these claims, the main inventive features center on denoising a protein backbone via diffusion sampling with structured covariance, using a GNN-based diffusion model together with SDE/Langevin or hybrid sampling, and generating an amino-acid sequence with a trained sequence generation model followed by manufacturing, with optional property-conditioning via energy components.
Structured diffusion sampling for denoised protein backbone
Transforming the initial state representing the protein backbone through a series of states to a final state representing a denoised protein backbone, the transforming performed by sampling using the trained diffusion model, wherein the sampling is performed using a reverse-time stochastic differential equation (SDE), a Langevin dynamics SDE, or a hybrid SDE combining both the reverse-time SDE and the Langevin dynamics SDE; and wherein the sampling is performed using a stochastic differential equation with a structured covariance enforcing protein chain and radius of gyration statistics.
GNN-based diffusion model with sparse edge structure
The trained diffusion model comprises a graph neural network (GNN) that comprises nodes for amino acid residues of the protein backbone and has a sparse edge structure.
Sequence generation from denoised backbone and manufacturing
Applying a trained sequence generation model to the denoised protein backbone to generate an amino acid sequence for the protein; and manufacturing the protein having the amino acid sequence.
Identical first and second GNN in sequence generation model
The trained diffusion model comprises a first graph neural network (GNN) that comprises nodes for the amino acid residues of the protein backbone; and the trained sequence generation model comprises a second GNN, wherein the first GNN and the second GNN are the same.
Protein property conditioning via multiple energy components for constraints and restraints
Applying a trained diffusion model wherein the sampling is performed using a diffusion energy component and one or more energy components for protein property conditions, with the protein property conditions defined by one or more selected constraints/restraints including a domain classifier constraint, secondary structure constraint, distance constraint, substructure root mean squared deviation (RMSD) constraint, substructure infilling restraint, shape constraint, symmetry constraint, and text caption restraint; and specifying an energy component for each constraint/restraint.
Across clm-00001, clm-00006, and clm-00010, the claims cover methods that denoise a 3D protein backbone via diffusion sampling using reverse-time SDE, Langevin dynamics SDE, or hybrid SDE, with a stochastic differential equation using a structured covariance enforcing protein chain and radius of gyration statistics. The diffusion model uses a GNN, the denoised backbone is converted to an amino-acid sequence via a trained sequence generation model, and the resulting protein is manufactured. Additional dependent refinements describe conditioning the sampling using energy components tied to explicit protein property constraints and restraints.
Stated Advantages
Improves likelihood and sample quality under ablation findings that indicate globular covariance and certain auxiliary/denoising losses improve likelihood and sample quality.
Uses refolding validation with agreement/confidence correlations (TM-score vs pLDDT) to assess refolding success trends.
Supports programmability for composable conditioner modules supporting both constraints and restraints via transformed coordinates/energy with Jacobian log-determinant correction.
Documented Applications
De-novo binders using constraint composition examples including symmetry, substructure constraints, distances, motifs, and shape constraints.
Enzyme miniaturization using programmable constraint/restraint conditioning examples.
Nanostructure control using programmable constraint/restraint conditioning examples for symmetry and shape.
Language-guided design using text caption conditioning.
Interested in licensing this patent?