De Novo Protein Design

The foundational generative paradigm that inverts the structure-to-sequence relationship. Instead of predicting how a sequence folds, inverse folding models take a target 3D backbone structure as input and generate an amino acid sequence predicted to fold into that exact conformation. Architectures like ProteinMPNN use message-passing neural networks on the protein graph to autoregressively decode a sequence, achieving remarkable recovery rates for native sequences. This approach is widely used for scaffold conditioning, where a designer specifies a desired structural motif and the model hallucinates a supporting sequence.

A generative framework adapted from image synthesis that learns to denoise protein structures. Frame-based diffusion models like RFdiffusion operate on the SE(3) group of rigid-body motions, progressively transforming random noise into valid protein backbones. Key mechanisms include:

• Forward process: Gradually corrupts backbone coordinates with Gaussian noise
• Reverse process: A denoising network trained on the Protein Data Bank learns to recover realistic structures
• Self-conditioning: The model's own previous prediction is fed back as input, dramatically improving designability This approach excels at unconditional generation and motif-scaffolding tasks.

A gradient-based optimization technique that uses protein language models (pLMs) as a learned prior over sequence space. Starting from a random sequence, the method iteratively mutates residues to maximize the model's confidence under a masked language modeling objective, while applying structural constraints. The trRosetta hallucination protocol couples a pLM with a structure prediction network, jointly optimizing for high-confidence folding predictions. This approach requires no training on structural data and can generate sequences that fold into novel topologies not seen in nature.

Methods for imposing functional requirements during the generation process to create proteins with specific biochemical properties. Common constraints include:

• Motif scaffolding: Fixing the backbone coordinates of a catalytic site or binding interface while generating the surrounding structure
• Symmetry conditioning: Enforcing cyclic, dihedral, or icosahedral symmetry for self-assembling nanoparticles
• Sequence constraints: Specifying conserved residues or charge distributions
• Binding target conditioning: Generating sequences predicted to bind a specific target protein using auxiliary potential functions Tools like LigandMPNN extend inverse folding to explicitly model small molecule and cofactor interactions.

An emerging paradigm that generates sequence and structure jointly rather than sequentially. EvoDiff combines discrete diffusion over amino acid identities with continuous diffusion over backbone coordinates in a unified framework. ProteinGenerator uses a flow-matching approach to simultaneously sample sequence-structure pairs from a learned distribution. Co-design avoids the information bottleneck of conditioning on a fixed backbone and can explore a broader functional landscape, discovering sequence-structure combinations that inverse folding alone might miss.

The critical wet-lab feedback loop that closes the design-build-test cycle. Generated sequences are filtered by designability metrics (self-consistency TM-score, pLDDT), then synthesized and characterized:

• High-throughput screening: Cell-free expression and fluorescence-based folding assays
• Biophysical characterization: Circular dichroism, size-exclusion chromatography, and thermal shift assays to confirm monodispersity and stability
• Structure determination: X-ray crystallography or cryo-EM to validate the designed fold at atomic resolution Leading groups report >50% success rates for designs that are soluble, monodisperse, and match the target structure.

The computational task of predicting an amino acid sequence that will fold into a specified target backbone structure. Unlike forward folding (structure prediction), inverse folding solves the design problem by mapping from 3D coordinates to sequence space. Models like ProteinMPNN use message-passing neural networks on protein graphs to generate sequences with high structural fidelity and expression yield. This approach is foundational for designing novel enzymes and binding proteins where the desired fold is known.

A generative framework that learns to reverse a gradual noising process to create novel protein structures. Starting from random coordinates, the model iteratively denoises atomic positions to produce physically plausible backbones. Architectures like RFdiffusion and FrameDiff operate on SE(3) frames, generating diverse and designable structures without a template. Key advantages include unconditional generation of novel folds and conditional generation guided by target motifs or binding pockets.

A method that optimizes random amino acid sequences to produce high-confidence structural predictions from models like AlphaFold or RoseTTAFold. By maximizing the predicted confidence metrics (pLDDT) while minimizing similarity to known structures, the model 'hallucinates' novel proteins. This approach was used to generate the first computationally designed proteins with no natural homologs, validating the ability to explore uncharted fold space.

Neural networks that respect the 3D symmetries of protein structures, ensuring predictions are invariant to rotation and translation. Architectures like SE(3)-Transformers and GVP-GNNs process vector features (atomic coordinates) alongside scalar features (amino acid types). This geometric deep learning approach is essential for both structure prediction and de novo design, as it guarantees physically consistent outputs regardless of the protein's orientation in space.

A design paradigm where a generative model builds a full protein structure around a fixed functional motif, such as an enzyme active site or a binding interface. The model must solve a constrained optimization problem: scaffolding the motif into a stable, foldable protein while preserving its precise geometry. RFdiffusion excels at this task, enabling the design of novel catalysts and therapeutic binders by grafting known functional sites into entirely new structural contexts.