De Novo Protein Design

The foundational step of constructing a novel 3D protein scaffold without a natural template. Modern methods use diffusion models and SE(3)-equivariant neural networks to generate physically plausible backbone geometries. Key approaches include:

• Frame-based diffusion: Iteratively denoises residue frames (translation and rotation) to produce valid backbones
• Constrained hallucination: Optimizes random sequences against structure prediction networks to yield novel folds
• Fragment assembly: Combines short peptide fragments from the PDB into new topologies The generated backbone must satisfy Ramachandran plot constraints and avoid steric clashes before sequence design proceeds.

The computational task of designing an amino acid sequence predicted to fold into a specified target backbone. Tools like ProteinMPNN use message-passing neural networks to predict optimal residue identities given local and global structural context.

• Rotamer libraries: Discrete side-chain conformations are sampled to pack the hydrophobic core
• Energy scoring: Rosetta energy functions or learned potentials evaluate sequence-structure compatibility
• Diversity constraints: Sequences are optimized for solubility and to avoid aggregation-prone motifs This step effectively solves the inverse protein folding problem, generating sequences that stabilize the designed backbone.

Proteins are designed to perform a specific biochemical function from scratch. This requires precise geometric specification of active sites, binding pockets, or catalytic residues before backbone generation.

• Binding site scaffolding: Matching a target ligand's geometry with complementary side-chain placement
• Theozyme construction: Placing catalytic residues in optimal geometries to stabilize transition states
• Parametric design: Using coiled-coil or repeat protein parameters to build modular functional scaffolds
• De novo enzymes: Creating active sites with non-natural catalytic mechanisms not found in biology Functional validation requires experimental characterization of binding affinity (KD) or catalytic efficiency (kcat/KM).

A technique where a structure prediction network like AlphaFold2 or RoseTTAFold is used in reverse to generate novel proteins. By optimizing a random input sequence to produce a confident predicted structure, the network 'hallucinates' new folds.

• Loss optimization: Maximizing pLDDT and minimizing PAE drives the network toward novel, well-folded structures
• Diversity sampling: Different random seeds produce structurally diverse outputs
• Inpainting: Specifying fixed functional motifs while hallucinating the surrounding scaffold This method leverages the implicit knowledge of folding landscapes learned by prediction models without requiring explicit energy functions.

Computational designs must be experimentally validated and refined. The design-build-test cycle integrates wet-lab characterization with computational feedback.

• Recombinant expression: Designs are synthesized as genes, expressed in E. coli or cell-free systems
• Biophysical characterization: Circular dichroism confirms secondary structure; SEC-MALS verifies monodispersity
• High-resolution structure determination: X-ray crystallography or cryo-EM confirms atomic-level accuracy
• Deep mutational scanning: High-throughput functional assays provide fitness landscapes for iterative redesign Failed designs inform improved energy functions and sampling strategies, closing the loop between computation and experiment.

Advanced methods that impose specific geometric or topological constraints during generation. These approaches enable the creation of proteins with predetermined shapes and symmetries.

• Symmetry-constrained design: Generating cyclic, dihedral, or icosahedral protein assemblies
• Topological specification: Designing proteins with predetermined knot types or chain entanglement patterns
• Parametric backbone curves: Defining the overall protein shape using mathematical splines before sequence design
• Pore and channel design: Creating transmembrane barrels with specified conductance properties These techniques are critical for designing protein nanomaterials, vaccine nanoparticles, and synthetic membrane proteins.

The computational task of designing an amino acid sequence that will fold into a specified target backbone structure. Unlike structure prediction (sequence → structure), inverse folding solves the reverse problem (structure → sequence).

• ProteinMPNN is a leading deep learning tool for this task
• Uses message-passing neural networks on protein graphs
• Outputs sequences with high recovery rates in experimental validation
• Critical for designing binders, enzymes, and self-assembling nanomaterials

A class of generative models that create novel protein structures by iteratively denoising random 3D coordinates. Inspired by image generation techniques like DALL-E, these models learn the distribution of valid protein backbones.

• FrameDiff and RFdiffusion are prominent implementations
• Operate on SE(3) equivariant representations of residue frames
• Generate structures conditioned on target motifs or binding pockets
• Enable unconditional generation of diverse, designable backbones

Large-scale neural networks trained on vast databases of protein sequences using self-supervised learning. Models like ESM-2 capture evolutionary, structural, and functional information in their internal representations.

• Trained via masked language modeling on millions of sequences
• Internal attention maps correlate with contact prediction
• ESMFold leverages pLM representations for structure prediction
• Used to score and filter designed sequences for foldability

A mathematical property ensuring a neural network's predictions transform consistently with the rotation and translation of input 3D coordinates. Essential for protein design where structures exist in continuous 3D space.

• Guarantees predictions are independent of the protein's orientation in space
• Implemented via tensor field networks and irreducible representations
• Core to architectures like SE(3)-Transformers and EquiDock
• Contrasts with invariance, which would lose directional information

A physics-based and knowledge-based scoring function that estimates the thermodynamic stability of a protein conformation. Used extensively to evaluate and refine computationally designed sequences.

• Combines van der Waals, electrostatics, solvation, and hydrogen bonding terms
• ref2015 is the current standard parameterization
• Guides fixed-backbone design and flexible backbone relaxation
• Complements deep learning methods by providing physical validation

The critical wet-lab confirmation that computationally designed proteins fold and function as intended. Bridges the gap between in silico prediction and real-world utility.

• Circular Dichroism (CD) confirms secondary structure and thermal stability
• X-ray crystallography or Cryo-EM provides atomic-resolution structures
• Binding assays (SPR, BLI) measure target affinity for designed binders
• High-throughput yeast display screens thousands of designs simultaneously