Antibody Structure Prediction

A physics-based or deep learning simulation that predicts the three-dimensional binding pose of an antibody relative to its target antigen. Rigid-body docking treats both partners as fixed, while flexible docking samples side-chain and backbone movements. Key challenges include accurately modeling the CDR-H3 loop flexibility and accounting for interface water molecules that mediate hydrogen bonds. Modern approaches like AlphaFold-Multimer and ClusPro integrate evolutionary information to improve success rates. The output is a ranked list of decoys scored by shape complementarity, electrostatic interactions, and desolvation energy.

A specialized deep learning model for rapid, template-free prediction of antibody variable domain structures directly from sequence. Unlike general protein folding models, IgFold is trained specifically on antibody structural data from the PDB and SAbDab databases. It uses a graph neural network architecture to predict inter-residue distances and orientations, then reconstructs the 3D coordinates. IgFold can predict an Fv structure in under 30 seconds, making it suitable for high-throughput screening of large immune repertoire datasets. It shows particular strength in modeling the framework regions while CDR loops, especially H3, remain the primary source of prediction error.

A transformer-based neural network pre-trained on vast repositories of antibody sequences to learn the underlying grammar of immune receptors. Models like AntiBERTy, AbLang, and ESM-IF use masked language modeling objectives to capture evolutionary constraints and structural propensities. These models generate dense vector representations (embeddings) that encode biophysical properties useful for downstream tasks:

• Variant effect prediction: scoring mutations for affinity impact
• Sequence generation: sampling novel CDR loops
• Anomaly detection: identifying unusual sequences with developability risks
• Paratope prediction: identifying antigen-contact residues from sequence alone

The application of generative models to create entirely novel antibody sequences and structures with desired binding properties. Diffusion models (e.g., RFdiffusion, AbDiffuser) learn to denoise random coordinates into valid protein backbones conditioned on target epitopes. Autoregressive models sample amino acids sequentially while maintaining structural consistency. Key capabilities include:

• Inpainting: redesigning only the CDR loops while preserving framework integrity
• Scaffolding: generating binders around specified motif residues
• Unconditional generation: exploring novel folds beyond natural repertoires This approach moves beyond mining natural immune repertoires toward true de novo design of therapeutic candidates.

A comprehensive computational screening cascade that aggregates predictions for multiple biophysical liabilities to rank antibody candidates for manufacturing suitability. Key assays modeled in silico include:

• Aggregation propensity: spatial aggregation propensity (SAP) scores and hydrophobic patch analysis
• Thermal stability: melting temperature (Tm) prediction from sequence and structure
• Chemical liabilities: deamidation (NG, NS motifs), isomerization (DG, DS), oxidation (M, W), and fragmentation sites
• Solubility: charge distribution and isoelectric point (pI) calculations
• Viscosity: intermolecular interaction modeling at high concentration These predictions filter out problematic candidates before costly experimental characterization.

A high-throughput experimental technique that measures the functional effect of thousands of single amino acid substitutions in an antibody. A variant library is created via saturation mutagenesis, subjected to a functional selection (e.g., binding, expression, stability), and sequenced before and after selection to quantify enrichment ratios for each variant. The resulting datasets provide:

• Fitness landscapes: comprehensive maps of mutational tolerance at each position
• Training data: supervised learning for variant effect predictors
• Validation: ground truth for benchmarking structure prediction accuracy
• Engineering insights: identification of affinity-enhancing mutations and destabilizing hotspots