IgFold

IgFold employs an Invariant Point Attention (IPA) mechanism adapted from AlphaFold2, which operates directly on 3D coordinates while maintaining SE(3) equivariance. This means the predicted structure is invariant to global rotation and translation—a critical property for physically meaningful protein representations. The IPA module iteratively refines backbone frames and residue orientations through attention operations that respect the geometric relationships between amino acids.

IgFold leverages AntiBERTy, a BERT-style transformer language model pre-trained on 558 million antibody sequences from the Observed Antibody Space (OAS) database. These deep contextual embeddings encode:

• Germline gene usage and allelic variation
• Somatic hypermutation patterns
• Paired chain co-evolution signals
• Structural propensity of individual residues This antibody-specific pre-training provides a powerful prior that dramatically improves structure prediction accuracy compared to general protein language models.

The broader field encompassing de novo computational generation of an antibody's three-dimensional structure from its amino acid sequence. The central challenge is accurately modeling the hypervariable CDR-H3 loop, which exhibits extreme sequence and conformational diversity. IgFold represents a state-of-the-art solution within this domain, prioritizing speed over physics-based simulation. Traditional methods like RosettaAntibody use homology modeling and energy minimization, while newer approaches leverage geometric deep learning.

The six hypervariable loops (L1, L2, L3, H1, H2, H3) forming the antigen-binding site. CDR-H3 is the most diverse in length, sequence, and structure, making it the primary determinant of binding specificity and the most difficult region to model. IgFold's architecture is specifically designed to handle the complexity of CDR-H3 prediction without relying on database lookups of canonical loop conformations, a limitation of older template-based methods.

Transformer-based neural networks pre-trained on vast repositories of antibody sequences to learn the underlying grammar of immune receptors. Models like AntiBERTy and AbLang generate dense vector representations capturing evolutionary and functional constraints. IgFold leverages these representations as input features, using the learned sequence embeddings to infer structural contacts and inter-residue geometries without requiring explicit multiple sequence alignments.

A class of neural networks that respect the symmetries of 3D space, ensuring predictions rotate and translate consistently with the input. SE(3)-equivariant architectures are crucial for protein structure prediction because they guarantee that the predicted coordinates transform correctly under rigid-body motion. IgFold's internal layers utilize these principles to predict residue-level frames and torsion angles directly from sequence embeddings, enforcing physical plausibility.

A downstream application that predicts the three-dimensional binding pose of an antibody relative to its target antigen. Rapid structure prediction by IgFold provides the necessary input for docking algorithms like HDOCK or ClusPro. The speed of IgFold enables high-throughput docking screens where thousands of predicted antibody structures are computationally bound to an antigen to prioritize candidates for experimental validation.

A computational screening cascade that aggregates predictions for aggregation propensity, thermal stability, and chemical liabilities. Accurate 3D structures from IgFold enable structure-based developability metrics, such as calculating spatial aggregation propensity (SAP) or identifying solvent-exposed hydrophobic patches. This allows teams to filter candidates early in discovery, before investing in synthesis and wet-lab characterization.