Generative Antibody Design

An adaptation of the RoseTTAFold Diffusion model for protein backbone generation, fine-tuned specifically for antibody design. It generates novel antibody structures conditioned on target antigen structures.

• Generates diverse CDR loop conformations not seen in nature
• Conditions generation on 3D epitope coordinates
• Validated experimentally for binding to influenza hemagglutinin and other targets
• Integrates with ProteinMPNN for inverse folding to design sequences for generated backbones

A denoising diffusion probabilistic model that jointly generates antibody CDR sequences and structures conditioned on the 3D structure of a target antigen epitope.

• Operates on residue-level representations with SE(3) equivariance
• Simultaneously designs sequence and structure for all six CDR loops
• Demonstrates improved binding affinity over native antibodies in silico
• Uses an equivariant graph neural network as the denoising backbone

A GPT-style autoregressive transformer trained on 558 million antibody sequences from the Observed Antibody Space (OAS) database. It generates full-length antibody sequences with controllable properties.

• Generates antibodies conditioned on species-of-origin and chain type
• Performs sequence infilling: redesigns specific CDR regions while preserving framework residues
• Captures germline lineage patterns in generated repertoires
• Enables zero-shot humanization by conditioning on human framework tokens

A Bidirectional Encoder Representations from Transformers model pre-trained on 558 million antibody sequences, designed for structure-aware antibody analysis rather than generation.

• Learns structural cluster representations corresponding to canonical CDR loop conformations
• Embeddings capture paratope-epitope interaction patterns without structural supervision
• Used for clustering antibody repertoires and identifying binding similarity
• Provides foundational embeddings for downstream generative fine-tuning tasks

An SE(3)-equivariant graph neural network for co-designing antibody sequences and 3D structures. It models the full antibody-antigen complex with geometric attention mechanisms.

• Processes multi-channel representations including residue types, coordinates, and pairwise distances
• Generates antibodies with higher predicted binding affinity than sequence-only methods
• Uses message passing on k-nearest neighbor graphs of the antigen surface
• Outputs both backbone coordinates and amino acid identities simultaneously

A masked language model trained on both heavy and light chain antibody sequences, designed to restore missing residues in incomplete antibody sequencing data and generate novel variants.

• Handles paired heavy and light chain representations simultaneously
• Restores missing germline residues from alignment gaps in repertoire sequencing
• Generates diverse CDR3 variants through iterative masked prediction
• Lightweight architecture suitable for high-throughput screening of generated libraries

A transformer-based neural network pre-trained on vast repositories of antibody sequences to learn the underlying grammar of immune receptors. These models capture evolutionary constraints, positional dependencies, and mutational tolerances.

• Trained on hundreds of millions of natural antibody sequences
• Used for variant effect prediction, sequence likelihood scoring, and generative design
• Examples include AntiBERTy, IgLM, and AbLang
• Enable zero-shot prediction of binding properties without structural data

The de novo computational generation of an antibody's three-dimensional structure from its amino acid sequence, with specific focus on accurately modeling the hypervariable CDR-H3 loop.

• CDR-H3 is the most diverse and functionally critical region
• Specialized tools like IgFold achieve sub-second prediction times
• Deep learning methods now rival experimentally determined structures
• Accurate structure prediction is a prerequisite for physics-based affinity optimization

A physics-based or deep learning simulation that predicts the three-dimensional binding pose and orientation of an antibody relative to its target antigen. This is the computational equivalent of determining how a key fits into a lock.

• Traditional methods use Fast Fourier Transform-based rigid-body docking
• Deep learning approaches like AlphaFold-Multimer predict complexes directly
• Critical for identifying paratope-epitope contact residues
• Scoring functions rank decoy poses to identify near-native binding modes

A simultaneous generative process that designs the antibody sequence and its corresponding 3D structure in a coupled fashion. This ensures the generated sequence is physically plausible and foldable.

• Diffusion models generate backbone coordinates and sequences jointly
• Addresses the sequence-structure gap where designed sequences fail to fold
• Enables design of novel topologies beyond natural antibody frameworks
• Frameworks like RFdiffusion and ProteinMPNN are leading this approach