AbBFN2

AbBFN2 is a generative AI model developed by InstaDeep for flexible, multi-objective antibody design and optimisation. Built on the Bayesian Flow Network (BFN) paradigm and extending the ProtBFN architecture into a multimodal framework, AbBFN2 is designed for pharmaceutical and biotech researchers seeking to consolidate complex antibody engineering tasks into a single unified generative framework. It is accessible via the DeepChain platform with minimal setup required.

Therapeutic antibody development is a resource-intensive, multi-objective optimisation challenge. The estimated germline antibody sequence space spans 10 to 100 billion possible sequences, and effective therapeutic candidates must simultaneously achieve precise target binding, low immunogenicity, stability, appropriate expression levels, and freedom from aggregation propensity. Traditional pipelines rely on separate, case-specific tools for each step—such as humanisation, developability assessment, and sequence generation—meaning optimising one property often compromises another. AbBFN2 addresses this by jointly modelling sequence, genetic, and biophysical attributes within a single framework, enabling scientists to optimise multiple properties simultaneously.

Core Architecture and Capabilities

• Built on the Bayesian Flow Network (BFN) paradigm, extending ProtBFN into a multimodal antibody-specific framework.
• Trained on diverse antibody sequences, modelling 45 different modalities including sequence, genetic, and biophysical attributes.
• Adapts to user-defined tasks by conditionally generating any subset of attributes when given values for others, enabling a steerable and flexible design approach.
• Does not require retraining or fine-tuning to accommodate new design tasks, greatly lowering the barrier to targeted antibody library generation.
• Combines interactive capabilities with domain expertise to enhance development campaigns and accelerate antibody discovery.

Unconditional Generation

• Generates biologically plausible antibody sequences without task-specific prompting or conditioning.
• Generated sequences closely match natural antibody distributions across CDR loop lengths, amino acid propensities, and sequence diversity.
• When folded using ImmuneBuilder, generated sequences align with structural motifs observed in natural antibodies, despite not being explicitly trained on structural information.
• Faithfully captures conditional distributions involving gene pairing propensities within and between variable heavy (VH) and variable light (VL) chain domains.
• Confirmed to generate sequences distinct from training data, demonstrating no memorisation and genuine generalisation.

Sequence Annotation

• Infers key attributes from sequence alone, including species, gene usage, and developability metrics.
• Achieves state-of-the-art performance across all evaluated annotation tasks, consistently matching or outperforming existing tools.
• Accurately predicts Therapeutic Antibody Profiler (TAP) flags—development liabilities—directly from sequence.
• Demonstrates implicit reasoning about structural effects despite not being explicitly trained on structural data.
• Accurately identifies heavy and light chain V, (D), and J genes, reflecting a strong understanding of antibody recombination patterns and lineage information.

Sequence Inpainting

• Completes antibody sequences when only partial sequence information is available, useful for redesigning specific regions such as CDR loops or the VH–VL interface.
• Successfully reconstructs natural VH–VL pairings, demonstrating understanding of intricate inter-chain dependencies.
• Conditioning on metadata such as framework sequence and gene segments significantly improves CDR prediction accuracy.
• Enables intelligent completion and editing of real-world antibody sequences while maintaining structural and functional integrity.

Sequence Humanisation

• Learns the likelihood that a given antibody will elicit an adverse immune reaction upon administration.
• Assessed humanness of 211 clinical-stage antibodies, finding that sequences rated as more human correlated with fewer observed immunogenic responses.
• Humanness score validated as a reliable signal for optimisation, correlating with real-world immune outcomes; over 99% of antibodies confidently classified as human present low to medium immunogenicity risks.
• Applied to 25 therapeutic antibodies with known experimentally humanised variants, accurately selecting human-compatible variants that closely mirrored experimentally derived sequences in both number and location of mutations.

Multi-Objective Optimisation

• Supports multi-round, multi-objective optimisation focusing on simultaneous humanisation and developability improvement.
• Often reaches a high probability of human-likeness within a single recycling iteration.
• Applied to 91 non-human sequences, achieving both human-likeness and developability objectives for 63 sequences within just 2.5 hours—a process that would traditionally take weeks to months per sequence experimentally.

Conditional Library Generation

• Generates antibody libraries enriched for rare and highly specific characteristics without additional model training.
• Conditioning parameters can include partial sequence context, target HV gene, defined CDR-L3 loop length, specific light chain locus, and favourable developability attributes.
• In benchmark testing, produced 1,715 out of 2,500 sequences meeting all specified requirements simultaneously, with sequences exhibiting natural-like behaviour across additional properties.
• Applied to broadly neutralising antibodies against HIV, achieving a more than 56,000-fold higher likelihood of satisfying all requested attributes compared to unconditional sampling by a standard protein language model or AbBFN2 itself.

AbBFN2 is available to explore via the DeepChain platform, with options to try a live demo, download the accompanying research paper, or access the open-source code. It is part of InstaDeep's broader suite of life sciences AI models on DeepChain, which also includes ProtBFN, Protein Folding, InstaNovo, and Nucleotide Transformers.