We’re excited to announce AIDO.StructureDiffusion, a generative AIDO module that combines precision structural control with high-quality generation of protein backbones. The model comes with the ability to design monomers and complexes, and can be fine-tuned for the generation of multi-chain binders and antibodies given a target.
AIDO.StructureDiffusion follows the RFDiffusion framework of training the head of a structure prediction model for structure generation, using a heavily modified version of the Boltz-1 diffusion module. It allows for arbitrary conditioning on desired structural classes directly on the residue level, enabling users to freely combine multiple motifs on a single or multiple chains. When de novo designing immunoglobulins, this unlocks the possibility to create precise CDR loop lengths, and also full, end-to-end generation of single-chain variable fragments (scFvs), producing structurally realistic multidomain immunoglobulin folds with correctly folded VH, linker, and VL regions. This enables rapid, automated creation of novel antibody candidates for therapeutics, diagnostics, and synthetic biology—without relying on predefined scaffolds or grafting.
The base model excels at full backbone generation, matching other state-of-the-art models in the fraction of designable and diverse structures resulting from each generation campaign.
With applications like binder design in mind, the base model was trained on a varied dataset of monomers and complexes, both from AFDB and RCSB. In order to ensure a varied training set, we assigned a weight to every chain and interface depending on the structural motifs found. This weighting scheme enables the model to generate monomers and complexes with diverse secondary structure content, and it can also be conditioned by dataset indicators, which were added during training.
Fine-Grained Structural Control by CAT Class Conditioning
AIDO.StructureDiffusion supports residue-wise structural conditioning, allowing precise structural control of the generated proteins. Using annotations from the CAT structural hierarchy, the generation process can be conditioned on particular CAT classes for separate domains on the same chain, allowing the generation of complex multi-domain proteins. This control can be coarse, using only the C level, but can also extend to the A and T levels. The model is notably able to produce discontinuous domains by providing non-contiguous stretches of the same structural conditioning values.
This approach to residue-wise and hierarchical structural conditioning of the diffusion process is naturally extendable to application specific conditioning, which we explore in antibody design below.
Antibody Design
Antibodies are a major component of modern approaches to drug discovery, making up more than a quarter of newly approved therapeutics in 2024. They are also one of the most promising modalities for AI-assisted protein design, where models that are able to reliably generate protein binders against disease relevant targets could potentially shorten drug development timelines and unlock new capabilities.
One major challenge in antibody engineering is developability. AI models have the potential to address key factors such as stability, aggregation, immunogenicity and expression by learning from the distribution of existing successful therapeutic antibodies. These models can be guided to design binders with more favorable developability profiles, increasing the likelihood of clinical success.
We fine-tuned the model on a mixture of antibody structures from SAbDab and interacting protein chains from the RCSB PDB. This targeted fine-tuning enabled the model to fully de novo generate backbones of nanobodies and variable fragments (Fv).
An area we targeted for improvement is the generic immunoglobulin CAT code, which encompasses a broad range of immunoglobulin superfamily proteins—including antibodies, T-cell receptors, MHC molecules, Fc receptors and others. As a result, using the unmodified CAT code limits the conditioning options specifically around antibody structures.
To address this, we expanded the CAT conditioning vocabulary to include fine-grained, immunoglobulin-specific features. This enhanced conditioning enables residue-level annotations for species, CDR and framework regions, and kappa (Κ) and lambda (λ) light chain loci. These annotations serve as supporting hierarchical enrichment of the immunoglobulin CAT code. With this additional context, the model can flexibly combine and control the lengths, shapes, and the number of structural elements in the generated antibodies—supporting more precise and diverse design.
Our conditioning framework lays the groundwork for incorporating developability-relevant features—shorter loop length, human of origin, Κ light chain—guiding the generation of antibodies that are not only structurally plausible but also more likely to succeed as therapeutics.
CDR Control
Current approaches typically rely on pre-selecting antibody scaffolds and then applying masked diffusion over fixed CDR loop lengths. Our method enables fully de novo antibody design by directly guiding the length of antibody regions. This strategy better aligns with the learned structural distribution of antibodies, allowing for a broader range of loop lengths without being constrained by predefined scaffolds. Importantly, this approach remains compatible with scaffold-based design if such constraints are desired.
Light Chain Locus Control
Guiding the model over the kappa (Κ) and lambda (λ) light chain loci offers the potential to influence epitope targeting preferences and improve developability profiles. In contrast to current diffusion-based approaches that often lock framework regions from existing therapeutics, limiting novelty and diversity, we demonstrate fully de novo diffusion conditioned on light chain locus, enabling greater flexibility and innovation in antibody structural space.
In natural antibody repertoires, the Κ locus is more frequently expressed than the λ locus—a pattern also reflected in our training data (see Figure above). To evaluate structure fidelity of our designs, we applied AbMPNN to create amino acid sequences for our generated backbones. In Κ- and λ- guided generations, the heavy and light chain sequences were correctly annotated in 100% and ~99% of cases, respectively, indicating strong consistency between structure and sequence.
Our model also demonstrated highly effective control over light chain loci. In Κ-guided generations, 99% of outputs were correctly assigned as the Κ chain, compared to just ~67% in unguided generations. Similarly, λ-guided designs resulted in ~72% λ-chain sequences, a substantial improvement over the ~24% observed without guidance.
These results highlight the model’s capacity to modulate over light chain loci with high precision—an essential capability for optimising therapeutic antibody function, improving biophysical properties, and expanding the diversity of candidates in drug discovery efforts.
For CDR and light chain locus identification, we employed AbMPNN, which may introduce its own biases and limitations. The flexible architecture of our model allows for future extensions, including all-atom structure generation, further broadening its applicability in structure-based antibody design.
Multi-Domain Single Chain Designs
Current antibody diffusion models typically generate heavy and light chains as separate polypeptides (Fv format). By integrating CAT and our immunoglobulin-specific guidance, our model is able to rationally design multi-domain constructs on a single chain. For example, we successfully generated single-chain variable fragments (scFvs) by prompting the model to produce immunoglobulin heavy and light variable domains connected by a 20-amino-acid linker. These scFvs exhibit correct chain loci, domain orientation towards the antigen, and accurately position all six CDRs in close proximity, forming a well-defined paratope.
When immunoglobulin guidance is omitted, the model no longer generates scFvs but instead defaults to producing VH-CH1 constructs—structures characteristic of Fab fragments. This behaviour reflects the model’s exposure to a training set heavily weighted towards Fab structures.
The ability to create single-chain, multi-domain designs in silico represents a major step forward in rational antibody design and beyond. In addition to conventional antibody formats, our guided generation framework opens up the possibility of designing more complex modalities, such as multispecific scFv and mini-proteins, immune cell engagers, all on a single polypeptide chain.
What’s next: Molecular Design in AIDO
Molecular design has entered a new era, with AI models being increasingly able to design binders to therapeutically relevant targets. Controllability of the design process is essential, and AIDO.StructureDiffusion enables new ways of guiding the model that users can leverage for generating structurally bespoke proteins. These additional capabilities are especially powerful in antibody design, where the extended structural vocabulary of AIDO.StructureDiffusion permits precise control of features relevant to developability, such as CDR length and light chain loci. It also enables the end-to-end generation of single-chain variable fragments.
While these contributions can already give direct benefit in real-world applications, the focus of ongoing and future work on AIDO.StructureDiffusion is to embed and steer the design process within the broader biological context of the emerging molecule. For this reason, AIDO.StructureDiffusion was developed as an AIDO module, allowing for integration with foundation models trained on other modalities, like protein and DNA sequences, and different scales, like cell expression. Within the AIDO, molecular design is not conceptualized as an isolated process, but an embedded component exchanging representations and feedback with its environment. We believe that by acknowledging biological complexity and making it a core modeling objective, this approach will unlock new capabilities to overcome key limitations of current drug discovery pipelines, like efficacy and toxicity.
