Pythia’s technology is designed to rapidly generate functional, full-length antibodies with the properties required for therapeutic development.
Our flow-based generative models are trained on large-scale experimental and synthetic data. They reason jointly over amino acid sequence, three-dimensional atomic structure, and explicit design specifications and constraints. Bringing these representations together in a single generative framework allows the models to translate high-level design objectives into fully specified molecules with atomic precision.
Many therapeutic target classes, including multi-pass membrane proteins, are exceptionally difficult to drug. Conformational dynamics, limited extracellular surface area, and poorly immunogenic or difficult-to-access epitopes often make conventional approaches such as immunization and library screening ineffective.
Pythia’s models are not confined to the molecular space accessible to conventional discovery. By exploring a broader design space, our technology can generate antibodies against targets that have historically been out of reach.
In a recent study, we demonstrated that our technology can generate full-length, high-affinity, functional antibodies against GPCRs, including both agonists and antagonists, without relying on iterative optimization or large experimental screens. The resulting antibodies also exhibited developability and manufacturability characteristics expected of therapeutic lead candidates.
Our technology provides control over where an antibody binds and the conformational state of the antigen it recognizes. This allows the intended mechanism of action to shape the design from the outset, so candidates can be created for a defined functional outcome, including agonist or antagonist activity.
Our models generate full-length IgGs with human-like CDRs. Germline frameworks can be specified, and the models can be constrained to avoid sequence liabilities. These capabilities bring key developability and manufacturability considerations into the initial design, reducing the need for extensive downstream engineering.
Our models can generate high-affinity, functional antibodies from scratch in a single design round. In our recent study, testing only a handful of designs per target yielded candidates with key properties expected of therapeutic leads, without iterative optimization or large experimental screens.
