Vilya-1: An all-atom foundation model for macrocycle structure prediction and design
Moving beyond canonical chemistry
While many methods exist to predict protein structure from natural sequence, these models struggle to accurately model peptides with non-canonical amino acids.
The design of therapeutics requires consideration of a suite of properties beyond just binding, and canonical amino acids often do not provide enough diversity to satisfy all of the required constraints for a drug discovery campaign. In order to deliver on challenging, intra-cellular targets, we need to move beyond the capabilities of existing design tools towards a previously unexplored chemical space.
Vilya-1 is a neural network built from first principles
Specifically tailored to model the structure of macrocycles with diverse chemistries.
It achieves state-of-the-art performance at recapturing a broad range of macrocycle structures—more than doubling the performance of the next best method. Its structural capabilities provide a foundation for many important problems in design, including enriching for permeable compounds and screening for protein binders using energy landscapes.
State-of-the-art structural modeling
On a combined benchmark of publicly-available and internal macrocycle crystal structures, Vilya-1 greatly outperforms all other methods at recovering crystal conformations. Physics- and rule-based methods successfully recover less than half of the conformations that Vilya-1 does, and deep-learning-based methods lag even farther behind, likely due to their inability to model chemistries different from their training sets.
From structure to function: transfer learning to downstream properties
Vilya-1 doesn’t just model structures: it can be fine-tuned to enrich for permeable compounds, and outperforms other architectures on the same task. In addition, it can rank its own sampled structures and enrich greatly for favorable conformations—a problem previously tackled only in the context of protein structure prediction, where most inputs adopt a single, globular fold.
Explore energy landscapes to prioritize the best designs
Designing macrocycles isn’t just about building the right structure. Vilya-1 lets us generate detailed conformational ensembles for a given molecule and evaluate the energy landscape around a design hypothesis. By analyzing these landscapes, we can identify which sequences are most likely to adopt the intended, binding-competent conformation.
Enabling design of arbitrary topologies
Accurately modeling structures is the foundation of design. We can use Vilya-1 to tackle myriad challenges in drug discovery—from screening potential binders to miniaturizing known hits as part of an optimization campaign. The molecules we create aren’t limited to canonical chemistries—and neither is our platform.
See the full technical report
Moving beyond the capabilities of existing design tools towards a previously unexplored chemical space.
Interested?
We are actively using Vilya-1 in multiple internal programs. Want access? Contact us!
Interested in joining us to push the boundary of AI-based design of macrocyclic peptides? Explore the open positions!
Next