NVIDIA-BioNeMo/Proteina-Complexa

What it solves

Proteina-Complexa addresses the challenge of designing novel protein binders—molecules that bind to specific target proteins or small-molecule ligands—by unifying two previously separate approaches: conditional generative modeling and sequence optimization (hallucination). It aims to create high-affinity binders with precise atomistic detail, including backbone geometry and side-chain conformations, while improving in-silico success rates and computational efficiency.

How it works

The system uses a flow-matching generative model that jointly models protein backbone geometry, side-chain conformations, and sequences. It was pretrained on "Teddymer," a large-scale synthetic dataset of binder-target pairs derived from predicted protein structures, and further refined with experimental multimer data.

During inference, the model employs a search-based optimization strategy. It uses reward models—such as AlphaFold2, RoseTTAFold3, and various force fields—to guide the generation process and filter for the best candidates. The framework also integrates sequence design tools like ProteinMPNN and SoluableMPNN to ensure the resulting proteins are viable.

Who it’s for

This tool is designed for computational biologists, drug discovery researchers, and protein engineers who need to design custom protein binders for therapeutic targets, enzyme design, or motif scaffolding.

Highlights

• Unified Framework: Combines generative priors with test-time optimization to outperform traditional hallucination methods.
• Versatile Design: Supports protein-protein binders, protein-ligand binders, and motif scaffolding (AME).
• Atomistic Precision: Models backbone and side-chains jointly for higher accuracy.
• Integrated Validation: Built-in support for evaluation using AlphaFold2, ESMFold, and RoseTTAFold3.
• Experimental Validation: Designs have been wet-lab validated to demonstrate real-world binding activity.