Summary

Proteina-Complexa is a generative model for atomistic protein binder design, unifying conditional generative modeling and sequence optimization. It targets researchers and developers in drug discovery and protein engineering, offering state-of-the-art performance in designing novel protein and small molecule binders.

How It Works

This project extends flow-based latent protein generation architectures using flow matching, unifying generative and "hallucination" methods via inference-time optimization. It jointly models backbone geometry, side-chain conformations, and sequences. Pretraining utilizes the Teddymer dataset, a large-scale collection of synthetic binder-target pairs derived from predicted protein structures and experimental multimers.

Quick Start & Requirements

• Installation: Recommended: UV environment (./env/build_uv_env.sh, source .venv/bin/activate). Alternative: Docker (docker build, docker run --gpus all).
• Prerequisites: Ubuntu 22.04+ for UV (or Docker). GPU access required for Docker. Python 3.12 has a tmol installation workaround. Post-installation requires complexa init and complexa download --all. Environment variables for reward models (AF2, RF3) and tools (Foldseek, MMseqs2, DSSP, SC) must be configured in .env.
• Links: Paper, Project Page.

Highlighted Details

• Achieves state-of-the-art binder design performance with higher in-silico success rates.
• Test-time optimization strategies outperform prior hallucination methods under normalized compute.
• Supports protein binder, ligand binder, and motif scaffolding (AME) design for small molecules.
• Designs experimentally validated, confirming in-silico success translates to binding activity.

Maintenance & Community

Associated with NVIDIA and academic institutions, with core contributors listed. Links to the paper and project page are provided. No explicit community channels are mentioned.

Licensing & Compatibility

License details are in a LICENSE file. The specific license type and compatibility for commercial use are not detailed in the README.

Limitations & Caveats

• Known tmol installation issue on Python 3.12 requires a workaround.
• TMOL reward model is unsupported for protein-ligand complexes in ligand binder/AME pipelines.
• AME input PDBs require strict chain (A for ligand, B for motif) and residue naming (L:0 for ligand) conventions to prevent errors.