Use Genie 3 Online
Commercially Available Genie 3 No-Code Web Server
Genie 3: Fast All-Atom SE(3)-Equivariant Diffusion for Protein Design, Binders and Motif Scaffolding
Genie 3 is an all-atom SE(3)-equivariant diffusion model for protein design, described in a 2026 bioRxiv preprint by Yeqing Lin, Minji Lee, Aakarsh Vermani, Ellena Jiang, Sebastiaan De Cooman, Matej Špeťko and Mohammed AlQuraishi. The code is released under the Apache-2.0 license and supports unconditional monomer generation, motif scaffolding and binder design.
Genie 3 treats proteins as branched polymers so that it can reason at atomic detail, including sidechains, while keeping SE(3)-equivariance.
The authors argue that protein design has lagged in two areas: modeling sidechain detail, and aligning generative design with the in silico oracle used to pick designs. Recent all-atom models give up SE(3)-equivariance, while hallucination methods are accurate but slow.
Genie 3 aims to close that gap with a diffusion approach that is both atomistic and fast, and the paper reports state-of-the-art results on binder design, motif scaffolding and unconditional generation.
Item | Detail |
|---|---|
Model type | All-atom SE(3)-equivariant diffusion for protein design |
Applications | Unconditional monomers, motif scaffolding, binder design |
Sampling | 100 steps instead of 1,000, about 10x faster with little loss in quality |
Training data | Confident AlphaFold Database monomers and Pinder dimers |
Venue | bioRxiv preprint, 2026 |
License | Apache-2.0 (code) |
How Genie 3 Designs Proteins at Atomic Detail
Representation: forward diffusion noises atom clouds; the reverse process uses Frenet-Serret frames along the backbone and sidechains.
Denoiser: an SE(3)-equivariant network with an input embedder, a latent transformer and an Invariant Point Attention decoder.
Partial atomization: motif residues are all-atom while scaffold residues are backbone-only; sidechains are sometimes masked in training so the model learns to reconstruct them.
Interface heuristics: extended hotspots expand the target interface, and convergent hotspots reuse residues shared by successful designs across rounds.
Directional scaling: lowers the sampling temperature for higher-quality designs.
Genie 3 Benchmark Results for Binder Design and Motif Scaffolding
Benchmark | Result |
|---|---|
Unconditional generation | Matches or beats Ambient and La-Proteina on monomers under 300 residues and outperforms them at 300-800 residues |
Motif scaffolding (MotifBench, 30 challenges) | Solves as many problems as Protpardelle-1c and achieves a higher MotifBench score |
Binder design (10 problems), 200-structure budget | Most successful designs on 7 of 10; BindCraft wins the other 3 |
Inference-time scaling vs Proteina-Complexa | Better on 7 problems, tied on 2, worse on 1 |
Nipah glycoprotein G experiment | 200 designs, 34 passed the filter, 8 tested and all expressed; one bound with KD about 92 nM |
Successful designs from Genie 3 and BindCraft are mostly structurally distinct, so the authors describe the methods as complementary. The authors caution that cross-group hit-rate comparisons in the Nipah experiment are not controlled.
What is Tamarind Bio?
Tamarind Bio is a no-code bioinformatics platform built to give life scientists and researchers access to powerful computational tools. Many cutting-edge machine learning models are hard to deploy and use. Tamarind provides an intuitive, web-based environment that removes the complexity of high-performance computing, software dependencies and command-line interfaces.
The platform is designed for biologists, chemists and other researchers who may not have a background in programming or cloud infrastructure but want to run models on their own data. Key features include:
A user-friendly graphical interface for setting up and launching experiments
A robust API for integration into existing research pipelines
An automated system for managing and scaling computational resources
Tamarind treats information and data security as a top priority, as detailed in its Trust Center and Terms of Service.
Accelerating Discovery with Genie 3 on Tamarind Bio
De novo binder design: generate binders against a target using hotspot or interface residues.
Motif scaffolding: build new scaffolds around a functional motif, with all-atom motif handling.
Unconditional monomer generation: sample novel monomers, including long ones beyond the training length range.
Iterative hotspot refinement: reuse convergent hotspots across rounds, as shown for TNF-α.
How to Use Genie 3 on Tamarind Bio
Log in and open the tool: go to tamarind.bio and select Genie 3.
Pick the design mode: unconditional monomer, motif scaffolding or binder design, if exposed.
Provide inputs: for scaffolding, a motif PDB and scaffold length ranges; for binders, a target PDB with hotspot or interface residues and a binder length such as 80-120 in the repo examples.
Set parameters: number of samples, length range and, if exposed, the direction scale (the repo recommends 0.8 for monomers up to 300 residues, 0.1 for motif scaffolding, 0.0 for binders).
Run the job and wait for generation and filtering to complete.
Download outputs: designed PDB files and metrics such as self-consistency RMSD and pLDDT in the results table.
Interpret and filter: keep designs that pass thresholds such as scRMSD under 2 Å for monomers.
Use downstream: test top designs experimentally or re-score them with additional structure predictors.
Things to Keep in Mind
Success is judged mainly by in silico oracles; the paper says their precision is at most 12% and they fail entirely for some targets such as H3, TGF-β and TIE2.
The evaluation pipeline uses AlphaFold-Multimer rather than AlphaFold3 because of licensing, and needs MSAs for hard targets.
Targets are assumed to be in the bound conformation, so induced-fit effects are ignored.
Genie 3 was trained only on dimers, and experimental validation is minimal: eight designs against one target.
Binder design requires hotspot or interface residue definitions for each target.
Source: Yeqing Lin, Minji Lee, Aakarsh Vermani, Ellena Jiang, Sebastiaan De Cooman, Matej Špeťko and Mohammed AlQuraishi, "Fast and Ultra-Capable Protein Design: Advancing the Frontier Through Atomistic SE(3)-Equivariance with Genie 3," bioRxiv 2026, DOI 10.64898/2026.05.01.722168. Code: https://github.com/aqlaboratory/genie3.