Genie 3

What is Genie 3?

Genie 3 is an all-atom SE(3)-equivariant diffusion model for protein design. Unlike backbone-only generators that produce only Cα traces, it models every heavy atom during generation, so output structures are ready for visualization, sequence inversion, and downstream validation without an extra full-atom reconstruction step.

The model supports three design tasks from a single checkpoint:

• Unconditional generation: creates novel protein structures from noise
• Motif scaffolding: holds fixed structural motifs and redesigns the surrounding scaffold
• Binder design: generates a candidate binder conditioned on hotspot residues of a target surface

Genie 3 is developed by the AQLaboratory. The implementation on ProteinIQ runs the official generation pipeline and returns each sampled structure as a PDB file.

How to use Genie 3 online

ProteinIQ hosts Genie 3 on GPU infrastructure and surfaces the three design modes through one form. Choose a mode, set the length or motif constraints, and request up to 20 samples. The job returns PDB files for every generated structure plus a file list, with no local installation of the model, sampling code, or evaluation dependencies.

Inputs

The required input depends on the selected design mode.

For motif scaffolding and binder design, the uploaded structure should be clean: one conformer per residue, no alternative locations, and residue numbering that matches the segment or hotspot specification. RCSB fetches retrieve the PDB file directly.

Core settings

Unconditional generation settings

If min_length equals max_length, every sample has the same length. If they differ, Genie 3 samples lengths across the range and generates n_sample structures per sampled length.

Motif scaffolding settings

The motif structure must contain a single protein chain. The total design length includes both the fixed motif residues and every generated flanking scaffold segment. The combined length must stay within the 256-residue limit.

Binder design settings

Hotspot residues define where the binder should contact the target. They must be present in the uploaded target structure and use the same chain IDs and residue numbers.

Outputs

Genie 3 returns one PDB file per generated structure. ProteinIQ renders the structures in an interactive 3D viewer and lists all files for download.

If predict_sequence is enabled, residue names in the generated PDB files reflect the predicted amino acid sequence. The returned structures are not energy-minimized or folded by a structure predictor by default.

How Genie 3 works

Genie 3 treats protein design as a diffusion process over 3D atomic coordinates. During training, the model learns to reverse a noise-corruption process that gradually perturbs real protein structures. At inference time, it starts from random coordinates and iteratively denoises them into physically plausible protein geometries.

All-atom SE(3)-equivariance

SE(3)-equivariance means the model's predictions transform consistently with rotations and translations of the input. If the input structure is rotated, the output rotates by the same amount without changing the internal geometry. This removes the need for data augmentation and lets the model learn geometric relationships in a coordinate-independent way.

Modeling all atoms, not just the backbone, matters most for design tasks where side-chain packing determines success. Binder interfaces, motif scaffolding constraints, and hydrogen-bond networks all depend on accurate heavy-atom geometry.

Guidance and sampling

The mode-specific direction scale controls classifier-free guidance strength during DDIM sampling. Genie 3's example configurations use 0.8 for unconditional short proteins, 0.1 for motif scaffolding, and 0.0 for binder design. Higher values follow the model direction more strongly; lower values allow more diversity.

n_sample_step sets the number of denoising iterations. 100 steps match the Genie 3 default and balance quality with runtime. Fewer steps run faster but may leave residual noise. More steps refine local geometry at increased cost.

Interpreting generated structures

Generated structures should be treated as design candidates, not final experimental structures. The diffusion model produces geometrically plausible backbones and side-chain conformations, but it does not guarantee foldability, stability, or binding affinity.

A practical validation workflow on ProteinIQ looks like this:

1. Inspect the PDB files in the interactive viewer. Look for closed, compact folds in unconditional designs and continuous backbone connectivity in scaffolds.
2. Run promising structures through AlphaFold 2, ESMFold, or Boltz-2 to check whether the sequence folds back to the designed structure.
3. For binders, predict the binder-target complex and inspect interface confidence. Low predicted aligned error at the interface and high interface pTM are positive signs.
4. Use ProteinMPNN or LigandMPNN to redesign sequences for generated backbones before experimental testing.

Common warning signs include:

• Clashing side chains or broken chain connectivity
• Extended, unstructured loops that dominate the design
• Binder designs where the generated chain does not contact the specified hotspots
• Self-consistency RMSD much larger than 2 Å when the sequence is folded back through a structure predictor

When to use Genie 3 vs alternatives

Genie 3 is a generative structure model. It is most useful when the goal is to explore new protein geometries rather than evaluate an existing sequence or complex.

Genie 3 vs BindCraft

Both tools generate protein binders, but their philosophies differ. Genie 3 samples from a diffusion model conditioned on target hotspots and returns raw structures quickly. BindCraft runs a longer optimization loop that uses AlphaFold2-Multimer as an objective function and filters designs with interface metrics. Genie 3 is better for rapid exploration; BindCraft is better when built-in in-silico filtering and relaxation are worth the extra compute.

Limitations

• No built-in folding validation: Genie 3 returns generated structures only. It does not run ESMFold, ColabFold, or AlphaFold2 to check self-consistency. Users must validate separately.
• Length cap: The current ProteinIQ deployment supports designs up to 256 residues and 2,048 tokens per batch.
• Sequence prediction is optional: Predicted sequences from predict_sequence are model proposals. They should be re-evaluated with an inverse-folding model or structure predictor before ordering genes.
• Binder design requires hotspot knowledge: The quality of binder designs depends heavily on choosing reasonable hotspot residues. Poorly chosen hotspots produce binders that miss the intended interface.
• No small-molecule or nucleic-acid support: Genie 3 designs proteins only. It cannot generate DNA, RNA, or ligand structures.
• Stochastic output: Results vary between runs unless the underlying sampling seed is fixed. Multiple samples are recommended to explore the design space.