Use Genie 3 Online

Genie 3: Advancing the Frontier of Generative Molecular Modeling Through Atomistic SE(3)-Equivariance

Genie 3 is a revolutionary structure diffusion model engineered to bridge and surpass the generation-hallucination gap in protein design. By introducing a branched polymer treatment of protein structures, Genie 3 achieves state-of-the-art capabilities across critical design tasks—unconditional monomer generation, complex motif scaffolding, and de novo high-affinity binder design—all while running orders of magnitude faster than traditional hallucination-based methods.

Developed by researchers at Columbia University, UC Berkeley, Puxano BV, and IT4Innovations National Supercomputing Center, Genie 3 marks a new frontier in structural biology and molecular architecture.

Key Capabilities

1. High-Affinity Binder Design

Genie 3 delivers state-of-the-art performance in producing high-affinity binders targeting complex proteins. Trained on both monomeric and multimeric datasets (including the interface-clustered Pinder complex dataset), it generates structures explicitly conditioned on a target's interface and hotspot residues. It reliably complies with user-specified hotspot inputs, avoiding the off-target binding typical of competing tools.

2. Precise Motif Scaffolding

Engineered with a partial atomization approach, Genie 3 scaffolds functional motifs (such as enzyme active sites) by maintaining full all-atom detail for the functional segment while efficiently generating backbone frameworks for scaffold segments. On the standardized MotifBench profile, Genie 3 outperforms leading methods by solving more complex problems with superior sidechain conformation concordance.

3. Out-of-Domain Long Monomer Generation

While trained on structures up to 256 residues, Genie 3 demonstrates strong out-of-domain generalization capabilities. It produces highly designable, diverse, and novel long monomers spanning up to 800 residues, outperforming alternative structural diffusion pipelines.

Performance & Speed Benchmarks

Genie 3 closes the gap between generative diffusion and costly sequence-optimization approaches like BindCraft, optimizing computational budgets by prioritizing structural diversity over diminishing sequence-level mutations.

• 10x Faster Sampling: By optimizing its reverse sampling setup with an implicit model reformulation (Δt=10$), Genie 3 slashes the number of structural denoising steps from 1,000 to 100 with negligible impact on final generative quality.

• Superior Efficiency: In binder design tasks across monomeric and multimeric targets, Genie 3 achieves the highest number of unique successes normalized per GPU-hour compared to BoltzGen, RFDiffusion, Proteina-Complexa, and BindCraft.

Per-Sample Structural Generation Runtimes (Seconds)

(Benchmarks conducted on a single Nvidia A100 40GB GPU).

Real-World Validation: The Nipah Virus Challenge

To demonstrate real-world therapeutic viability, Genie 3 was evaluated in Adaptyv Bio’s Nipah Competition to engineer a de novo binder against the under-characterized Nipah virus Glycoprotein G (NiV-G)—a tetrameric protein critical for host cell attachment and pathogenesis.

• Nanomolar Affinity: After generating 200 structural designs and filtering via an in silico structure prediction pipeline, the top candidates were expressed in vitro using a cell-free system. Genie 3 successfully produced a verified binder with a highly potent binding affinity (KD) of 92 nM.

• 12.5% Experimental Success Rate: Out of only 8 attempts engineered and tested via Surface Plasmon Resonance (SPR), Genie 3 yielded a 1/8 experimental hit rate, noticeably outperforming aggregate candidate pools from other architectures.

Underlying Technology

Genie 3 re-examines the role of physical symmetry in deep learning by establishing that structural models do not need to sacrifice efficiency to achieve detailed sidechain configuration reasoning.

• Branched Polymer Structure Representation: Proteins are modeled as continuous branched polymers, anchoring Frenet-Serret (FS) frames at the Cα backbone line as well as across sidechain heavy atoms following the atom14 configuration convention.

• True SE(3)-Equivariant Denoising: Unlike other modern networks that drop physical symmetry and depend on expensive data augmentation with random rotations and translations, Genie 3 implements an explicitly SE(3)-equivariant denoiser.

• Bidirectional Latent Transformer: Features a sophisticated architecture where single-residue and residue-residue pair representations continuously communicate information bidirectionally at every single layer via Invariant Point Attention (IPA), attention pair biases, and outer product projections.

• Global Tokens: Integrates global stream tokens during structural modeling to enforce broad topological reasoning across massive molecular scales.

• Directional Noise Scaling Heuristic: Leverages an implicit reverse diffusion sampling scheme adjusted by a directional scale parameter (eta), yielding rigid, stable, and highly designable protein backbones without inference-time distribution mismatch.

What is Tamarind Bio?

Tamarind Bio provides a next-generation computational platform that simplifies advanced molecular biology tools for researchers and developers. By rendering complex backend deep-learning models accessible via web-based, zero-setup interfaces and cloud APIs, Tamarind Bio accelerates breakthroughs in drug discovery, enzyme engineering, and therapeutic design.

How to Use Genie 3 on Tamarind Bio

Genie 3 is a specialized language model developed to predict and optimize peptide sequences that interact selectively with target proteins. Running Genie 3 on Tamarind Bio requires just a few streamlined actions:

1. Upload Target Context: Provide the structural file (such as a PDB) or sequence of the target protein on the Tamarind Bio platform.

2. Define Binding Conditions: Highlight or select target hotspot residues or target interface regions where binding needs to be engineered.

3. Configure Parameter Metrics: Set your preferred generation limits, diversity factors, and sequence constraints.

4. Initiate Generation: Submit the prompt to execute the backend pipeline. Genie 3 will automatically compute and report optimized candidate peptide sequences with high predicted affinity.

Frequently Asked Questions

How does Genie 3 handle input conditioning for binder design?

Genie 3 utilizes chain identifiers and binding interface mask tokens within its input embedder. During generation, it implements two advanced interface expansion heuristics—"extended hotspots" (for monomeric targets) and "convergent hotspots" (for multimeric targets)—to enlarge user-provided hotspot residue sets into a robust interface prior that matches its training distribution.

What datasets was the model trained on?

Genie 3 was trained using a multi-task regimen featuring a 9:1 ratio between motif scaffolding and binder design tasks. It leverages Foldseek-clustered structural configurations from the AlphaFold Protein Structure Database (filtered for high confidence, pLDDT >= 80) for monomer data, and the interface-clustered Pinder dataset for experimentally determined protein dimeric complexes.

Does Genie 3 co-generate sequences and structures?

Genie 3 focuses on generating structural configurations. Downstream sequence design is handled effectively via an inverse folding step utilizing ProteinMPNN.