Genie 3 also introduces a new inference-time scaling paradigm, where information from past design rounds is injected into later ones, altering the underlying sampling distribution as compute scales. This yields dramatic gains, turning nearly intractable → tractable problems. 6/8
Genie 3 brings two architectural innovations:
1. All-atom SE(3)-equivariance using branching frames for sidechains
2. A LatentTransformer that tightly couples single and pair latents
They combine to give a much stronger foundation model for geometric reasoning over proteins. 2/8
We see this in unconditional generation, where Genie 3 generalizes to 800 residues despite being trained only on proteins <250 residues in length.
We see this in motif scaffolding, where Genie 3 is an outlier on MotifBench relative to other methods. 3/8
We’re currently experimentally testing Genie 3 designs across a broad range of binder targets. We also submitted 8 designs to the AdaptyBio Nipah challenge, all of which expressed, with one achieving a Kd of 92 nM. 7/8
On vanilla inference-time scaling, where scaling = just more sampling, Genie 3 outperforms Proteina-Complexa on most tasks. On multimeric tasks, it shows robust scaling behavior in all instances, even when Proteina-Complexa does not. 5/8
We'll update the preprint with all experimental results once they're finalized. For now we wanted to get Genie 3 into people’s hands right away. As usual, code + model weights are under an Apache 2.0 license.
Preprint: www.biorxiv.org/content/10.6...
GitHub: github.com/aqlaboratory...
8/8
In binder design, Genie 3 yields qualitative gains on the AlphaProteo benchmark set. Compute-normalized or not, it's almost always the best model, with up to 20x gains on some hard multimeric tasks. Its lowest success rate is >1% per GPU hour, while other models can fail entirely. 4/8
Introducing Genie 3, a generative protein model that substantially advances the state-of-the-art for binder design, increasing in silico success rates by up to 20x on hard multimeric targets. It also debuts a form of inference-time scaling unobserved in other design models. 🧵1/8
