Yeqing Lin

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: biorxiv.org/content/10.648… GitHub: github.com/aqlaboratory/g… 8/8

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

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

@_ncyxx @MoAlQuraishi (2/2) However, when using a sampling temperature of 0.6, the designability is significantly increased to 96%.

@_ncyxx @MoAlQuraishi (1/2) We found that lowering the sampling temperature helps to boost designability. For reference, when using a sampling temperature of 1, we observe that the designability is less than 40%, which is the percentage of designable structures in our subsampled AFDB structures.

Introducing Genie 2, our latest protein design model! Genie 2 sets a new state of the art on key design metrics and supports motif scaffolding, including that of multiple motifs with undefined geometric relationships, a feature new to protein diffusion models. (1/5)

@RolandDunbrack @MoAlQuraishi @m1nj12 @zhaozhang_hpc The code for Genie 2 are now available at github.com/aqlaboratory/g….

@MoAlQuraishi @lin_yeqing @m1nj12 @zhaozhang_hpc Looking forward to the code/weights to try it out (only license and readme on github so far).

@rravi @MoAlQuraishi @m1nj12 @zhaozhang_hpc The code for Genie 2 are now available at github.com/aqlaboratory/g….

@MoAlQuraishi @lin_yeqing @m1nj12 @zhaozhang_hpc Thanks. The linked github doesn't have code yet. Any timelines?

@MoAlQuraishi @m1nj12 @zhaozhang_hpc The code for Genie 2 are now available at github.com/aqlaboratory/g….

Code for training and inference, and model weights, here: github.com/aqlaboratory/g… Preprint: arxiv.org/abs/2405.15489 Actual work was done by phenomenal students @lin_yeqing and @m1nj12 and with helpful support from @zhaozhang_hpc at Rutgers. Thanks to all involved! (5/5)

@AjasjaLjubetic @MoAlQuraishi We choose to use ESMFold since it does not require MSA and in terms of self-consistency evaluations, the performance does not vary a lot between using AF2 and ESMFold.

@MoAlQuraishi Looks very interesting, especially the multi scaffolding feature! Is there a reason you chose ESMFold over AF2 (or OpenFold), apart from lower computational cost? ESMFold finished 83rd in CASP15. I realize de novo proteins don't have MSAs, but AF2 still seems to perform very well

@RolandDunbrack @MoAlQuraishi @m1nj12 @zhaozhang_hpc We are currently cleaning up our codebase and adding documentations. We will release the training and inference code, together with our weights, in a few day.