📢We are excited to share SynCoGen—the first generative model that co-generates 🔷building-block graphs,🔷reaction edges and 🔷full 3-D coordinates, so every molecule comes with both a synthesis plan and a physically plausible shape.
On the whole, I’m incredibly optimistic about SynCoGen’s potential to design potent and synthesizable binders directly in 3D, and we’re just getting started—plenty more experiments are in the works, so stay tuned!
🚀 Thrilled to share that SynCoGen has been accepted to ICLR 2026!
Since the initial submission, we’ve been busy pushing the model further (see the updated paper here: https://t.co/R1o41Z8PiB. Here’s what we’ve put together in the meantime…
💊 We’ve also published trained weights for both unconditional and conditional generation:
• Unconditional: https://t.co/w9fFFFeJWQ
• Conditional: https://t.co/DvU5Z6q9dM
there’s a particular tech-adjacent solipsism where when they discover something it’s the first time it’s ever been discovered, and when they see something it’s the only way it’s ever been seen, and these tendencies occur in tandem
Internet cooks everyone’s brain but not in the same way. Some people get deep fried, some people get charred beyond recognition. For some people it’s a precisely calibrated sous vide. As for myself, I have achieved a perfect sear with a delicious, lacy golden crust
📜 Preprint: https://t.co/R1o41Z9n89
🧑🏻💻 Code: Soon!
If you're interested, I'd be happy to chat at the ICML GenBio workshop! Kudos to my amazing co-authors @MirunaCretu2, Dmytro Shevchuk, Pietro Lio, Rob Batey, Mike Tyers, Michal Koziarski, Vignesh Ram Somnath, and Cheng-Hao Liu!
📢We are excited to share SynCoGen—the first generative model that co-generates 🔷building-block graphs,🔷reaction edges and 🔷full 3-D coordinates, so every molecule comes with both a synthesis plan and a physically plausible shape.
📊 Results:
SynCoGen outperforms SOTA 3D molecular generation models on both structural quality and synthesizability while automatically proposing bench-ready reaction plans. Without any retraining, the model can also inpaint missing fragments between fixed substructures!