I've been working on variants of this problem for my entire career, and have only dreamed of designing molecules on the computer with success rates this high. Proud of @chaidiscovery for pulling this off, and doing it with the speed, rigor and grace that was only imaginable ✨.

Thank you for having us on the show! I've been a long time listener of @NoPriorsPod. It was fun to catch up with @saranormous and @_jackdent to talk more about Chai-2 and what's coming next.

This is a massive understatement. @Mattmcpartlon1’s brilliance is unparalleled. But he’s also an incredibly thoughtful, disciplined and caring leader. Every day working with @Mattmcpartlon1 challenges me & the company to think bigger, work harder, and show up stronger. 🖤.

RT @byersblake: I appreciate Chai’s hard work to ensure my predictions are accurate.

RT @alexkaplan0: We've just passed the Singularity moment for molecular engineering. This might be the most important thing you read today.….

RT @NaveenGRao: THIS is what gets me excited about AI! Using generative techniques to solves previously computationally intractable problem….

RT @AndrewE_Dunn: NEW: The OpenAI-backed biotech startup Chai Discovery announced a new suite of AI models today, called Chai-2. The main r….

RT @modal_labs: Some use cases are cooler than others, and it's safe to say zero-shot antibody design is one of them. Congrats on launchin….

I'll be going on @tbpn at 1:30pm PST today to talk about Chai-2 and what's next for us @chaidiscovery. Make sure to tune in!.

RT @Sino_Biological: Congratulations to @chaidiscovery on this incredible milestone. We are proud to have partnered with the Chai team on s….

RT @adaptyvbio: Chai Discovery just released their new model Chai-2, validated on Adaptyv! . Chai-2 enables generation of de novo proteins….

RT @chaidiscovery: We're excited to partner with the UK Government on OpenBind, an effort to create the largest open dataset of drug-protei….

Welcome to the team, Zhuoran! @ZhuoranQ is an AI-bio legend, having led the development of NeuralPlexer1-3. I've been following Zhuoran's work for a long time and it's great pleasure to now work together at @chaidiscovery. Welcome aboard! 🚀.

RT @SeqeraLabs: Exciting news! 🎉You can now try Chai-1, the SOTA model for biomolecular structure prediction by @chaidiscovery, with Seqera….

RT @StephaneRedon: Excited to announce the general availability of Chai-1 in SAMSON! Like AlphaFold 3, Chai-1 is a state-of-the-art foundat….

We are excited to see continuous improvement on models and evaluation. Important takeaways:.- Small sample sizes can result in misleading conclusions.- Error bars are critical for accurate benchmarking.- AI moves fast, but we need to be careful about critical performance claims.

With expanded evaluation sets, the error bars tighten up. On a set of 665 recent PDBs, we found a noticeable gap between the methods: Chai-1 has the best performance, while Boltz-1 and AlphaFold2 are roughly equivalent.

For example, we found that if we look at the best rather than the mean prediction – left, consistent with the methodology presented in the Boltz-1 report – the order of the results changes, highlighting variance in this experiment and the need for more rigorous comparisons.

We compared DockQ>=0.23 on the remaining 13 targets, and found that since the number of targets is so small, the error bars are large and overlapping. High p-values from comparisons indicate insufficient statistical significance to draw conclusions.

We ran Chai-1, AlphaFold2.3 and Boltz-1 on the 20 multimeric targets in CASP15 with <2048 tokens. The official DockQ implementation (v2.1.2) failed to compute metrics on 4 examples, and 3 OOM’d on Boltz-1 (on an 80GB A100). We report mean across 5 random samples for all models.