When life gives you lemons: "Here, we show a workflow we used for the identification of a contaminant from a cryoEM grid without prior knowledge of protein sequence..." in other words "direct sequence identification from the cryoEM map"

Great interview. scientificamerican.com/article/breakt…

Having some fun with Manifold leading up to AACR this week. manifold.markets/NiklasRindtorf… (platform is play money only)

The 10 most striking Kaplan-Meier curves (that I have seen) 🧵 1. Dual GLP1/GIP agonist, Tirzepatide, for prevention of type 2 diabetes in people with obesity It's easy to miss the dark blue line (15 mg group), because it never leaves the X axis:

What are other targets that have notoriously lousy commercial antibodies right now?

New #AI-ready dataset = a leap forward for drug discovery & energy technologies. AI at Meta + Berkeley Lab co-led #OMol25: the most diverse open-source chemistry dataset ever built. Science just got a major upgrade. @lbnlcs Berkeley Lab ETA #OpenScience newscenter.lbl.gov/2025/05/14/com…

From Likelihood to Fitness: Improving Variant Effect Prediction in Protein and Genome Language Models １．This study introduces Likelihood-Fitness Bridging (LFB), a method that improves variant effect prediction in protein and genome language models (pLMs and gLMs) by averaging

This is an incredible progress story An aggressive type of cancer (Philadelphia chromosome–positive acute lymphoblastic leukemia) used to carry a dismal prognosis, with 50% of patients dying in a year New therapies have transformed it to a cancer that most people are cured of

Here's the key graph, comparing performance by ether0 and frontier models across the chemistry tasks we trained + evaluated on. Performance increases on the open-ended tasks are pretty strong. Interestingly, we found that improvements are harder to see on the multiple choice than

If the largest HPC in 2030 will consume 9 GW (the largest nuclear plants in the world can barely deliver this), what we consider to be noteworthy compute (down to 2 OOM below that number) will still draw as much power as a medium size solar setup / small coal plant produces

FreeTimeGS: Free Gaussian Primitives at Anytime Anywhere for Dynamic Scene Reconstruction Paper + demos + code zju3dv.github.io/freetimegs/

Seems like a lot of the pre- vs. post-training compute requirements can be boiled down to "The lower the rate of observing useful training information, the lower the bandwith between nodes can be." Also great to see Prime Intellect getting a shoutout!

o3 pro is very very good.

Microsoft researchers achieved a breakthrough in the accuracy of DFT, a method for predicting the properties of molecules and materials, by using deep learning. This work can lead to better batteries, green fertilizers, precision drug discovery, and more. msft.it/6011SQwKX

Catenane-like binding modes when designing cyclic peptides with Boltz-2: The target protein is not cyclic, so not a catenane in sensu stricto. That said, something I haven't seen before.

It is only called Feynman-Kac when it is from the Feynman-Kac region of France, otherwise it is just sparkling Sequential Monte Carlo.

cathedrals everywhere, for those with eyes to see