New week, new paper from the lab. This time we go back to our favorite protein-protein interactions and their specificity - G proteins and RGS proteins: rdcu.be/cpHQI The following thread explains (1/7):

Our recent paper explained

Don't miss the interview with the first author, Anna Fassler Bakhman. '... I volunteered ... as a first-aid ambulance responder, where I encountered a wide range of medical situations and saw how biological processes directly affect people's lives.' doi.org/10.1242/dmm.052933

🚨 JUST IN: A migratory bird just shattered world records — flying 8,425 miles (13,560 km) NON-STOP across the Pacific without landing once. The bar-tailed godwit doesn’t stop to eat, drink, or sleep during its migration across the Pacific Ocean. Its journey from Alaska to Australia takes roughly 11 days of continuous flight, covering over 13,000 kilometers through storms, headwinds, and open ocean with zero land beneath it the entire time. Before departure, it does something almost surgical to its own body. It shrinks its digestive organs down to almost nothing, converting the stomach, intestines, and liver into raw fuel. The bird essentially eats its own gut to make room for fat reserves that will power its wings for nearly two weeks straight. The brain doesn’t fully sleep either. Half of it stays active while the other half rests, alternating in shifts mid-flight at altitude over the open Pacific. The godwit is simultaneously unconscious and navigating with magnetic field sensitivity that no human instrument in the 18th century could replicate. What makes this genuinely staggering beyond the physical record is the navigational precision involved. The bird leaves Alaska and arrives in New Zealand with accuracy that would embarrass early GPS systems. It reads Earth’s magnetic field, atmospheric pressure gradients, star positions, and potentially quantum-level compass mechanisms inside its eye that literally let it see magnetic field lines overlaid on its visual field. Evolution spent millions of years building an aerospace navigation system inside a 300 gram animal. We spend billions engineering machines that do what this bird does on instinct, fat reserves, and half a sleeping brain. The longest recorded non-stop flight by a commercial aircraft is around 20 hours. This bird does 11 days. Without a runway.

~20 years ago Holm & Sander said we should map the protein universe. Now AI can power the mapping, offering an unprecedent global look at protein similarities. Many more cool details in the pre-print, including sub-sequence embedding and comparisons to ESM3, ProstT5, and ProTrek

We can map protein space to study its properties. Here are maps with structural class, metal, and co-factor binding. There is a striking split between the region of domains with few vs. those with many cofactor associations [but metal binding domains are all over space]

Domains of the same architecture and even the same structural class are next to each other. This was all learned with self-supervision. No ECOD labels were used in training. The agreement with the structural labels is astonishing (for both structure and sequence modalities).

We used it to create maps of protein space. Here is a map of ~30,000 ECOD representative domains colored by architecture (projected to 2D with tSNE). We’re flipping between structure, sequence (and subsequence) embeddings: the sequence/structure modalities co-embed very well.

Our new AI model CLSS that can co-embed protein sequences and structures is out at tinyurl.com/4wsvswut. We use a CLIP-like lean contrastive model (~35M million trained parameters!) and have super-exciting results. With Guy Yanai, Gaby Axel, @LiamLongo, and Nir Ben-Tal.

If you were looking for a tool to run massive protein complex structure alignments, here’s Foldseek Multimer, based on the already classic Foldseek👇

Nature research paper: De novo designed proteins neutralize lethal snake venom toxins go.nature.com/4jt16RY

This is the first generative protein language model trained on three protein features: sequence, structure, and function. Using GFP as a test case, the model generated a bright functional protein with low sequence homology to known proteins – estimated to be equivalent to a simulation of five hundred million years of evolution. Congrats to the @pdhsu Lab for being part of this project! @tranvinq @BartieLiam @mnemeth101

Our paper doi.org/10.1073/pnas.2410003… just came out 🥳🥳🥳 We used a deep generative network to study codon usage in two eukaryotes: S. cerevisiae and S. pombe and two bacteria: E. coli and B. subtilis. This is what we did: 🧵 @TomerSidi197763

A popular FDA-approved drug to treat hypertension, hydralazine works, at least in part, by modulating RGS proteins! Honored to be a part of a collaborative team led by @MegaMatthewsLab! biorxiv.org/content/10.1101/…

Predicting absolute protein folding stability using generative models More details 🦋👇

Journal club scheduled

Very interesting, in @eLife: The membrane domains of mammalian adenylyl cyclases are lipid receptors doi.org/10.7554/eLife.101483…

Gene behind orange fur in cats found at last | Science | And guess what, it regulates G protein signaling ☺️ science.org/content/article/…

Thrilled to announce Boltz-1, the first open-source and commercially available model to achieve AlphaFold3-level accuracy on biomolecular structure prediction! An exciting collaboration with @jeremyWohlwend, @pas_saro and an amazing team at MIT and Genesis Therapeutics. A thread!

For a PhD student, choosing a good lab is 10 times more important than choosing a particular topic to study

Come see the scientists that should've shared the Nobel Prize for "Computational Protein Design", speak: ipd.uw.edu/protein-design-th…