The proteina-complexa paper is the 2nd this month suggesting that high K+E content, a staple of ProteinMPNN designs, is predictive of poor expression (the bits in bio benchmark paper showed a few weeks ago). In their case, they measure poor sequence recovery in phage display

I feel strongly this is a terrible idea. I battle it with my students all the time. Examples: - Can you quickly make this minor change to this figure? That'll take 30 min. to rerun all the preprocessing. - Can you send me the raw data to this figure that contains 5 points? That'll be 10 TB of data.

The Google Scholar preprint bug strikes again: I now think this is a version of it. GS can't find the PMC version of my article, even though it clearly exists, and hence harasses me about failing accessibility mandates. 1/

Excellent point. If you haven't personally cured cancer, don't opine on whether AI can do it.

Google Scholar claims an article we posted on bioRxiv is not publicly available. Has anybody encountered this? Is there something to be done? @richardsever.bsky.social

Why is this a news story? Cargill uses computer vision to make their butchering process 0.5% more efficient. Next up: Chemical engineer at DuPont discovers that a slight adjustment in buffer chemistry results in 0.5% less waste during nylon production.

The reverse is also true though. Americans want to do "Europe" in a week and it's absolutely too big for that.

Protein language models are bad at mutational effect prediction blog.genesmindsmachines.com/p/protein-la...

I have a joke about grant funding, but it didn't make the payline.

I always feel uncomfortable when I'm in one of these gondolas. They're attached to the cable with a strong spring. It seems a bit sketchy. (Though in fairness, gondola crashes are extremely rare.) www.bbc.com/news/article...

Wow, looks great! Looking forward to nicely formatted equations in ggplot2 axis labels.

NVidia announces new chip for space data centers. The chip works great in space. How you get it there, how you power it, how you cool it, those are your problems. But once you've figured them out, the NVidia chip will work just fine.

Some innovative ideas for sequence generation with protein language models.

Yup, this is the SpringerNature policy. If you're NIH funded you _must_ pay for Gold OA. Think twice before sending your paper there.

In biological data finding meaningful ways to do training-test splits without leaking information can be tricky as readouts can be highly correlated, causing over-optimism. Here a nice example based on protein language models 👇

Good news, protein engineering researchers, mutation effect predictions remain an unsolved problem.

New paper showing that much of the apparent success of protein language models in predicting mutational effects is a mirage: These models mostly memorize sites. 1/ www.biorxiv.org/content/10.6...

Looks like a very important check on LLMs in protein structure/function prediction.

Pretty much the same reason all the gene regulatory network inference models don't work as well as people think they do

How often an appropriate choice of benchmark is absent from a paper announcing a new technique or approach, and it takes papers like this to make the comparisons ;)

One secondary observation from this work: Note how bad ESM C is for viral data. Look at part B, second column, yellow box plots. Predictions for viral data are terrible, while predictions for cellular data are quite good. Don't use ESM C if you're working with viruses.

New paper showing that much of the apparent success of protein language models in predicting mutational effects is a mirage: These models mostly memorize sites. 1/ www.biorxiv.org/content/10.6...

Using Quarto to write (and typeset) a book.

If you use #QuartoPub for any PDF documents for teaching and you need to meet impending PDF accessibility rules, upgrade to v1.9, add format: typst: pdf-standard: ua-1 to the YAML front matter, and it'll work!

Nothing to see here. Just the normal ups and downs of markets.

New paper from former PhD student @tkschulze.bsky.social on supervised learning of protein variant effects across large-scale mutagenesis datasets MAVE/DMS experiments provide large amounts of data for benchmarking variant effect predictors, but may be difficult to use in supervised learning. 1/5

Effective payline for NIGMS: 0% 👀

My uncle was an early IT entrepreneur. I remember him telling me in the 1990s that programming was not going to be a career for much longer because of progress in no-code software design tools. Yes, that was 40 years ago.

Here’s a full draft of the upcoming second edition of my “Data Visualization: A Practical Introduction”: socviz.co

#PostdocRecruiting: My group at the University of Texas at Austin is recruiting a Postdoc in virology and structural biology. Please see more details here: utaustin.wd1.myworkdayjobs.com/UTstaff/job/... Informal inquiries are encouraged, and see more lab info here: zunlongke-lab.org

Results from an impressive world-wide binder design competition.

2026 has been a remarkable year for me so far (despite the current world events): opening my lab at the University of Zurich, assembling an amazing group of people in the group, and celebrating it with a fantastic lab warming party. Figured it's also time for a small profile refresh.

There's a city called Bangor just down the road, but that's too simple. Let's instead report on the weather in Llanfairpwllgwyngyllgogerychwyrndrobwllllantysiliogogogoch.

This is fantastic. Must watch. www.youtube.com/watch?v=eEbz...

This is fantastic. Must watch. www.youtube.com/watch?v=eEbz...

My assignments have an AI disclosure statement where students need to explain how they used AI. AI as a support tool is allowed. But AI is not allowed to write entire assignments. Will Einstein lie or will it properly disclose that the entire assignment was done by AI in violation of the syllabus?

Three postdoctoral fellowships in quantitative biology are available as part of a new Quantitative Biology Initiative in the Department of Biology at the University of Maryland. Best consideration date: 3/14 Job Ad: umd.wd1.myworkdayjobs.com/en-US/UMCP/j... Department: biology.umd.edu/people

Some thoughts prompted by the recent incident of an AI agent trying to smear and shame a matplotlib developer. We need to think about building AI with empathy. Sociopathic AI (which is what we have today) will turn into a nightmare. blog.genesmindsmachines.com/p/sociopathi...

Some thoughts prompted by the recent incident of an AI agent trying to smear and shame a matplotlib developer. We need to think about building AI with empathy. Sociopathic AI (which is what we have today) will turn into a nightmare. blog.genesmindsmachines.com/p/sociopathi...

This is the first time that I’ve seen an actual, mechanistic explanation for how the shingles vaccine may protect against dementia. Very interesting! If you’re in your 50s, definitely get your vaccine. profbubak.substack.com/p/how-the-sh...

Olympics aren't what they used to be without Simone.

Challenges for open ecosystems as "first-of-its-kind case study of misaligned AI behavior in the wild...raises serious concerns about agents executing blackmail threats" theshamblog.com/an-ai-agent-...

"Blackmail is a known theoretical issue with AI agents. In internal testing at the major AI lab Anthropic last year, they tried to avoid being shut down by threatening to expose extramarital affairs, leaking confidential information, and taking lethal actions." 🙄

GitHub just posted this about some new changes they’ve just made + changes they have planned to try limiting PRs from LLMs, but thst wont stop wild stories like this rogue agent launching smear campaigns

Wait this is insane. An autonomous AI agent has its code rejected, so it writes a public blog post attacking attempting to smear the editor who rejected it. We have already let this go way too far.

This happened to the maintainer of matplotlib btw. Not some little-known, obscure project.

AI agent goes nuts on open source maintainer after having its pull request denied. This is a pretty insane story. Open source development as we've been used to for the last few decades is likely over. 1/ theshamblog.com/an-ai-agent-...

I once watched "Professional drummer hears Enter Sandman for the first time" and my gut reaction was how is this even possible? How can you live in the civilized world and be a professional musician and you've never heard this song? Hard to believe. 1/2

Just a regular reminder that there is absolutely no scientific evidence that caffeine (in moderation) is in any way problematic. Caffeine is this thing that you think should be bad for you but it's just not. Most studies suggest it has meaningful benefits. See reply for a recent example.

Let's be clear: Musk does not currently have a vehicle with the lift capacity and reusability required to build a city on Mars. He doesn't have a vehicle to build a city on the Moon either.