Pedro Beltrao

Wow^2

I'm thrilled to be heading back to the always-wonderful Biology of Genomes conference this year. Hope to see many of you there!

A special opportunity for lovers of mass spectrometry proteomics to join a like-minded team at PTI.

Join a collaborative initiative to enable direct protein analysis at unprecedented scale, in partnership with leading instrument developers, academics, and industry leaders.

1/2

SVG would be great. it sometimes can do some bits of an SVG but the quality is so poor. The capabilities of these models is so uneven. "Jagged intelligence" as they call it.

When I was playing around with Gemini generation and manipulation of scientific images, I also tried this beautification aspect and noted (as do the authors here) that there is a high risk of distorting the actual data.

He does go over those analogies. He starts very much on the idea that new technologies are often derided for not being proper art. He tries to make a more subtle point that art is the consumption but the creation. We grow when we try to make art in a way that (current) AI models don't.

Swiss Personalized Health Network (SPHN) unlocks data from 800,000 patients Researchers can now easily discover what clinical health data exists in Switzerland and how to access it, while fully respecting patient privacy. This major milestone is enabled by the release of stan...

🎉 Researchers can now explore a foundational resource for data-driven healthcare solutions – and see how to access the data for personalized health studies.
The milestone is enabled by release of metadata for the Swiss Personalized Health Network (SPHN) Federated Clinical Routine Dataset. More👇

YouTube video by Brandon Sanderson We Are The Art | Brandon Sanderson’s Keynote Speech

I don't agree with everything but this is a good take from a book writer on the AI and art. A bit long and geeky. Why do we root for a scify humanoid (Data in StarTrek) trying to learn about art but dislike LLMs doing it? Data was doing it because he wanted to grow
www.youtube.com/watch?v=mb3u...

Updated SNSF grant rules, valid for the next round
- Max 2 project grants , 1 being a Lead Agency/Weave/ICIS project.
- Max 1 proposal per 12 month period
- Max personnel costs 200k per year, max group total budget 250k per year as before
- Max 3 million per project (= max 3 PIs?)

EiraBio

Our company has a new name, Eira Bio (www.eirabio.com) (formerly K2 Therapeutics), but the same important mission to apply our advanced synthetic biology, directed evolution, and protein design technologies to discover and develop next-generation antibody-based medicines for a range of diseases!

We have used the changes in phosphorylation of substrates of kinases to predict kinase activity changes. We then refer to this as predicted or inferred kinase activity. We could always write changes in phosphorylation of a set of target sites of kinase X but it would make writing very convoluted.

Right, a bunch of adverts,big figures, marketing and viral posts the culture is much more of self promotion. I end up falling into it as well and I don't like it.

Yep. I wonder how much of it comes from the algorithmic feed. It makes it more annoying to view things as we end up getting viral posts but then it may help propagate some of the science better too.

I don't like that LinkedIn has become the most active social media for science. It feels less personal than twitter/bluesky.

Research Group Leader Do you want to lead groundbreaking research in computational biology? Join us at EMBL-EBI! EMBL's European Bioinformatics Institute (EMBL-EBI) is seeking talented and highly-motivated scientists to jo...

We are hiring for group leaders again — EBI is a great place to start your research group!

embl.wd103.myworkdayjobs.com/EMBL/job/Hin...

Six new National Centres of Competence in Research #NCCR announced today in strategically important areas such as medicine, quantum technology and climate: Children & Cancer, CLIM+, Genesis, Muoniverse, Precision, and Separations.

www.wbf.admin.ch/en/newnsb/JE...

@snsf.ch @sbfi.admin.ch

Multiple protein structure alignment at scale with FoldMason Protein structure is conserved beyond sequence, making multiple structural alignment (MSTA) essential for analyzing distantly related proteins. Computational prediction methods have vastly extended ou...

FoldMason is out now in @science.org. It generates accurate multiple structure alignments for thousands of protein structures in seconds. Great work by Cameron L. M. Gilchrist and @milot.bsky.social.
📄 www.science.org/doi/10.1126/...
🌐 search.foldseek.com/foldmason
💾 github.com/steineggerla...

A Chemical-Genetic Interaction Matrix Reveals Drug Mechanism and Genetic Architecture https://www.biorxiv.org/content/10.64898/2026.01.28.701823v1

Combining the genetics and trait-linked differential expression is on the list and we did cite your preprint in the discussion in that context. We though of a few options for how to do it but we haven't tried to compile the necessary gene expression data across a large set of traits/diseases.

The convergence at protein interaction modules has also been studied in a preprint by the Ideker lab pubmed.ncbi.nlm.nih.gov/40666368 with a different method. Some continuations for us include supervising the latent space with known pairs of related traits, or combining with transcriptional data.

Finally, we can take advantage of the generative aspects of the VAE by averaging neighbor traits and decode this to predict candidate trait/disease gene. Essentially, transferring information from related traits. We explore this in mode detail for heart rhythm disorder - Long QT Syndrome (LQTS).

Based on this we find that indeed common and rare variant studies tend to converge on common pathways/complexes. For a hearing loss, shown below, we see that even in a small network module common and rare linked genes highlight different stereocilial substructures.

Encoding human traits, mouse KO phenotypes and GO terms organizes these traits and terms into neighborhoods related by the distances of their gene sets projected in an interaction network. If we enconde a drug via its known targets we find it is placed near its known disease indication(s).

We combined a protein interaction network embedding and a VAE to encode traits into a lower dimension representation where traits should be closer if their linked genes are near in the network. As expected the same or similar traits are indeed placed closer together independently of gene overlaps.

We find that, for the same trait, the genes identified by common and rare variant studies are mostly non overlapping, in part due to biases. The Pritchard lab studied this for burden and GWAS in detail pubmed.ncbi.nlm.nih.gov/41193809. Are the different genes at least related to each other?

New lab preprint - Common and rare variant studies for the same trait identify different genes and here Diederik Laman Trip developed a protein network AI enconding to investigate if traits studied by different approaches converge on the same molecular pathways
www.biorxiv.org/content/10.6...

I acknowledge the financial issues which is not something I had encountered before when talking with USA faculty about this. The most typical concern has been them feeling entitled to having the job for life. In most countries in Europe there would be a centralized pension scheme and health care.

There is the obvious fraud potential, but I think something much more fundamental is going to happen, too. It will change the meaning of what a "research publication" is. The reason why people read papers is to learn something that they couldn't reproduce in seconds by pushing a few buttons.

In the specific case of Switzerland there is an insurance based health system (privately paid) and most of the pension is contributed over course of the life by the employer+employee. More generally, I favour tax-based universal health care & enforced personal pension savings with social support.