Fabian Fröhlich

Out now! Our very first paper@crick.ac.uk, an impressive feat by @tobyandrews.bsky.social where we show how a developing heart grows and scales up its morphological complexity to keep beating...an excellent summary below ⬇️
www.cell.com/developmenta...

Yes, fully agreed that we should keep (investing in) public transport. But I think the issue with Desmonds opinion is that he is making straw man arguments against the metro, not that he is overstating the impact of AI on personal vehicles.

But self driving cars are a reality in some (small) parts of the US. Given the track record of AI in vision and the amount of training data, it’s not too far fetched that they could replace public transport in a decade or so. Isn’t the issue that we probably do not want to replace it?

Deep-learning-based gene perturbation effect prediction does not yet outperform simple linear baselines - Nature Methods The analysis presented in this Brief Communication shows that, despite their complexity, current deep learning models do not outperform linear baselines in predicting gene perturbation effects, thus e...

An analysis shows that current deep learning models do not beat linear baselines in predicting gene perturbation effects, thus emphasizing the importance of further method development and evaluation. @const-ae.bsky.social @wkhuber.bsky.social @s-anders.bsky.social

www.nature.com/articles/s41...

At 17, Hannah Cairo Solved a Major Math Mystery | Quanta Magazine After finding the homeschooling life confining, the teen petitioned her way into a graduate class at Berkeley, where she ended up disproving a 40-year-old conjecture.

What a fantastic accomplishment -- and what a fantastic story! www.quantamagazine.org/at-17-hannah...

How Long Contexts Fail Taking care of your context is the key to building successful agents. Just because there’s a 1 million token context window doesn’t mean you should fill it.

www.dbreunig.com/2025/06/22/h...

Proteomic analyses identify targets, pathways, and cellular consequences of oncogenic KRAS signaling Proteomic analyses offer insights into how KRAS inhibitors affect oncogenic KRAS signaling.

Excited to share our latest work published in #ScienceSignaling! 🚀
www.science.org/doi/10.1126/...
(1/4)

This project has been in the making for quite some time. CORNETO not only integrates key concepts and methodologies in biological network inference, but also introduces a novel framework for multi-condition analysis. Congrats to the team, and especially to @pablormier.bsky.social for leading this.

Which Kind of Science Reform What hope is there for science reform, if we can't agree on what to reform? Right now, principles are more important than practices.

How can we reform science? I have some ideas. But I am not sure you’ll like them, because they don’t promise much. elevanth.org/blog/2025/07...

The HTML presentation framework | reveal.js Documentation and demos for the open source reveal.js HTML presentation framework.

revealjs.com

Design of intrinsically disordered region binding proteins Intrinsically disordered proteins and peptides play key roles in biology, but a lack of defined structures and high variability in sequence and conformational preferences have made targeting such syst...

www.science.org/doi/10.1126/...

What Has a Foundation Model Found? Using Inductive Bias to Probe for World Models Foundation models are premised on the idea that sequence prediction can uncover deeper domain understanding, much like how Kepler's predictions of planetary motion later led to the discovery of Newton...

I do not think you could invent Quantum Mechanics solely from 1900 data. And it could be very much architecture dependent : LLMs are not able to even infer Newton's laws from a massive number simulations arxiv.org/abs/2507.06952

Schmidt Fellowships for #AI in #cancer research 2025 round is now open. These are pathway to independence fellowships between @icr.ac.uk & @imperialcollegeldn.bsky.social that support fellows to be competitive for faculty jobs at the end of the 2 year funding. www.imperial.ac.uk/jobs/search-...

join us at ISMB iscb.bsky.social & @dreamchallenges.bsky.social - benchmarking foundation models w. @anshulkundaje.bsky.social @oldbap.bsky.social G. Stolovitzky @bowang87.bsky.social @mariabrbic.bsky.social K. Kalantar, J. Guinney supported by @ebi.embl.org @cziscience.bsky.social Tempus AI

But it also eliminates of the business model for search engine optimisation and web search ads, which are the primary reasons why we already have such an enormous amount of low-quality content on the internet.

Oh the LLMs have already started usin en dashes

🚨Job alert in theory/math devbio 🚨
I’m currently advertising postdoc positions to join my group. We’re building reaction-diffusion models to understand some of the crazy things embryos can do. We are based in a beautiful part of Scotland :). Please share! www.abdnjobs.co.uk/vacancy/rese...

Good news for all em-dash enthusiasts: ChatGPT now uses en dashes – with spaces, of course.

I remain sceptical about AI’s role in other aspects of science, but this strongly reinforces the idea that without an AI-first coding strategy (including appropriate training), one risks being left far behind in the coming years. www.microsoft.com/en-us/resear...

Novel Research in Biological Artificial Intelligence Expressions of interest and nominations for a Max Planck Director to lead research in the field of Biological Artificial Intelligence.

Novel Research in Biological Artificial Intelligence: Expressions of interest and nominations for a Max Planck Director to lead research in the field of Biological Artificial Intelligence @mpi-cbg.de and @csbdresden.bsky.social. More information: www.mpi-cbg.de/news-outreac...

Change of venue: fission-yeast cell-division cues actually initiate in the nucleus In yeast, the CDK and cyclin proteins that drive cell division activate first in the nucleus — a different location in the cell from where was previously thought.

In yeast, the CDK and cyclin proteins that drive cell division activate first in the nucleus — a different location in the cell from where was previously thought

https://go.nature.com/3GdXuEz

Excited to launch our AlphaGenome API goo.gle/3ZPUeFX along with the preprint goo.gle/45AkUyc describing and evaluating our latest DNA sequence model powering the API. Looking forward to seeing how scientists use it! @googledeepmind

Ah interesting points, wasn‘t aware about the bit with math formulas. As a probably below average writer l’m okay with average writing. It‘s also worthwhile to read up on the connection between essays and class/socioeconomic background and consider implications for inclusivity.

Introducing Whisper We’ve trained and are open-sourcing a neural net called Whisper that approaches human level robustness and accuracy on English speech recognition.

Audio recording + openai.com/index/whisper/ + any LLM you fancy?

More convinced of writing but still not on board with ubiquity of essay/paper writing. Very asymmetric communication and easily biased/masked by language. And I don't want to outsource the decision-finding on whom to employ to the applicant.

I’m not disagreeing with you, my point is that both models are suited to their respective fields, and there’s value in each learning from the other. What I often see in biology, though, is that ideas are communicated with a lot of flair but little underlying logic ;)

After having spent some time in the world of biology, I have to say that most maths/ml papers are pretty awful in terms of their english language. But that's fine for the field because everything is in the equations.

For me what ChatGPT is most helpful with is turning my thoughts into text that is comprehensible by others. It appears that those two processes are less entangled for me than for you, which is probably the result of our education and part the reason for our different levels of AI skepticism.

In the soft sciences, and biology hard and soft components, thinking is more fuzzy and cannot be assessed and or trained in the same way and doesn't just rely on logic but has other components.

I think this is very much a disconnect between soft and hard sciences as well as semantics what we mean by "thinking". In maths you can define problems that require logic (and thats what the hard sciences mean by thinking) with unique solutions that you can test and thereby also train idependently.