Sign in

Fabian Fröhlich

@frohlichlab.com
1.1K followers 409 following 135 posts

Dynamics of Living Systems (www.frohlichlab.com) group leader @thecrick.bsky.social Understanding signalling & cell state dynamics through mathematical modelling and machine learning.

PostsRepliesMedia
Reposted by Fabian Fröhlich
The Francis Crick Institute @crick.ac.uk · 29/09/2026
Our AI for Biological Discovery online seminar series continues with “Communities of AI agents for scientific discoveries” by James Zou, Stanford University, on 7 October at 15:00 BST. Register now for the webinar: www.eventbrite.co.uk/e/ai-in-biol...
eventbrite.co.uk
AI in Biology Seminar | Professor James Zou
A new AI for Biological Discovery Seminar Series, presented by the Francis Crick Institute.
164
Fabian Fröhlich @frohlichlab.com · 29/09/2026
1) we are in an attention economy, if I need to verify I can also /goal /loop it myself. 2) where it gets interesting is when you have multiple models to predict the same Y, because you wouldn’t necessarily expect them to have the same failure mode
100
Reposted by Fabian Fröhlich
The Francis Crick Institute @crick.ac.uk · 28/09/2026
Applications are open for our new AI/ML clinical fellowships, combining formal MSc-level training in AI and machine learning with PhD or postdoctoral research at the Crick. Find out more about the programme and apply by 11 November: bit.ly/4hxmAgn
crick.ac.uk
AI/ML clinical fellows
Our clinical training pathway integrating AI and machine learning with discovery research.
022
Fabian Fröhlich @frohlichlab.com · 28/09/2026
Mechanistic World Models for biology are (almost) here: www.biorxiv.org/content/10.6...
biorxiv.org
GEMOT: Towards Mechanistic World Models for Biology
Scientific discovery seeks mechanisms that explain observations and predict beyond the measurements that produced them. Whereas large language models (LLMs) encode knowledge implicitly, mechanistic wo...
020
Fabian Fröhlich @frohlichlab.com · 25/09/2026
assuming this is connected to arxiv.org/abs/2609.253...
arxiv.org
Scoring Grant Applications with Large Language Models
Purpose: Assessing grant applications is time-consuming and difficult, adding to the overall burden of academic peer review. Whilst funders are exploring whether AI can help, there is no published res...
120
Fabian Fröhlich @frohlichlab.com · 25/09/2026
I cannot endorse what UKRI did here but at least they used (presumably local) open weight models so they respected privacy and inputs cannot be used for training.
110
Reposted by Fabian Fröhlich
Andrew Krause @blindmath.bsky.social · 24/09/2026
"What are we even trying to do here?" Is what I have been asking many people in academia more widely lately. It is long past time for us, across *all* fields, to rethink the faux-meritocracy of numerical metrics (# of papers, citations, etc). We really should decide and reward what we truly value.
031
Reposted by Fabian Fröhlich
Chris Tape @christophertape.bsky.social · 24/09/2026
PhD opportunity in our lab at @ucl.ac.uk! 🧬 How do colorectal tumours spatially adapt to treatment? Study cancer cell plasticity using organoids, single-cell, & spatial biology. 4 years | Jan 2027 (flexible) | Stipend + UK home fees ucl.ac.uk/work-at-ucl/... Please share!
072
Reposted by Fabian Fröhlich
The Francis Crick Institute @crick.ac.uk · 23/09/2026
Discover the latest issue of The Crick magazine, featuring stories on cancer, antimicrobial resistance, life after a Nobel Prize, liver health, heat and the human body, and more from across the Crick. Read now: www.crick.ac.uk/news-and-features/t…
084
Reposted by Fabian Fröhlich
Richard Sever @richardsever.bsky.social · 20/09/2026
"All corners of academia will soon be flooded with AI-generated research....Maths and CS go first because their entire output is symbol manipulation...Biology comes later not because biologists are cleverer but because the rate-limiting step is physical" giorgio.gilest.ro/the-weimar-o...
giorgio.gilest.ro
The Weimar of knowledge – la bottega
56640
Fabian Fröhlich @frohlichlab.com · 20/09/2026
I wish I could always get the same answer from LLMs by simply setting the temperature to 0 ...
011
Fabian Fröhlich @frohlichlab.com · 18/09/2026
Or we have to come to terms with the idea that a contribution that can be replicated by a student with AI in a matter of days isn't really a meaningful contribution anymore ...
000
Fabian Fröhlich @frohlichlab.com · 17/09/2026
I think the supplement is the place where we will make AI reading AI generated text _really_ happen
010
Fabian Fröhlich @frohlichlab.com · 14/09/2026
The general attitude of our community towards a technology that can [...] suggests that these operationalizations of our values are at best incomplete.
010
Fabian Fröhlich @frohlichlab.com · 13/09/2026
Yes, and the way this has been historically been addressed is regulatory capture (see above), vertical integration (discounts for science, enterprise deals, ...) and diversification (will anthropic turn into a pharma company?).
020
Fabian Fröhlich @frohlichlab.com · 13/09/2026
I would expect them to have realised that we‘ve lost control of AI as a technology, not as tool. Reminds me of the passage below from Against the Day (replace silver with AI). We‘ve been on this trajectory for a while I don‘t see us changing course.
000
Fabian Fröhlich @frohlichlab.com · 13/09/2026
How much of this is genuine risk vs regulatory capture though? Wouldn‘t be a first that manipulating public perception is used to stifle competition/innovation (en.wikipedia.org/wiki/War_of_...).
en.wikipedia.org
War of the currents - Wikipedia
240
Reposted by Fabian Fröhlich
Terence Tao @teorth.bsky.social · 11/09/2026
A group of 25 Fields Medalists, including myself, have made a joint declaration on Math and AI: mathandai.org . We welcome additional signatories. See also this article in the Economist announcing the declaration: www.economist.com/science-and-...
mathandai.org
Declaration — Math and AI
Read the declaration and add your name.
422052926
Fabian Fröhlich @frohlichlab.com · 11/09/2026
in other words, more worried about humans being stupid than AI being smart …
020
Fabian Fröhlich @frohlichlab.com · 11/09/2026
They forgot to turn off the internet though … www.anthropic.com/research/ali...
anthropic.com
An alignment assessment of recent cybersecurity incidents
We present an alignment assessment of four incidents in which Claude models gained unauthorized access to real third-party systems.
131
Fabian Fröhlich @frohlichlab.com · 10/09/2026
Right, but its presumably cheaper and can be turned off.
010
Reposted by Fabian Fröhlich
Joshua Bull @joshuabull.bsky.social · 10/09/2026
I'm really excited to be heading to the Crick to start the Quantitative Spatial Biology lab! We'll be doing all things #spatialbiology, mathematical modelling, and AI, to get a deep understanding of the biological mechanisms underlying how spatial structures form in health and disease #newPI 🔬🧬🧪🧮
091
Fabian Fröhlich @frohlichlab.com · 10/09/2026
Replace them with AI
100
Reposted by Fabian Fröhlich
Ruben Perez-Carrasco @perez-carrasco.bsky.social · 09/09/2026
The new AC-DC paper is out! ⚡ One small gene circuit, more than 30 distinct dynamical behaviours. Showing how small regulatory networks can hide a surprising amount of dynamical complexity. www.cell.com/cell-systems...
cell.com
Let there be multifunctionality: Uncovering the dynamical zoo of the AC-DC genetic circuit
Small gene regulatory circuits can encode a surprising diversity of dynamical behaviors. A biophysical model of the AC-DC motif reveals over thirty distinct bifurcation structures and shows how molecu...
0134
Reposted by Fabian Fröhlich
James Pirruccello 🎃🦇🧹 @jamespirruccello.com · 06/09/2026
openai.com/index/resear...
openai.com
Research acceleration: The view inside OpenAI
Inside OpenAI, coding agents are reshaping AI research. Explore early data on agent usage, experiment velocity, task complexity, and research acceleration.
152
Reposted by Fabian Fröhlich
Nozomu Yachie @nzmyachie.bsky.social · 04/09/2026
Maybe it’s time to say goodbye to UMAP and other low-dimensional embeddings (at least in many single-cell analyses). We developed MILK to encode single-cell populations into tree representations at unprecedented scales. www.biorxiv.org/content/10.6... Here’s what MILK is: 🧵
14425
Reposted by Fabian Fröhlich
Nature Biotechnology @natbiotech.nature.com · 19/08/2026
A blinded, prospective benchmark of in silico antibody discovery anchored to experimental affinity and developability www.nature.com/articles/s41...
nature.com
A blinded, prospective benchmark of in silico antibody discovery anchored to experimental affinity and developability - Nature Biotechnology
A prospective, blinded competition is conducted for artificial-intelligence-generated and optimized antibody discovery.
282
Fabian Fröhlich @frohlichlab.com · 27/08/2026
🧬 One cell whispers code 🔗 neural nets meet ODEs 🔬 come decode with us
crick.wd3.myworkdayjobs.com
Postdoctoral Fellow - Frohlich Lab
Salary for this Role: From £47,500, with benefits, subject to skills and experience. Job Title: Postdoctoral Fellow - Frohlich Lab Reports to: Fabian Frohlich Closing Date: 24/Sept/2026 23.59 GMT Job ...
000
Reposted by Fabian Fröhlich
Reese Richardson @reeserichardson.bsky.social · 25/08/2026
A massive update: At least ~15 companies~ are selling scientists antibodies using faked validation data. We've documented 18,000+ manipulated images on 17,000+ products sold by leading laboratory suppliers including Thermo Fisher, Abcam, Santa Cruz Biotechnology, Millipore Sigma and Bio-Techne. 1/🧵
18454329
Reposted by Fabian Fröhlich
The Francis Crick Institute @crick.ac.uk · 24/08/2026
Join us for our first online AI for Biological Discovery Seminar, where our Chief AI Scientist Mihaela van der Schaar will host Caroline Uhler to discuss her research at the intersection of machine learning, statistics and genomics. Learn more and book your free ticket: bit.ly/4xAvNLL
bit.ly
AI in Biology Seminar | Professor Caroline Uhler
A new AI for Biological Discovery Seminar Series, presented by the Francis Crick Institute.
0117
Fabian Fröhlich @frohlichlab.com · 14/08/2026
well you can have claude write a biorxiv scraper on your phone while taking a dump and reasonably presume everything is fine because you're the only one doing it, right?
251
Reposted by Fabian Fröhlich
derek guy @dieworkwear.bsky.social · 13/08/2026
I enjoyed this. ordinaryabundance.com
ordinaryabundance.com
Ordinary Abundance
A walk through a modern apartment, through the eyes of the people for whom everything in it was new.
753423907
Fabian Fröhlich @frohlichlab.com · 13/08/2026
Curriculum learning is good, curriculum multiple shooting is better: new preprint on how to train ODE (including UDE and NODE) models: arxiv.org/abs/2608.05777
arxiv.org
Curriculum Multiple Shooting for Robust Training of Neural and Universal Differential Equations
Neural ordinary differential equations (NODEs) and universal differential equations (UDEs) provide flexible and popular frameworks for learning interpretable dynamical systems from noisy time-series d...
010
Reposted by Fabian Fröhlich
Daniel Litt @littmath.bsky.social · 11/08/2026
Yesterday I gave a talk at an OpenAI summit on the future of mathematics. I was asked to talk about the bad future, where humans become mathematically disempowered. Here's a blog post based on the talk: www.daniellitt.com/blog/2026/8/...
daniellitt.com
The End of Mathematics — Daniel Litt
126820
Reposted by Fabian Fröhlich
David P. Reichert @david-p-reichert.bsky.social · 03/08/2026
Many don't see how much AI has changed recently. Coding agents aren't just chatbots. We can’t deal with the challenges of AI if we don’t understand it. Here's a post to provide some evidence (Claude doing small ML experiments) and form intuitions. davidpreichert.substack.com/p/if-you-hav... 🧵..
davidpreichert.substack.com
If you haven’t recently used Claude Code*, you might not understand where AI is at
A report on a series of mini machine learning projects, executed with, and mostly by, Claude
98912
Reposted by Fabian Fröhlich
National Trust @nationaltrust.org.uk · 12/08/2026
ANNOUNCEMENT: All our sundials will be out of order this evening between about 6pm and 8pm, we apologise for any inconvenience this may cause. Normal service will resume tomorrow. In unrelated news, enjoy the eclipse later... Photos: Mike Selby, Annapurna Mellor, Chris Davies, John Millar
A wall-mounted sundial inside a walled garden. Text on the sundial reads: "fear god, obey ye king"A traditional sundial on a plinth, formal gardens surround itThe sun beams down on a traditional sundial, with heather visible in the backgroundA shot of a sundial from above. It's mounted on a stone plinth. Text on the side reads "KEW BRIDGE"
201712458
Fabian Fröhlich @frohlichlab.com · 12/08/2026
I would also push back on the proposition that large, non-identifiable models cannot be interpreted causally; see doi.org/10.1016/j.ce... and doi.org/10.15252/msb....
doi.org
Redirecting
010
Fabian Fröhlich @frohlichlab.com · 12/08/2026
More generally, unless a simplification follows from first principles or from datasets spanning large families of biological backgrounds (species, genotypes, and so on), identifiability is almost always bought with generalisability. I know of very few datasets with that coverage.
110
Fabian Fröhlich @frohlichlab.com · 12/08/2026
I think the argument I developed with the toy model applies equally to the cell cycle example in the paper: any intervention that materially affects the operation of the PP2A–ENSA–GWL module cannot be represented by the simplified model, and that is the price paid for identifiability.
100
Fabian Fröhlich @frohlichlab.com · 12/08/2026
I appreciate the point, and the approach is interesting, but what it establishes is causal necessity within the model, not in the system it represents—and the two should not be conflated. There is also no guarantee that the simplified model is unique, and with it the causal structure it implies.
110
Fabian Fröhlich @frohlichlab.com · 11/08/2026
If a biological system actually implements three mechanisms controlling a phenotype, each dominant to a different degree across genetic backgrounds, and the data are only sufficient to pin down one of them, the answer should not be to simplify the model but to collect more data.
120
Fabian Fröhlich @frohlichlab.com · 11/08/2026
However, there is no guarantee that a simplified model tells you anything about the real world, and you risk exactly what you caution against at the outset: a model that only reproduces its input data, whose interpretability is an artefact of the structure rather than a reflection of the biology.
140
Reposted by Fabian Fröhlich
Ezequiel Galpern @eag91.bsky.social · 05/08/2026
1/ New preprint! with @solersanchisx.bsky.social @cwjpugh.bsky.social @federicobilleci.bsky.social @jonnyfrazer.bsky.social and Mafalda Dias, we introduce a scalable framework to improve stability prediction and separate folding from functional constraints. www.biorxiv.org/content/10.6...
biorxiv.org
Blending physics-based and inverse folding models to disentangle variant effects on stability and function
Protein sequences are constrained not only by the need to fold into stable structures, but also by specific functional requirements imposed by natural selection. Yet predictions of how amino-acid chan...
1139
Fabian Fröhlich @frohlichlab.com · 05/08/2026
Huge thanks to @giacomofabrini.bsky.social who saw this was doable and built it, @erc.europa.eu for the funding and @crick.ac.uk for an environment where work like this can happen.
010
Fabian Fröhlich @frohlichlab.com · 05/08/2026
Data-driven integration calls breast cancer subtype important; supply EGFR levels to the ODE instead and the subtype signal vanishes. Two explanations with same predictive power so the data cannot arbitrate between a molecular and a systems lens. DMMs surface the ambiguity and let us be the judge.
120
Fabian Fröhlich @frohlichlab.com · 05/08/2026
Per usual, mechanistic models stay informative where they fail: DMMs systematically underpredict phospho-ERK under MEK inhibition, which could reflect unmodelled crosstalk with AMPK signalling.
110
Fabian Fröhlich @frohlichlab.com · 05/08/2026
The unexpected result: across 63 mammary cell lines, most heterogeneity sits at the inputs and outputs rather than in the core machinery. Baseline ERBB2 activation and ERK-to-RSK gain emerge as the major axes, likely set by endocytic, cytoskeletal and calcium programmes.
110
Fabian Fröhlich @frohlichlab.com · 05/08/2026
DMMs couple semi-supervised representation learning to an ODE model of EGFR/MAPK signalling, end-to-end. The encoder proposes cell-line-specific parameters; the ODE model tests them against dynamic perturbation data. Representation and mechanism constrain each other rather than sitting side by side.
110
Fabian Fröhlich @frohlichlab.com · 05/08/2026
Mechanistic ODE models encode pathway topology but are blind to anything outside the pathway. Representation learning compresses data into low-dimensional embeddings but says nothing about mechanism. Deep Mechanistic Models (DMMs) are our attempt at both in one model www.biorxiv.org/content/10.6...
biorxiv.org
Deep Mechanistic Models reveal pathway-extrinsic drivers of mammary MAPK signalling heterogeneity
Cells sense and respond to their environment through signalling pathways, and the dynamics of these pathways shape cell fate even within genetically identical populations. Two largely separate computa...
1101
Fabian Fröhlich @frohlichlab.com · 04/08/2026
The result we didn't expect: most heterogeneity across the 63 mammary cell lines manifests at the level of inputs and outputs but isn’t generated by core machinery. Tune MAPK model parameters and you'll get a fit, but you may be crediting the cascade for variation that lives beside it.
000