How many papers are truly irreproducible and how many show a possible causal path which often isn’t the limiting factor and is conditional on important context (eg mutations, genetics, environment of the experiment)? Thinking about omnigenic model, “mechanistic bias” in medicine, personalised PRS.

Hepatic acetyl-CoA metabolism modulates neuroinflammation and depression susceptibility via acetate @cp-cellmetabolism.bsky.social www.cell.com/cell-metabol...

Very impressive work and interesting modelling ideas 💡:

By training GPN-Star on vertebrate, mammal, and primate alignments, we reveal task-dependent advantages of modeling deeper versus more recent evolution. These findings offer new biological insights and practical guidance for developing future gLMs and evolutionary models. (6/n)

This is truly an incredible breakthrough IMO. Really exemplifies what you get when deep domain expertise (popgen/evolution/disease genetics in this case) fuses with cleverly crafted ML. What u get r sleek, well thought out architectures that absolutely destroy the behemoths. Wow!! 1/

If you feel like you need noise canceling headphones 🎧 - you probably actually need to buy them right now. Don’t delay this decision and loose time trying to operate without them. Noise cancelling headphones are like your work email+slack - you may be able to work without them but it’s pretty hard.

Has anyone ever seen a house/flat in the UK/Cambridge that had an HRV or ERV system? If you can’t have windows open at night due to security concerns how are you supposed to ventilate? I tried closing windows and even with the fan on max CO2 levels never got below 850ppm (bad for sleep quality).

With unprecedented 38× coverage, our atlas resolves single neuron clusters and shows that separate waves of neurogenesis use different modes of molecular identity encoding: discrete in early born and continuous in late born. 3/8

Neuronal diversity is written in transcriptional codes 🧬. But what is the logic of these codes that define cell types and wiring patterns? To find out we built a #scRNAseq developmental atlas of the Drosophila nerve cord and linked it to the #connectome 🪰🧠 #preprint thread ⬇️1/8

And he made it again!...tour the force from JIa-Wei´s Wang lab using comparative sc-transcriptomes, and developing new analytical tools, to tackle evolutionary trajectories of plant cell types!. Amazing work as usual! www.cell.com/cell/abstrac...

Ended up coming back to these points to articulate state of the GRN field in the paper/fellowship introduction. Introductions need to be specific about the task specification and model-task mismatch in existing work. Task specification need to match biology terms at the right level of abstraction.

👀 we are involved in projects 1 & 6 on the list: 1️⃣ Multicellular molecular characterisation of inflammatory bowel disease 6️⃣ Designing Efficient Single-Cell Perturbation Experiments with Lab-in-the-Loop AI Agents 📅 Application deadline is 30th September ❗

Welcome @timcoorens.bsky.social ‪ ‬🇳🇱, our new Research Group Leader. Find out how Tim’s group is exploring using large-scale single-cell and spatial data to trace cell lineages, understand cancer origins, and uncover how mutations drive disease. www.ebi.ac.uk/about/news/p... 🖥️🧬

Looks like this platform became almost as unusable as Twitter/X. I rarely open it. If you want to reach out - you can use slack or email or maybe LinkedIn or for code issues - GitHub Issues. Please don’t email about the code issues - GitHub Issues is much more useful for managing them.

Biologist folk (especially in evolutionary biology and/or ecology, but it don’t matter): Can you give me your favorite examples of trade offs in biology? Organism or system don’t matter. Primary literature or reviews preferred.

I think small focused fully remote research conferences should exist. You can ask participants to record 1, 5, 20 & 60 min talks that others can watch in advance, then provide a structure for scheduling meetings with the authors for questions. You can even just email the authors to ask for a call.

How do enhancer-gene contacts relate to transcription activation? In mouse embryonic stem cells, a combo of STED, oligoFISH & RNA FISH shows that 3% of Dppa3 alleles transcribe. For these, enhancers and promoters come closer (panel E). Stumberger ... Leonhardt & Harz www.biorxiv.org/content/10.1...

If you are interested in peptide-based protein binders, check out this new @biorxivpreprint.bsky.social by the group of @sebastianpomplun.bsky.social of our @led3hub.bsky.social on their computer-based design.

Unpopular opinion. You can’t expect scientists to spend time on code engineering before seeing any results that illustrate why spending 6 months in human hours on engineering is warranted. Same re coming up with optimal ways to work. Those things in academia need to be supported by dedicated teams.

Are world models necessary to achieve human-level agents, or is there a model-free short-cut? Our new #ICML2025 paper tackles this question from first principles, and finds a surprising answer, agents _are_ world models… 🧵 arxiv.org/abs/2506.01622

Hakime, as part of @elsa-ai.eu, is organizing a dedicated Health Privacy Session which will feature a keynote talk, a panel discussion, and the recap of the Health Privacy Challenge with contributed presentations from the challenge participants. #CAMDA 📅 July 24 | 🕑 11:20-15:00 | 📍 Room 01C

John will share how to achieve higher throughput and fidelity in screening via flexible, high-accuracy barcode decoding. He presents a principled, information theory-based software solution that unifies diverse, demanding decoding tasks within one robust framework. #HiTSeq 📅 July 23 | 📍 Poster C-192

@elihei.bsky.social is gonna present our new tool segger, developed in collaboration with Moritz Gerstung and @danapeer.bsky.social, for fast and accurate segmentation of spatial transcriptomics using attention-based graph neural networks. #MLCSB 📅 July 22 | 🕑 15:50 | 📍 Room 01A

@thbec.bsky.social is going to share preliminary results on Meta-DeepRVAT, a new approach for deep learning based meta-analysis improving the power of rare variant association studies using population scale external control cohorts. #MLCSB 📅 July 21 |📍 Poster A-312

Excited to present our ongoing work @steglelab.bsky.social @sangerinstitute.bsky.social @bayraktarlab.bsky.social @mhaniffa.bsky.social @teichlab.bsky.social on whole embryo modeling of how TFs interact to drive specific CRE element activation in human organogenesis. #RegSys Wed 23 July, 14:40, 11BC

@vitaliikl.bsky.social will discuss how biophysical deep learning improves understanding of embryogenesis. He applied cell2state, his new method leveraging TF-motif interactions and biophysical constraints, on an unpublished multimodal sc and spatial atlas. #RegSys 📅 July 23 | 🕑 14:40 | 📍 Room 11BC

🚨 Join us tomorrow for this really exciting talk on the interplay between morphogen signalling and tissue organization! 🚨

German food hospitality is pure incompetence. That’s the tweet. I don’t think I ever had a positive experience.

Cambridge busses are useless. Scheduled 2 minutes before trains arrive not after. Drivers stop where they want not on the stops. Taxi or driving is the only way to get around. @stagecoachwales.bsky.social * But what about 🚲 ? Cycling is extremely dangerous for people without advanced experience.

Night science: semi overlapping clouds ⛅️ 🌧️ ☁️ of ideas 💡 Day science: PyTorch computation graph Challenge is going from imprecise cloud ☁️ to precisely connecting all project parts into a process that results in paper figures and intermediate results.

@iscb.bsky.social 2025 morning yoga sessions are a nice touch. I hope I can wake up in time! Looking forward to presenting our work modelling how TFs specify genomic DNA accessibility state in every cell - as cell types develop in organogenesis (cell2state model + multiome of >1000 populations).

New work on dynamics of morphogen signaling: Cells decode BMP gradient via temporal integration of signaling level, not instantaneous thresholds. While GRN gates response to provide additional spatial input Morphogen signalling dynamics + GRN for tissue patterning www.biorxiv.org/content/10.1...

New method for reducing high-dimensional neuronal responses in visual cortex to low-d latents in 🐁 BUT while this works with simple stimuli eg drifting gratings, natural image responses cannot be similarly reduced... 👀🧠 #neuroskyence www.biorxiv.org/content/10.1...

The "rubber hand illusion" in mice... @lucestebanez.bsky.social &co use automated videography to show that mice display quantifiable behavioral markers of the embodiment of an #ArtificialLimb, opening the door to future research into human #BodyOwnership disorders @plosbiology.org 🧪 plos.io/4jHsESn

Fine-tuning improved PromoterAI’s ability to predict the direction of motif effects — a known issue of multitask models. The model often recognized motifs before fine-tuning, but got the direction wrong. After fine-tuning, its predictions aligned better with the data. (8/)

Excited to share my first contribution here at Illumina! We developed PromoterAI, a deep neural network that accurately identifies non-coding promoter variants that disrupt gene expression.🧵 (1/)

Specialised models are a part of science programme of the field - steps on a path to deeper understanding - not software you make & sell. No effort towards understanding -> the models & tasks they solve are not assessed for how well they match the science programme -> recipe for wrong & not useful.

Cell communication is one of the most complex problems in biology - which requires very precise formulation of tasks. Eg cell density prediction given … with 10s of components listed under “given …”. Missing one of these 10s can mean big divergence between the task and meaning of biology terms.

It’s quite hard to systematise where models can go wrong but here is an attempt: 1. Wrong math 2. Wrong mapping of math to code (or vice versa) 3. Correct math and code but wrong interpretation given in text: code solves one task but paper presents it as another task 4. Poorly understood tasks 5. …

This phrase was coined in 1976 when the models were a lot simpler than the models published now. Linear mixed models and PCA can be described this way because as general methods they are wrong in less complex ways. Specialised models can be wrong in quite misleading ways as well as useless.

Re all models are wrong but some are useful: Useful models are designed with a good match between the model architecture and the science question aka ML task. Big gap between the model and what biologists mean with the cell comm questions -> both useless and likely wrong model.

The last part of @paulknabl.bsky.social 's PhD is now available at www.biorxiv.org/content/10.1... Paul showed that BMP signaling is active in the diffuse nervous system of the sea anemone Nematostella and the box jellyfish Tripedalia. Moreover, suppression of BMP signaling in Nematostella leads...

The firing of neural populations is high-dim even if their subthreshold activity is low-dim! This work by @bio-emergent.bsky.social and @haydari.bsky.social shows how, with a solvable model, a data analysis technique, and data from mouse visual cortex: www.biorxiv.org/content/10.1...

Nice work from Du & Warmflash using micropatterned hESCs to dissect midbrain-hindbrain boundary formation Timing of WNT signalling, not strength or duration, specifies midbrain-hindbrain boundary Good example of recapitulating developmental signalling in vitro www.biorxiv.org/content/10.1...

A regular reminder that 55um Visium spots/locations are not cells. Cells need to be isolated as individual cells physically or by image segmentation. “Cell communication” between Visium spots is like fusing two power plug 🔌 pins, inserting into the socket & expecting your device to still work.

We currently have a call for support that has gone out to European labs, to support FlyBase-UK. We are asking our colleagues from labs in the US and other countries to wait for a similar call to them that will go out in the near future, to support the US sites. We thank you for your patience.

URGENT: FlyBase has lost practically all its funding overnight; even user fees are tied up in denied grant funding. 🤬🤯 Any lab using @flybase.bsky.social please donate using the link in post below. This incredible community, on whose backs our #Drosophila labs depend, can't be left out to dry.

Social media is like recurrent intrusive thoughts but generated outside of your brain and algorithmically selected to be more recurrent and more intrusive.

Awesome paper. Simple post-hoc trick (averaging over homologs) with elegant evolution theory dramatically improves zero shot coding variant effect prediction in pLMs & actually delivers better results from the larger models (inverting the trend with raw likelihoods). 1/