Just presented our new multimodal histopathology method "SpotWhisperer" at ICML, one of the largest AI conference. SpotWhisperer enables spatially resolved annotation of histopathology images using natural language. We achieved this by "transferring" annotations from transcriptomic data. More soon!

After an exciting 3.5 years at CeMM and the AI Institute, it's time for me to move on to the next adventure. I am grateful for the friendships I made and the science I got to contribute to. I also wrote a brief recap of my science and activities during that time. Enjoy! moritzs.de/cemm

Having fun at work... :) Made possible by an amazing bunch of colleagues at my institute.

Great behind-the-scenes overview on 'Flow matching and OT in cell biology' hrovatin.github.io/posts/sc_flo...

THANK YOU to everyone who joined the #March for #Science in #Vienna today! 💐💐💐 >500 people in front of @univie.ac.at, ~1000 people at Votivpark! 💚💚💚 Science is for everyone! #standupforscience2025 #scienceforall #sciencenotsilence

SPEAKER LINEUP FOR STAND UP FOR SCIENCE ☀️ Here's our list of current speakers—and it's still growing! Come tomorrow to hear some amazing speakers talk about science, their experiences, and the future. #standupforscience2025

Yes. LLMs should rather be seen as "interface modules" of more complex AI systems: Their core role will be to *translate* between AI language (bits and bytes) and human-readable language.

Visiting the #38C3 conference was an absolute blast again! Four days of great interactions and talks around politics, society, information security and science - My highlights post on LinkedIn: www.linkedin.com/posts/moritz... - I gave a talk too: lnkd.in/dXRF2mai - All talks here: media.ccc.de

I'm hiring: 1. Research associate (wet-lab w/ phd) to generate mpra perturbation data 2. ML postdoc to build multimodal generative AI for DNA (eg diffusion and LLMs) 3. Bioinformatician (any level) to process and harmonize functional genomics data to train foundation models DM me if interested!

If you've never donated to Wikipedia, now's the time. Elon Musk is attacking it because it's one of the only sources of honest, curated information that he can't control. www.newsweek.com/elon-musk-ta...

Thrilled to see our work with the @bocklab.bsky.social featured in 'The Scientist'! It's a really great read and builds on an interview I had with the @kamalnahas.bsky.social. www.the-scientist.com/an-ai-lab-pa...

Had a fantastic time last week at CSHL's #biodata24, with amazing discussions around AI and bioinformatics and presenting my work on CellWhisperer. Thanks @mike_schatz and all organizers, looking forward to being back in 2026!

How come everyone is joining Bluesky these days? I thought it dead already.

Congratulations to @findycang.bsky.social from @jhubiostat.bsky.social who did an amazing job presenting her work at #biodata24 tonight! Preprint coming soon. ☺️ #ProudPI #stats #spatial #omics #transferlearning 🧪🧬🖥️🧠📈

What if we could look at a human tissue image🔬and quantify its biological age?🧬 And how could we get the same insights without collecting the tissue sample? In our new @biorxivpreprint, we create "tissue clocks" ⏳to do just that! www.biorxiv.org/content/10.1... 📝 #aging #tissue thread below⬇️

"Contextual AI models for single-cell protein biology" 🧪🧬🖥️ - Protein embeddings are powerful. - But: they are usually cell-type agnostic - This paper: Embedding cell-type-aware protein representations via GNNs on scRNA-seq-pruned protein interaction graphs www.nature.com/articles/s41...

"Transformers need glasses! Information over-squashing in language tasks" - In typical LLMs every token only 'sees' previous tokens (causal attention) - This leads to information inbalance across tokens - Thereby, later tokens 'vanish' in long sequences arxiv.org/pdf/2406.04267

'The primitive endoderm (PrE) supports lineage plasticity to enable regulative development' 🧬️🧪 - PrE is extra-embryonic tissue (at day 4) but can still recreate a full blastocyst (i.e. also embryo) - JAK/STAT supports this plasticity - www.cell.com/cell/abstrac...

"Why is exercise good for you?" - Humans, as hunter-gatherers, evolved to 'always move' - Exercise alters gene expression throughout the body - With exercise, muscles contribute to anti-inflammation through IL-6 production www.nature.com/articles/d41... 🧬🖥️🧪

Cool study... Post-January 6th deplatforming on Twitter reduced the reach of misinformation Hats off to my UCR colleagues Kevin Esterling and Diogo Ferrari and their co-authors... 🧪sociology polisky policysky www.nature.com/articles/s41...

openreview.net/pdf?id=KIeU4... - Most drug design methods assume rigid proteins - Simple docking experiments show: Flexible side chains enhance modeling - New method FlexFlow: Flow matching of small molecules in addition to side chain configurations Looking forward to full release 🧬 & 🖥️

I go 90% with the stated opinion about OpenAI etc. Yet, it's a real pity that some arguments needed to be overdrawn. I use GPT-4 daily and find it extremely useful. Also, highly unlikely that LLMs will worsen due to data pollution. One more example of how polarisation wins the attention market..

- A common problem in ML is to generalize to cases unseen during training - Here, GPT- and protein embeddings enable perturbation effect prediction for unseen genes - My take: Great direction and I anticipate more 'embedding arithmetics' (e.g. 'gene1-gene2' for overexpression and KO)

The original 'oversquashing' paper is very accessibly written: - In GNNs, when data needs to pass many nodes, the data gets exponentially 'diluted' - Adding a simply fully-connected edge layer, improved upon QM9 by 42% - This foundational and simple concept is only from 3 years ago

"Cell subtype-specific effects of genetic variation in the Alzheimer’s disease brain" - scRNA-seq 🧬 allows high-throughput variant effect description 🖥️ - Here: brain tissue from 424 individuals, revealing unprecedented number of eQTLs - Analysis mostly promoter-driven. More Hi-C data needed?

"Understanding Biology in the Age of Artificial Intelligence" arxiv.org/abs/2403.04106 AI models are uncommon to uncover strict "laws" in biology. How are they still useful? An epistemological perspective 3-bullet-point review in the comment below

cos_sim(a, b) = dot(a_norm, b_norm) But is it always better for comparing embeddings? - No: cos_sim may yield arbitrary results - Cause: Model expression flexibility - Remedy: Train with cosine similarity, or normalize dataset *before* embedding learning

Apple's insights to training multimodal (image) LLMs: - LLaVA-like approach (fine-tune LLM on tokenized images) - Image encoder model is key. More information (image feature dimensions and tokens) = better - Augment text-only or image-interleaved documents to prevent deteriorating LLM performance

Geometrical explanation of LayerNorm [(x-μ)/σ] arxiv.org/abs/2305.025... - implicitly projects inputs orthogonal to 1-vector and scales to norm of sqrt(d) - enables transformers to attend to all keys simultaneously (sometimes useful) - ensures that attention keys are not too "small" to be selectable

www.biorxiv.org/content/10.1... - Aggregate scRNA-seq dataset (human and mouse) with annotated cell potency (i.e. totipotent, pluripotent, multipotent, unipotent). - Train a neural network to predict level of potency. - Neural network has binary weights, allowing to read out predictive genes