David van Dijk

David van Dijk

@vandijklab.bsky.social

Learning the rules of life. Assistant Professor of Medicine and Computer Science @ Yale

115 followers46 following51 posts

View full thread
View full thread
David van Dijk
@vandijklab.bsky.social

What's next for C2S-Scale? • True Multimodality: Integrating proteomics, epigenomics, imaging data 🖼️ • Deeper Biology: Modeling cell interactions, dynamics, & development ⏳ • Enhanced Trust: Improving interpretability & reliability ✅ • Community Tools: Building shared benchmarks & platforms 🏆

View full thread
David van Dijk
@vandijklab.bsky.social

Size matters! 📈 We observed clear scaling laws: As model size increased from 410M → 27 Billion parameters, performance consistently improved across tasks. This confirms that LLMs learn better biological representations at scale using the C2S approach. Even works with efficient LoRA tuning! 💪

David van Dijk
@vandijklab.bsky.social

Beyond standard training, we used Reinforcement Learning (RL) 🤖 to fine-tune C2S-Scale. Using GRPO + biological rewards, we specifically improved: • Perturbation prediction accuracy 🧪 • Biological Q&A relevance ❓ Aligning LLMs with biological goals! ✅

View full thread
David van Dijk
@vandijklab.bsky.social

To truly "teach" biology to LLMs, we built a massive corpus: Over 1 BILLION tokens! 📚 This wasn't just cell sentences – it included: • 🧬 50M+ cell profiles (human/mouse) • 🏷️ Annotations & Metadata • 📄 Biological Text (abstracts, etc.) Result? One model, many tasks!

David van Dijk
@vandijklab.bsky.social

And it works! 🎉 C2S-Scale achieves SOTA performance, surpassing specialized single-cell models AND general LLMs: • 🎯 Cell type annotation • 🧪 Predicting perturbation responses • ✍️ Generating dataset summaries from cells • 🗺️ Inferring spatial relationships • ❓ Answering complex biological questions

David van Dijk
@vandijklab.bsky.social

We enable LLMs to "read" biology via Cell2Sentence (C2S) 🧬➡️📝: ranking genes creates text. This lets us leverage massive pre-trained models, unifying transcriptomic data with biological text (annotations, papers) for richer understanding.

View full thread
View full thread
David van Dijk
@vandijklab.bsky.social

This work highlights the power of CLM-based intelligent adaptive solvers for scalable operator learning of dynamical systems. Imagine more efficient and accurate simulations for everything from fluid dynamics to climate modeling! 🌍

David van Dijk
@vandijklab.bsky.social

Current ML methods for PDEs often use fixed time steps, which is inefficient, especially for complex dynamics. COAST, powered by a causal language model (CLM), predicts both the solution AND the optimal time step. 🧠

David van Dijk
@vandijklab.bsky.social

🔑 Key finding: COAST generates variable step sizes that intelligently adapt to the current system's complexity! Smaller steps in complex regions, larger steps in simpler ones. Across systems, more complex dynamics get finer time resolution.

David van Dijk reposted

Kendall Reid
@kendallreid.bsky.social

1 billion cells feels like a hype number. I’d much rather see a fraction of that with higher quality and much more variation.

10x Genomics@10xgenomics.bsky.social

Creating the largest #singlecell dataset ever will take dedication, collaboration & high-performance tools. We’re thrilled the @chanzuckerberg.bsky.social Billion Cells Project will pair GEM-X tech with Ultima Genomics sequencing to achieve this landmark scale. Learn more: bit.ly/4aRaScr

David van Dijk reposted

Nick Banovich
@nebanovich.bsky.social

Excited to post my first #skySplain about our recent work published yesterday in Nature Genetics! www.nature.com/articles/s41.... One of the first authors of this study - Annika Vannan – actually wrote this breakdown, but she's not yet over here on bluesky and asked me to post!

nature.comSpatial transcriptomics identifies molecular niche dysregulation associated with distal lung remodeling in pulmonary fibrosis - Nature GeneticsXenium spatial transcriptomic profiling of pulmonary fibrosis characterizes cell composition dynamics and histopathological features associated with the disease.

David van Dijk reposted

Eric Topol
@erictopol.bsky.social

New: The largest medical A.I. randomized controlled trial yet performed, enrolling >100,000 women undergoing mammography screening The use of AI led to 29% higher detection of cancer, no increase of false positives, and reduced workload compared with radiologists w/o AI thelancet.com/journals/lan...

thelancet.comScreening performance and characteristics of breast cancer detected in the Mammography Screening with Artificial Intelligence trial (MASAI): a randomised, controlled, parallel-group, non-inferiority, ...The findings suggest that AI contributes to the early detection of clinically relevant breast cancer and reduces screen-reading workload without increasing false positives.

David van Dijk reposted

Jacob Schreiber
@jmschreiber91.bsky.social

Very cool work showing the promise of integrating several sources of data to directly address human health conditions. Synchronized long-read genome, methylome, epigenome and transcriptome profiling resolve a Mendelian condition www.nature.com/articles/s41...

nature.comSynchronized long-read genome, methylome, epigenome and transcriptome profiling resolve a Mendelian condition - Nature GeneticsSimultaneous profiling of the genome, methylome, epigenome and transcriptome using single-molecule chromatin fiber sequencing and multiplexed arrays isoform sequencing identifies the genetic and molec...
David van Dijk
@vandijklab.bsky.social

We do see strong scaling laws in single cell foundation models

Jacob Schreiber@jmschreiber91.bsky.social

Genomic Foundationless Models: Pretraining Does Not Promise Performance I've long believed genomic foundation models are not as useful as claimed. In my mind, there isn't enough training data to justify their size. Interesting to see more work in this direction. www.biorxiv.org/content/10.1...

biorxiv.orgGenomic Foundationless Models: Pretraining Does Not Promise PerformanceThe success of Large Language Models has inspired the development of Genomic Foundation Models (GFMs) through similar pretraining techniques. However, the relationship between pretraining performance ...

David van Dijk reposted

James Zou
@jameszou.bsky.social

Excited to share our new study of #genomics🧬, #EHR📈 + treatment outcomes of 78K patients across 20 #cancers! We identified >700 mutations predicting which drugs💊 are effective for individual patients. Our #ML model predicts who responds well to immunotherapies. #precisioncancer

David van Dijk reposted

Anshul Kundaje
@anshulkundaje.bsky.social

Our ChromBPNet preprint out! www.biorxiv.org/content/10.1... Huge congrats to Anusri! This was quite a slog (for both of us) but we r very proud of this one! It is a long read but worth it IMHO. Methods r in the supp. materials. Bluetorial coming soon below 1/

biorxiv.orgChromBPNet: bias factorized, base-resolution deep learning models of chromatin accessibility reveal cis-regulatory sequence syntax, transcription factor footprints and regulatory variantsDespite extensive mapping of cis-regulatory elements (cREs) across cellular contexts with chromatin accessibility assays, the sequence syntax and genetic variants that regulate transcription factor (T...
Anshul Kundaje@anshulkundaje.bsky.social

Aiming to release our long awaited ChromBPNet preprint by early next week as well. I'm recovering from back to back infections for the past 6 weeks. But we're almost there. We have some solid variant prediction benchmarks in there against large supervised models and lots more. Stay tuned.

David van Dijk reposted

Steven Sloan
@sloanlab.bsky.social

First 🦋 post! Very excited to share our lab's latest work exploring the shared and divergent aspects of human astrocyte development and glioblastoma. This effort spans many fields from developmental #glial biology to #stem cells and #tumor biology. rdcu.be/d5ADz A 🧵

rdcu.beMapping the developmental trajectory of human astrocytes reveals divergence in glioblastomaNature Cell Biology - Sojka et al. analyse the transcriptomic and epigenomic landscape of human astrocyte maturation and identify an epigenetically regulated intermediate state associated with...

David van Dijk reposted

Caleb Lareau
@caleblareau.bsky.social

Out today in @naturegenet.bsky.social -- PERFF-seq! With @tsionabay.bsky.social , @ronanchaligne.bsky.social, Bob Stickels, Meril Takizawa, + Ansu Satpathy, we describe this new assay to study rare populations with programmable nucleic acid cytometry. 1/n www.nature.com/articles/s41...

nature.comTranscript-specific enrichment enables profiling of rare cell states via single-cell RNA sequencing - Nature GeneticsProgrammable Enrichment via RNA FlowFISH by sequencing (PERFF-seq) isolates rare cells based on RNA marker transcripts for single-cell RNA sequencing profiling of complex tissues, with applicability t...

David van Dijk reposted

Arc Institute
@arcinstitute.org

Seeking pioneering scientists in immunology, neuroscience, and machine learning! Join Arc Institute as a Core Investigator + Stanford Bioengineering as Associate/Full Professor. Full lab funding and cutting-edge facilities in Palo Alto. Apply by Jan 15 for full consideration: shorturl.at/CuH1K

arcinstitute.orgAssociate or Full Professor Rank Search to join the Arc Institute as a Core Investigator and the Stanford Department of Bioengineering | Arc InstituteArc Institute is a independent nonprofit research organization headquartered in Palo Alto, California.

You are up to date.