Science Shutter Time

Science Shutter Time

By Vivekanandan RamalingamScienceLife Sciences
Download on the App Store

Science Shutter Time episodes

  • From avocado chairs to chromatin states: diffusion models for single-molecule footprinting (SMFNet)

    The diffusion model architecture that gave avocado chairs can also generate chromatin states of DNA molecules. Nate Diamant (PhD student, Computer Science, Kundaje lab, Stanford) and Till Schwämmle (postdoc, Krebs lab, EMBL and Stanford) talk about SMFNet, a masked discrete diffusion model that, given only a DNA sequence, generates realistic single-molecule footprinting reads: which sites on each individual molecule are open, bound by a transcription factor, or wrapped in a nucleosome. Sequence-to-function models usually predict one average signal, but the same average can hide very different mixtures of molecular states. They explain how single-molecule footprinting reads those states, how SMFNet is trained and samples a molecule by revealing it in steps, what it shows about co-binding and substates at enhancers and promoters, how it compares with the companion models ChAI-Net and Mola, and where the field goes next: more cell types, kinetics, and designing sequences for a target distribution of states.Nate Diamant: https://nathanieldiamant.com (@ndiamant on X and Bluesky: https://x.com/ndiamant)

    Till Schwämmle: https://www.embl.org/groups/krebs/members/

    Kundaje lab: https://kundajelab.stanford.edu

    Krebs lab (EMBL): https://www.embl.org/groups/krebs/

    Preprint: https://www.biorxiv.org/content/10.64898/2026.10.30.762954v1Code: https://github.com/kundajelab/SMFNet (paper analysis code: https://github.com/kundajelab/SMFNet_paper_analysis_code)Data and model: https://huggingface.co/datasets/kundajelab/smfnet-data-ont-mesc and https://huggingface.co/kundajelab/smfnet-ont-mescChapters00:00 Trailer01:06 Why use diffusion models for gene regulation?04:22 One average, many molecular states10:10 How single-molecule footprinting works13:39 SMF vs ATAC-seq19:44 What distributions mean21:51 How SMFNet works29:49 Unmasking order32:27 What it reveals: co-binding, promoter states37:06 Histone marks and footprints on one molecule39:18 Companion models: ChAI-Net and Mola44:08 Bigger models, more cell types, more data52:46 Too fast to footprint57:06 Sequences for a state distribution

    1 hr 3 min
  • Can AI help regrow an arm? Regeneration, exploding cells and gene regulatory grammar

    Cut a planarian into pieces and each piece grows back into a whole worm. Bo Wang (Associate Professor of Bioengineering, Stanford) explains what regeneration needs: information from the genome, from development, and from the adult body itself, such as positional cues held in muscle. We discuss why some animals regenerate and others do not, ruptoblasts (planarian immune cells that explode to kill), and how to match cell types across species that diverged half a billion years ago. The motif vocabulary of cell types are often conserved while the exact grammar and the wiring between transcription factors and their target genes can be reshuffled. We close with how AI models of regulatory DNA help study non-model organisms, design new regulatory sequences, and where virtual cells meet perturbation and causation.Bo Wang: https://bioengineering.stanford.edu/people/bo-wang

    Wang lab: https://wanglab.stanford.edu (@PhysPlanaria: https://x.com/PhysPlanaria)

    Kundaje lab: https://kundajelab.stanford.edu

    Regulatory grammar across evolution, Chai, Gibson et al., Nature Ecology & Evolution 2026 -https://doi.org/10.1038/s41559-026-03164-5

    Ruptoblasts (exploding cells), Chai et al., Cell 2026: https://doi.org/10.1016/j.cell.2026.05.008

    Check out my websites: https://scholar.google.com/citations?user=2IzutB8AAAAJhttps://www.shuttertimewithvivek.com

    Chapters


    00:00 Trailer02:47 Cut a worm, get new worms04:33 Human limbs?10:23 Regeneration information in muscle11:41 Regeneration across animals21:06 Gene regulatory networks: kernels and switches23:18 Exploding immune cells26:51 Matching cell types across species37:48 Vocabulary vs grammar of regulatory DNA39:13 Motif turnover: flies vs mammals44:43 Terminal selectors and shuffled wiring48:58 Developmental hourglass51:23 AI for non-model organisms54:38 Designing regulatory sequences1:00:33 Virtual cells, perturbation and causation

    1 hr 5 min
  • How to interpret deep learning models, and extend DeepLIFT to AlphaGenome and beyond.

    Wish you could do fast and more useful interpretation of large models like AlphaGenome? How can we tell what a deep learning model has learned? Ruchir Rastogi and Valeh Amiri from the Kundaje lab at Stanford walk through the main ways to interpret sequence-to-function models: in silico mutagenesis, gradient x input, integrated gradients, Shapley values, DeepSHAP and DeepLIFT. They explain why each method gives the answer it does (saturation, the role of a reference..), and how their work with Salil Deshpande extends DeepLIFT to virtually any deep learning architecture, including state-of-the-art transformer-based models such as AlphaGenome: by decomposing modules into simple primitives or by using local integrated gradients.

    1 hr 31 min
  • Understanding DNA Sequence Made Easy: A PyTorch Reimplementation of AlphaGenome

    How do DNA foundation models help us understand DNA at scale? In this video, I discuss how modern sequence-to-function DNA models can uncover biological meaning from genomic sequence with our guests Alejandro Buendia and Danila Bredikhin from Anshul Kundaje’s lab at Stanford. We also discuss their PyTorch reimplementation of the foundational DNA sequence-to-function model, AlphaGenome, that makes it easier for the community to use and build on. This work is featured through genomicsxai, a peer-reviewed blog-style alternative to traditional publishing.

    Links

    genomicsxai: https://genomicsxai.github.io

    AlphaGenome in PyTorch: https://genomicsxai.github.io/blogs/2...

    scverse: https://scverse.org

    1 hr 1 min
  • What Can Worms Teach Us in the Age of AI?

    In the age of AI and virtual cells, what do model organisms such as C. elegans still have to offer? In this video, I speak with Dr. David Sherwood, Professor and Associate Chair of Biology at Duke University, about why worms remain a powerful system for biological discovery. We discuss the unique advantages of C. elegans, including its rapid life cycle, transparency, genetic tractability, invariant cell lineage with the fate of every somatic cell mapped, and that it is the only organism with a completely mapped neural connectome. We also explore how the Sherwood Lab uses genome editing, genetics, and live imaging to study the fascinating biology of basement membranes, including anchor cell invasion, a process with important parallels to invasive behavior in cancer.

    1 hr

About Science Shutter Time

From the publisher's feed

We do deep dives with scientists on cutting-edge science