iiinova — where your digital self keeps evolving

And a huge thanks to the Mirny, Maehr, and Dekker labs, as well as the feedback and support of our multi-PI center over the course of Phase II of the 4DN Consortium.

040

This all was no small feat, but the outcome of tireless effort across a team spanning 4 labs! Congratulations to all involved, including lead authors Xiangru, Denis, Thomas and Liyan, as well as Johan Gibcus, @renemaehr.bsky.social , @leonidmirny.bsky.social , and @jobdekker.bsky.social

151

These transitions are also tied to regulation. Using joint decomposition of long-range interactions across all five stages, we isolated a trajectory where H3K27me3 changes from sharp peaks in to broad domains covering coinciding with silencing of genes marking alternative, non-hepatic lineages.

3. LCHi-C shows that chromatin interactions stabilize over differentiation and do so in a defined order: speckle-associated domains harden at commitment, while the rest of the genome hardens during maturation.

2. At hepatic commitment, compartmentalization increases dramatically, driven by segregation of speckle-associated from lamina-associated chromatin. H3K9me2 gets deposited and LADs peripheralize, speckles consolidate, and B-type lamins get replaced by lamin A/C.

161

1. In hESCs, chromosomes form diffuse, extensively intermingled territories. On exit from pluripotency, they individualize while centromeres make a transient excursion to the nuclear periphery that reverses at the next stage.

151

We found that nuclear reorganization during development does not occur as one coordinated switch but as at least 3 major temporally separable transitions, each with a distinct molecular signature:

To probe nuclear dynamics during differentiation on a common timeline, we developed a five-stage differentiation system of H1 human ESCs to hepatocyte-like cells (HLCs), and profiled every stage using a battery of techniques. We further resolved lineage commitment at single-day resolution.

140

Nuclear organization differs profoundly between pluripotent and differentiated cells, but the order in which properties change and their how they are related to each other are not yet understood.

130

Nezar Abdennur @nvictus.bsky.social · 16/07/2026

The 4DN Center for 3D Structure and Physics of the Genome proudly presents: "Stepwise reorganization of chromosome conformation and nuclear organization during stem cell differentiation". www.biorxiv.org/content/10.6...

14419

Also excited to kick off a new initiative with @jmschreiber91.bsky.social and the polars-bio (biodatageeks.org/polars-bio/) team, where we are working towards a better agent-ready substrate for genomic analysis and ML: composeq.dev

composeq.dev

CompoSeq

Composable foundations for genomic data.

050

Nezar Abdennur @nvictus.bsky.social · 16/07/2026

Really enjoyed presenting our genomics talk "Declare, don't parse" at #SciPy2026 with Conrad Bzura! tl; dr Problem #1: Genomics is coupled to file formats, not data Problem #2: Most genomic operations aren't even "genomic" Solution: Be composable! Slides here: abdenlab.org/scipy26

abdenlab.org

Declare, Don't Parse: Toward composable genomic analysis - Slidev

SciPy 2026 — Nezar Abdennur and Conrad Bzura

151

My fav example in the teaser is the attraction-repulsion spectrum. You can see how a single hyperparameter reconstitutes a close approximation of the major neighbor-based embedding methods: Laplacian eigenmaps -> ForceAtlas2 -> UMAP -> t-SNE.

dtour is built around two types of workflow: - revealing structure in a single high-dimensional space through hyperdimensional tours, and - validating or comparing embeddings through sequential tours over a series of 2D projections.

100

dtour brings it all into one exploratory interface: you can slide through a overview of projection “keyframes”, a guided cyclical tour along geodesic paths connecting keyframes, fine manual axis manipulation for user-driven excursions, and a wandering grand tour for serendipitous discovery.

100

Some tools help by letting you traverse or “tour” multiple projections, but they restrict the choice of tour modality up front and provide little to no steerability.

100

Other methods try to bring the whole manifold structure of your data into view by heuristically squeezing it into a 2D plane, which can muddle genuine structure with layout algorithm artifacts.

Some dimensionality reduction methods help you find interesting angles to look at your data but hide everything orthogonal to the projection you’re looking at.

100

To develop a better "feel" for big datasets without being fooled by artifacts, no single representation is enough: you need to explore! And to explore, you need navigability!

dtour was a really fun project with Fritz resulting from years of discussion about manifold learning and dimensionality reduction, the limitations of state-of-the-art tooling, and what we could do to make it better.

100

Elphege Nora Lab at UCSF @elphegenoralab.bsky.social · 13/05/2026

High-dimensional data is hard to understand. But is it truly cursed? To help you build better intuition for high-dimensional datasets, @lekschas.de and I developed dtour, a visualization tool for smoothly navigating through data projections.

1167

Reposted by Nezar Abdennur

23/ Check out this viewer by @nvictus.bsky.social to get a sense of how few cohesins make enhancer-promoter bridges 🤯 abdenlab.org/cohesin-brid...

162

Nezar Abdennur @nvictus.bsky.social · 03/02/2026

Many of you may have received pessimism from me over the years about using Hi-C to predict distal regulatory effects, mired with paradoxical optimism that cohesin loop extrusion is key. Well, the theoretical story underlying my seemingly paradoxical rambling is finally ready to be told. 👇

1173

💡 SciPy is a vibrant, interdisciplinary conference that brings together researchers, developers, and innovators who are shaping the future of computing, data science, and AI for scientific research (and a love for Python!). Get a taste of past talks on YT: www.youtube.com/@SciPy-Conf/...

youtube.com

SciPy

000

Nezar Abdennur @nvictus.bsky.social · 03/02/2026

🌟 Sign up for office hours to learn more about preparing your proposal 👉 www.scipy2026.scipy.org#officehours

scipy2026.scipy.org

SciPy 2026

100

Nezar Abdennur @nvictus.bsky.social · 03/02/2026

I’m excited to be co-chairing the Biological & Medical Sciences track for the #SciPy2026 conference 🎉 in July, along with Aishwarya Chander and Jacqueline Valeri. www.scipy2026.scipy.org 🗓️Call for proposals is open! ***Submission deadline: Feb 25*** 🚀 Submit here 👉 pretalx.com/scipy-2026/cfp

scipy2026.scipy.org

SciPy 2026

100

Reposted by Nezar Abdennur

Jacob Schreiber @jmschreiber91.bsky.social · 27/08/2025

In the genomics community, we have focused pretty heavily on achieving state-of-the-art predictive performance. While undoubtedly important, how we *use* these models after training is potentially even more important. tangermeme v1.0.0 is out now. Hope you find it useful!

14414

Nezar Abdennur @nvictus.bsky.social · 25/08/2025

My talk on #Composability in genomic software at #SciPy2025 is up on YouTube where I showcase both #anywidget and #oxbow. Thank you to the organizers for the opportunity to present this to both computational biologists and the wider scientific computing community! www.youtube.com/watch?v=G22_...

youtube.com

Nezar Abdennur - Accelerating Genomic Data Science and AI/ML with Composability | SciPy 2025

YouTube video by SciPy

031

Reposted by Nezar Abdennur

trevor manz @manzt.sh · 07/08/2025

if interested in creating anywidgets of your own, our tutorial was finally shared to youtube: www.youtube.com/watch?v=frEo...

youtube.com

Abdennur, Lekschas, & Manz - Bring your __repr__’s to life with anywidget | SciPy 2024

YouTube video by SciPy

273

Nezar Abdennur @nvictus.bsky.social · 19/08/2025

Our #anywidget tutorial from last year's #SciPy conf was uploaded to youtube! Check it out for a hands-on walkthrough to create your own web-based widgets.

We anticipate that joint dimensionality reduction and projection will become a foundational norm for comparative and integrative analysis of long-range interaction profiles in Hi-C/3C+ data. e.g. existing methods for working with classic A/B vectors can be extended to joint higher-order embeddings.

000

We jointly-hic to create an atlas of 89 human Hi-C samples, uncovering distinct patterns of nuclear architecture associated with heterochromatin composition and demonstrating how higher-order principal components capture missing information about gene expression and regulatory element activity.

100

jointly-hic accomplishes this using mini-batch incremental PCA, allowing for joint decomposition of arbitrarily many contact matrices at any resolution with constant memory.

100

Joint decomposition allows for robust and directly comparable low dimensional representations of arbitrarily many contact maps, providing insights into genome organization across diverse biological contexts, from different tissues to developmental stages.

100

The classic A/B compartment track comes from matrix factorization of a contact matrix into eigenvectors or PCs. Done separately, each map is projected onto a different coordinate system. Comparing such vectors directly is problematic, especially if seeking info from **higher-order** components.

Nezar Abdennur @nvictus.bsky.social · 09/07/2025

Yes, and more recently Zarr too academic.oup.com/gigascience/... While oxbow makes legacy data more accessible, it is a good conduit to more general-purpose persistent storage.

academic.oup.com

Analysis-ready VCF at Biobank scale using Zarr

AbstractBackground. Variant Call Format (VCF) is the standard file format for interchanging genetic variation data and associated quality control metrics.

020

Reposted by Nezar Abdennur

Jacob Schreiber @jmschreiber91.bsky.social · 13/05/2025

A huge challenge I face when doing ML + genomics analysis is *friction*: the stupid error messages (wrong device!) and dumb implementation issues that snap you out of the zone. I wrote a vignette on how tangermeme has helped me reduce this friction: tangermeme.readthedocs.io/en/latest/ho...

tangermeme.readthedocs.io

How To: Reduce Friction and Save Time with Tangermeme — tangermeme v0.1.0 documentation

1143

Reposted by Nezar Abdennur

Jacob Schreiber @jmschreiber91.bsky.social · 30/06/2025

(4) bpnet-lite: Load official Chrom/BPNet models into PyTorch for downstream tangermeme integration. Improved command-line tools + docs. Still concerns about perf of models trained from scratch -- will be resolved next version! github.com/jmschrei/bpn... bsky.app/profile/jmsc...

github.com

GitHub - jmschrei/bpnet-lite: This repository hosts a minimal version of a Python API for BPNet.

This repository hosts a minimal version of a Python API for BPNet. - jmschrei/bpnet-lite

121

We’re excited and eager for feedback, so please give oxbow a try! `pip install oxbow`

020

It also supports: * Column projection and pushdown (parsing only the fields you need) * Complex and nested field types (e.g. alignment tags, variant genotype call data, etc.) * Genomic range-based queries via an index * User-defined transports and file systems

This update (v0.4.x) provides complete #ApacheArrow data models for 11 file formats and counting, including the GA4GH/htslib formats and UCSC’s BigWig/BigBed.

We revamped the #rustlang backend and implemented a new "DataSource" API in #Python, which allows for streaming conventional #genomic files – in-memory, on-disk, or in the cloud – into the modern data tools you use regularly, including #Pandas, #Polars, #DuckDB, and #Dask.