Axel Visel
LightNews LightNews
banner
axelvisel.bsky.social
Axel Visel
@axelvisel.bsky.social
3.5K followers 600 following 260 posts
Genome Scientist. Studies how DNA makes humans, mice, plants, microbes. At Lawrence Berkeley National Lab and Joint Genome Institute. Views are my own.
Posts Replies Media Videos
axelvisel.bsky.social
Soils contain an amazing diversity of functions encoded in plasmids.

The Global Soil Plasmidome Resource: 98,728 soil plasmids from 6,860 samples.

Led by @mattlabguy.bsky.social and @apcamargo.bsky.social at @jgi.doe.gov @biosci.lbl.gov @berkeleylab.lbl.gov

www.nature.com/articles/s41...
A stylized infographic showing the workflow for building a global soil plasmidome resource on the left and a textured world map on the right. The workflow depicts three input data streams from metagenomic datasets and isolate plasmids, which pass through steps like quality control, clustering, functional annotation, CRISPR analysis, host assignment, and detection of gene categories such as biosynthetic clusters, antimicrobial resistance, antimicrobial peptides, and CAZymes. All outputs feed into a central SQL database. The world map shows sample locations across the globe as teal circles of varying size, highlighting regions from many plasmids were recovered. Adapted from Fig 1A and 1B in Fiamenghi et al., doi:10.1038/s41467-025-65102-6
November 18, 2025 at 7:18 PM
axelvisel.bsky.social
Here is our speaker lineup:
jgi.doe.gov/work-with-us...

Featuring @nekton4plankton.bsky.social @annedekas.bsky.social @microyunha.bsky.social @masarunobu.bsky.social and more
Headshots of eight featured speakers at the 2025 NeLLi Symposium. Top row, left to right: John M. Archibald (Dalhousie University, keynote), Harriet Alexander (Woods Hole Oceanographic Institution), Matthew Brown (Mississippi State University), Anne Dekas (Stanford University). Bottom row, left to right: Yunha Hwang (Harvard University), Masaru Nobu (JAMSTEC), Vivek Natarajan (Google), Olga Zhaxybayeva (Dartmouth College).
August 25, 2025 at 9:47 PM
axelvisel.bsky.social
Postdoc opportunity with @tomasbruna.bsky.social here at @jgi.doe.gov @berkeleylab.lbl.gov:

Develop, benchmark, and apply Plant Genomic Language Models (gLMs) 🌿🧬💻

jobs.lbl.gov/jobs/computa...

#AI #PlantGenomics #gLMs #Postdoc
Graphic announcing a Computational Postdoctoral Fellow position. Text: Computational Postdoctoral Fellow – Plant Genomic Language Models (gLMs). JGI is seeking a Computational Postdoctoral Fellow to join a funded project at Berkeley Lab’s Joint Genome Institute applying genomic language models (gLMs) to unlock new biological insights in plants. Join one of the world’s leading genomics facilities to pioneer AI applications in plant biology, with access to cutting-edge computational resources and diverse genomic datasets. This project will focus on developing and exploring applications of gLMs to plant genomes, from identifying novel regulatory elements to revealing evolutionary patterns across species, with particular interest in cross-species and data-limited scenarios. In this role, you will design tasks and datasets to test and extend gLM capabilities, apply models to generate new biological hypotheses, and share findings through high-impact publications and presentations.
August 21, 2025 at 12:05 AM
axelvisel.bsky.social
On the cover of @natplants.nature.com this month:

@leobaumgart.bsky.social @greensi.bsky.social @abmora.bsky.social @omalley-regulome.bsky.social et al. map binding sites for 360 TFs across 10 🌿plant🌿 species using a new multiDAP approach

Here is Sharon's explainer 🧵:
bsky.app/profile/gree...
Nature Plants August 2025 cover. Abstract artwork of overlapping colorful waveforms in red, orange, yellow, and purple on a black background, resembling mountain ridges or transcription factor binding profiles. The caption reads “Cistromes uncovered.”
August 20, 2025 at 12:45 AM
axelvisel.bsky.social
In 2026, @jgi.doe.gov and EMSL will (for the first time!) hold a joint user meeting in Seattle from March 3-5

Program details to be announced, but our registration site just launched - so you can already secure your spot:

jgi.doe.gov/work-with-us...
Promotional banner for the 2026 EMSL-JGI User Meeting. Theme: “Empowering Automated Laboratories – Integrating Experimentation with Data.” Event dates: March 3-5, 2026. Location: Seattle Marriott Waterfront, 2100 Alaskan Way, Seattle. The banner features a skyline of Seattle with Mount Rainier in the background, surrounded by colorful hexagonal graphics showing scientific imagery such as DNA, lab automation, microbial structures, and plants. EMSL and JGI logos are shown at the top.
August 19, 2025 at 5:30 PM
axelvisel.bsky.social
Yes, absolutely! They are clearly not all artifacts. But some of them are and the problem is more pronounced in plants than in animals. Fortunately, the preprint also highlights how a unified re-annotation strategy can reduce these artifacts, making whatever PAV remains even more interesting!
Screenshot of text from the preprint: "Unsurprisingly, within both flowering plants and amniotes, gene PAV becomes increasingly common with increasing divergence time—more diverged genomes share fewer orthologs, likely due to sequence deletions, pseudogenization and other forms of large-scale mutation followed by selection or drift. However, the absolute amount of gene PAV and the degree that PAV is affected by divergence time was remarkably different between plants and animals." https://www.biorxiv.org/content/10.1101/2025.08.14.670405v1.full
August 18, 2025 at 7:31 PM
axelvisel.bsky.social
🌿Plant pangenomes🌿 have a LOT of genomic presence-absence-variation.

Or do they???

A new @jgi.doe.gov preprint by @tomasbruna.bsky.social @jotlovell.bsky.social @avril-m-harder.bsky.social Avinash Sreedasyam takes a closer look

www.biorxiv.org/content/10.1...
A two-panel Scooby-Doo meme. In the top panel, Fred pulls the mask off a villain labeled “Presence-absence-variation in plant genomes.” In the bottom panel, the unmasked villain is revealed to be “Annotation artifacts.”
August 18, 2025 at 6:28 PM
axelvisel.bsky.social
Of a chipmunk
Screenshot of a ChatGPT 5 conversation. Prompt: "Make a schematic drawing of the abdominal cavity of a chipmunk, frontal view, and label all organs." The result shows a cartoon chipmunk with a cartoon of human inner organs superimposed. The organs are overall misplaced, with the liver approximately where the lung should be. Most organs are mislabeled. The liver is labeled as both liver and kidney, the large intestine as small intestine and spleen and urinary bladder. Only the stomach is labeled correctly.
August 15, 2025 at 12:55 PM
axelvisel.bsky.social
Congratulations to Uma Grover on an outstanding poster presentation! 🎉

Uma joined @b-coli.bsky.social's and my group at @jgi.doe.gov @biosci.lbl.gov this summer as a @berkeleylab.lbl.gov SULI intern and carried out exciting work on single-cell profiling of plant-fungal interactions.
Axel Visel, Uma Grover, and Benjamin Cole stand in front of Uma's poster at Berkeley Lab's 2025 Workforce Development & Education (WD&E) Summer Intern Poster Session. Her poster is entitled "Single-cel profiling of sorghum / AMF interaction"
August 9, 2025 at 1:25 PM
axelvisel.bsky.social
Enhancers are "just DNA" - finding all of them in the genome is hard.

www.nature.com/articles/s41...
August 8, 2025 at 6:25 PM
axelvisel.bsky.social
Hiding in plain sight - how close are we to mapping ALL 🧬enhancers🧬 in the genome?

Our new paper by Mannion et al. takes a systematic look at "hidden enhancers" and why they remain so hard to find. With @mosterwalder.bsky.social, @jlopezrios.bsky.social & many more

www.nature.com/articles/s41...
A busy tool wall in a shed. At the bottom there are instructions saying "Find the 10 hidden enhancers!" Across the wall between the tools are 10 enhancers, represented as DNA helices, but they are difficult to find in the style of a "hidden object" puzzle. Original photo by Lachlan Donald, https://www.flickr.com/photos/lox/9408028555
August 8, 2025 at 6:10 PM
axelvisel.bsky.social
🧬 @jgi.doe.gov is hiring! We're seeking an Advanced Analysis Manager to lead data infrastructure and software engineering supporting the generation, analysis, and dissemination of genomics data

Hybrid | SF Bay Area | Apply by Aug 7

jobs.lbl.gov/jobs/advance...

@biosci.lbl.gov @berkeleylab.lbl.gov
Buildings on the Lawrence Berkeley National Lab campus, with the San Francisco Bay and the Golden Gate Bridge in the background. It's sunset time, the sky is orange. On the left is the Integrative Genomics Building (IGB), which houses the JGI, NMDC and KBase. On the right is the BioEPIC building.
July 29, 2025 at 11:49 PM
axelvisel.bsky.social
Catalan got it right, though:

"ratpenat" = the "feathered rat"

There is even one in Valencia's coat of arms
Coat of arms of Valencia, Spain, featuring a gold crown topped with a black bat with outstretched wings. Below the crown is a blue and red shield with vertical red and yellow stripes and ornate golden floral patterns
July 16, 2025 at 6:48 PM
axelvisel.bsky.social
Getting the message across from a distance:

🧬REX🧬 is a “Range EXtender” element that can turn short-range into looooooooooooong-range enhancers

Great collaboration led by @gracebower.bsky.social and @evgenykvon.bsky.social

www.nature.com/articles/s41...

@biosci.lbl.gov @berkeleylab.lbl.gov
Two-panel cartoon set in a desert with mountains and cacti. Top panel: A green box labeled “ENHANCER” stands on the left, shouting weakly “Turn on, please?!?” toward a distant blue spiky blob labeled “GENE” on the right, who is asleep with a “Zzz.” Bottom panel: The enhancer now uses a microphone connected to a large purple speaker labeled “REX,” which amplifies the message “Turn on!!” in a bold speech bubble across the desert. The gene is now awake and alert with wide eyes and a happy expression. Image credit: Emma Visel
July 2, 2025 at 4:03 PM
axelvisel.bsky.social
💯Forgot to cover that in the thread. Yes, it's remarkable how well it worked with chromatin accessibility data alone (and to think how much better it might work with more data modalities and using single-cell data).
Screenshot from the discussion of the paper: "Although our models achieved high positive predictive value and considerable sensitivity, further research is needed to improve on this result. First, future models should be trained directly on enhancer activity measures, instead of on open-chromatin signal, which is just a proxy for activity. Second, models derived from the exact cell type in which activity was observed will probably perform better than models trained on bulk data or data from closely related cell types. Third, the existence of degenerate TF motifs and discrepancies between results obtained from contribution scores and direct signal change prediction imply that the models may contain more information than we are currently able to capture."
June 18, 2025 at 7:40 PM
axelvisel.bsky.social
With respect to the “inner logic” of enhancers: Some enhancers were uniformly (across all structures) affected by loss-of-function mutations, while others had structure-specific motifs for subregions. Shown here for a heart enhancer.
Two rows of mouse embryo heart close-up images showing a series of changes in LacZ reporter staining with progressively reduced activity due to stepwise mutations, with expression first disappearing in left atrium, the right atrium, then left ventricle, then right ventricle, and finally in the outflow tract.
June 18, 2025 at 6:09 PM
axelvisel.bsky.social
Now the real question: Can the best-fit models correctly predict consequences of 1bp mutations? Yes - at least in several cases we tested. This has important implications for interpreting noncoding variants observed in clinical genetics data.
Side-by-side comparisons of wild-type and point-mutated enhancers in mouse embryos, shown with corresponding DNA motif logos. Left: Heart closeups show LacZ staining for a reference embryo (blue signal) and a mutant (A320G) with disrupted MEF2 motif, resulting in loss of expression. Right: Full embryos show reference and C60T mutant with gain of enhancer activity. Associated SOX and SNAI motifs are shown, with a pink bar indicating the nucleotide mutated in the repressive SNAI site.
June 18, 2025 at 6:06 PM
axelvisel.bsky.social
To further tweak this, we found that in some cases combining models from different tissues and/or considering degenerate (suboptimal) motifs improved prediction performance.
Visualization of transcription factor motif predictions from three machine learning models (limb, hindbrain, and glutamatergic neuron) aligned to the same enhancer sequence. Each row displays motif matches for different factors, such as HAND2/TWIST1, SOX, SNAI, NFI, NR/RAR, and Homeobox, highlighted with sequence logos. At the bottom, a block diagram shows the in vivo mutagenesis results for each 12-bp block: wild-type (blue), minor/major loss (light blue), full loss (white), and gain (orange). Gain blocks correspond to repressive motif predictions, suggesting functional repression.
June 18, 2025 at 6:05 PM
axelvisel.bsky.social
We built models from relevant epigenomic data and used our in vivo data to benchmark their performance. We found that the DL models tell us the truth, but not the whole truth: Most predicted functional sites are real (PPV 88%), but they still miss ~40% of functional sites.
Two pie charts summarizing the performance of motif predictions. The left chart shows 88% predictive value, indicating that most predicted motifs were found in functional blocks. The right chart shows 59% sensitivity, meaning that the majority, but not all functional blocks had a predicted motif. Text between the charts reads: “Almost all predicted motifs were in functional blocks” and “Majority of functional blocks had a motif prediction.”
June 18, 2025 at 6:04 PM
axelvisel.bsky.social
But who cares about blocks? We need generalizable, basepair-resolution predictions and insights about TF motifs. Enter @anshulkundaje.bsky.social @anusri.bsky.social and their ChromBPNet deep learning model of chromatin accessibility.

www.biorxiv.org/content/10.1...
A schematic comparing machine learning predictions and in vivo enhancer mutagenesis results. The top row shows predicted transcription factor binding motifs (RFX, SP/KLF, and two E-boxes) along a DNA sequence. Some regions between motifs are labeled with question marks, indicating uncertainty. The bottom row is a bar showing results from in vivo 12-bp block mutagenesis, with blocks shaded from light to dark blue indicating increasing functional impact. Some blocks also have question marks, suggesting functional elements not identified by the model.
June 18, 2025 at 6:04 PM
axelvisel.bsky.social
But the overall sensitivity varied widely across the enhancers. One was hypersensitive - every block mattered. Most had “Achilles’ heel” blocks - single 12bp hits that wiped out function entirely. Only three had gain-of-function blocks, presumably corresponding to repressor sites.
A schematic summarizing the effects of 12-bp block mutations across seven enhancers. Each row represents an enhancer, and each column represents a block. Color-coded blocks indicate the functional impact of each mutation: dark blue for wild-type (no effect), light blue for minor loss, pale blue for major loss, white for full loss, and orange for gain of function. Annotations highlight features such as “Achilles’ heel” blocks (where single mutations abolish activity), regions with multiple gain blocks (suggesting repressors), and stretches with no apparent function. A legend on the right explains the color coding.
June 18, 2025 at 6:02 PM
axelvisel.bsky.social
Turns out functional sequence features are scattered across the entire length of enhancers, and there are MANY of them. About two thirds of all blocks caused changes in function - mostly losses, but also some gains, i.e. ectopic enhancer activity in the wrong tissue.
On the left, a pie chart titled "Distribution of changes" shows proportions of enhancer mutagenesis outcomes: just over half of them show "loss" of function (white), about a third show "no change" (blue) and about a tenth show "gain" (orange). On the right, images of mouse embryos illustrate differences in enhancer-driven LacZ reporter activity. Three columns compare wild-type, transition mutagenesis, and transversion mutagenesis. The wild-type shows strong tissue-specific staining; transition and transversion mutants show reductions or alterations in staining patterns, indicating changes in enhancer activity.
June 18, 2025 at 6:02 PM
axelvisel.bsky.social
Why 12bp blocks? The combined enhancers covered >2kb. That would require a LOT of experiments with 1bp mutations. Using block mutatagenesis, we covered all basepairs efficiently. It was still a LOT of experiments.
Diagram explaining the mutagenesis strategy for enhancer dissection. Left: A DNA sequence is divided into consecutive 12 base pair (bp) blocks, each of which is individually mutagenized. Right: The mutagenesis uses transition substitutions, where A↔G and T↔C. An example sequence is shown before and after mutation: “CCACAGATAAGA” becomes “TTGTGAGCGGAG,” with mutated bases highlighted in red.
June 18, 2025 at 6:01 PM
axelvisel.bsky.social
To find out, we systematically mutagenized 7 human enhancers in 12bp blocks and checked if these mutations change enhancer activity in developing mouse embryos.
A panel of mouse embryos showing LacZ staining patterns driven by wild-type enhancer alleles, illustrating tissue-specific expression. Right: A diagram of possible mutation outcomes with corresponding embryo images. Wild-type expression (blue frame) can shift to gain of function (orange frame), minor loss (light blue), major loss (lighter blue), or full loss (black frame) depending on the mutation. Arrows connect the wild-type to each possible outcome, showing the diversity of enhancer mutation effects in vivo.
June 18, 2025 at 6:00 PM
axelvisel.bsky.social
Textbook knowledge: Enhancers consist of transcription factor binding sites (TFBSs). Once the right TFs have bound, the enhancer turns on.

But which of these sites are actually important? Do we know all the sites? What about redundant sites? Questions over questions.
A schematic illustration of an enhancer with various transcription factors bound. Colored protein in teal, orange, and yellow are interspersed with gray shapes along a black DNA strand. Blue arrows point to the proteins with the following questions: “Combinatorial effects?”, “Which binding sites are important for function?”, and “Redundant binding sites?” The image highlights the complexity of interpreting enhancer function based on transcription factor binding.
June 18, 2025 at 5:58 PM