Sign in

Carl de Boer

@carldeboer.bsky.social
570 followers 188 following 120 posts

Associate Professor, UBC school of Biomedical Engineering. Trying to enable personalized medicine by solving gene regulatory code.

PostsRepliesMedia
Carl de Boer @carldeboer.bsky.social · 5h
Not the death metal I thought I was in for, but still cool.
000
Carl de Boer @carldeboer.bsky.social · 08/10/2026
AI generated? Mixing up the graphics for coding and non coding...
100
Reposted by Carl de Boer
Ben de Bivort @debivort.bsky.social · 22/09/2026
Hello friends, the Harvard Life Science departments have banded together to hire a new colleague working at the intersection of Life Sciences and AI. Could be primarily an experimentalist, or could be primarily a compute/theorist. academicpositions.harvard.edu/postings/16773
academicpositions.harvard.edu
Tenure-Track Professor in Life Science and AI
The Life Sciences Departments in Harvard Faculty of Arts and Sciences seek to recruit a tenure-track professor whose research is at the interface of AI and life sciences, from methodological developme...
34254
Reposted by Carl de Boer
Maitreya Dunham @maitreya.bsky.social · 24/09/2026
Exciting news: @uwgenome.bsky.social will be hiring for two tenure-track faculty positions this year, one open rank and one assistant professor. Both are “Genome Sciences” interpreted broadly so we welcome applicants working in a variety of fields and methods. Ad will be out shortly.
1107105
Carl de Boer @carldeboer.bsky.social · 23/09/2026
Depressing to see all the AI generated grad school applications. These slop grenades waste our time and even if "successful" (getting a position) it will be a terrible match since the applicant was actually the AI. If I wanted a lab full of AI researchers, no people would be needed.
031
Reposted by Carl de Boer
Arnaud Krebs @arnaudkr.bsky.social · 21/09/2026
We are recruiting! Wet&Dry either combined or separate. If you are excited about gene regulation, new methods in genomics and AI-based interpretation! Talk to us!
0126
Reposted by Carl de Boer
Keystone Symposia @keystonesymposia.bsky.social · 16/09/2026
Are you a postdoc or early-career scientist? The Keystone Symposia Fellows Program may be for you. Career development. SAB meetings with leaders in your field. 1:1 mentorship. A community of Fellows committed to each other's growth. Applications close in 2 weeks on Sept. 28.
134
Reposted by Carl de Boer
Vijay Ramani @vram142.bsky.social · 14/09/2026
Everything you wanted to know about single-molecule epigenomics but were too afraid to ask! Wonderful to co-write (h/t ENORMOUS lift by @arnaudkr.bsky.social) this primer on our nascent field. Also, a 💯 example of wonderful colleagues building something *together*, not in competition =)
nature.com
A practical guide to studying genome function using single-molecule genomics
Nature Reviews Molecular Cell Biology - Single-molecule genomics methods are used to study the activity of regulatory factors on individual DNA molecules genome-wide, thereby enabling...
07526
Reposted by Carl de Boer
Wendy Bickmore @wbickmor.bsky.social · 08/09/2026
It was great working with students Kun and Ryan and @hannahlong.bsky.social, combing the literature to try and find examples of bona fide disease-causing variants in non-coding elements - promoters, enhancers and silencers. www.nature.com/articles/s41...
nature.com
Mechanisms underlying disease-causing variants in promoters and enhancers - Nature Genetics
This Review discusses how rare-disease-causing variants in the noncoding genome impact gene regulation, why these examples are so few and how new approaches could accelerate discovery of noncoding var...
16424
Reposted by Carl de Boer
Nozomu Yachie @nzmyachie.bsky.social · 04/09/2026
Maybe it’s time to say goodbye to UMAP and other low-dimensional embeddings (at least in many single-cell analyses). We developed MILK to encode single-cell populations into tree representations at unprecedented scales. www.biorxiv.org/content/10.6... Here’s what MILK is: 🧵
14425
Reposted by Carl de Boer
Emma Dann @emmamarydann.bsky.social · 28/08/2026
Our work on systematic perturb-seq of primary human T cells is now out in Cell 🎉 www.cell.com/cell/fulltex... It's been a privilege to work with @ronghuizhu.bsky.social between @jkpritch.bsky.social @marsonlab.bsky.social labs, with a dream-team of co-authors ❤️ Highlights in preprint thread👇
cell.com
Genome-scale perturb-seq in primary human CD4+ T cells maps context-specific regulators of T cell programs and human immune traits
A dynamic atlas of gene regulation was generated by perturbing every expressed gene across 22 million primary human CD4+ T cells under resting conditions and following re-stimulation. The resulting ma...
14620
Reposted by Carl de Boer
Fred Hutch @fredhutch.org · 15/08/2026
🔬 Fred Hutch Cancer Center is hiring an Assistant or Associate Professor in the Basic Sciences Division @basicsci.fredhutch.org. Join a world-class research community in Seattle and help advance discoveries in fundamental biology. Apply: apply.interfolio.com/190640
02322
Carl de Boer @carldeboer.bsky.social · 10/08/2026
Watched this develop over the years and it is super cool!
164
Reposted by Carl de Boer
Alejandro Montenegro @aemonten.bsky.social · 05/08/2026
"Here we describe a systematic effort (‘Codebook’) to determine the sequence specificity of 332 putative and poorly characterized human TFs." www.nature.com/articles/s41...
nature.com
An expanded codebook of human transcription factor DNA-binding specificity - Nature
Results from a panel of assays that analyse different aspects of DNA sequence specificity reveal more than 100 new motifs to aid the characterization of putative human transcription factors.
14312
Reposted by Carl de Boer
Centre de Regulació Genòmica (CRG) @crg.eu · 03/06/2026
Congratulations to CRG's @benlehner.bsky.social on receiving the 2026 Premio Rei Jaume I for Biomedical Research. The award recognises his pioneering discoveries on genetic variation, evolution and disease, helping transform our understanding of biology and human health. #genomics #genetics #CRG
Ben Lehner at the CRG
0114
Reposted by Carl de Boer
Jay Shendure @jshendure.bsky.social · 30/07/2026
New preprint from Shendure Lab on Locus-Scale aka Long-@$$ MPRAs led by the amazing Abby McGee & @carinabiar.bsky.social Most MPRAs test ~300 bp fragments next to a promoter. But real enhancers are bigger, act combinatorially and from a distance. 1/n www.biorxiv.org/content/10.6...
26437
Reposted by Carl de Boer
Lior Pachter @lpachter.bsky.social · 30/07/2026
One of the major pains of academia is having to click through labyrinths of webpages to submit papers. Now, thanks to Joe Rich (and AI), problem solved! His tool is called PaperPush and is available at github.com/pachterlab/p... It's super easy to use! 1/🧵
github.com
GitHub - pachterlab/paperpush
Contribute to pachterlab/paperpush development by creating an account on GitHub.
36430
Reposted by Carl de Boer
Jenny Chen @jennychen.bsky.social · 29/07/2026
The Chen Lab @ EMBL Rome (current population: 1) is hiring! Looking for an experimental Lab Manager w/ a PhD and experience in molecular bio and mouse-based neuro research. Apply if you like: mouse behaviors, high-throughput sequencing, good coffee & ancient history tinyurl.com/48h66vfh
tinyurl.com
Lab Manager
The European Molecular Biology Laboratory (EMBL) is one of the world’s leading life science research organizations. With sites across Europe, EMBL fosters an international, collaborative, and inclusiv...
085
Reposted by Carl de Boer
Albert Blanch-Asensio @ablanchasensio.bsky.social · 20/07/2026
The STRAIGHT-IN (Dual) kit is now available via @addgene.bsky.social! 🧬 A complete 24-plasmid toolkit for installing up to two orthogonal CLYBL landing pads, DASIT-based enrichment without FACS, and efficient large DNA integration in human cells. Link: www.addgene.org/kits/gallowa... 🧵
11911
Reposted by Carl de Boer
bioRxiv Genomics @biorxiv-genomic.bsky.social · 19/07/2026
scWeave: A deep learning model that bidirectionally translates between gene expression and chromatin structure at single cell resolution www.biorxiv.org/content/10.64898/20…
031
Carl de Boer @carldeboer.bsky.social · 18/07/2026
Talk to references. Not an email. Definitely not a letter. Phone or zoom.
130
Carl de Boer @carldeboer.bsky.social · 14/07/2026
Well all my grants go to GMX and it is the only one I feel comfortable reviewing for
010
Carl de Boer @carldeboer.bsky.social · 10/07/2026
They need a mechanism to both review and submit in the same cycle and same panel (e.g. by splitting the panel in half). As funding rates go down (or fewer app cycles are offered), more people will have apps under consideration. If we eliminate them, few will be left to review.
100
Reposted by Carl de Boer
FragileNucleosome @fnucleosome.bsky.social · 17/06/2026
We just started our summer break, but we're also planning our September-December seminar schedule! If you have an exciting chromatin/transcription story that will be ready for the spotlight by then, use the survey below to nominate yourself to give a #FNucleosome talk 🧬🧬 forms.gle/1EKYN2Nt79ar...
forms.gle
Application to speak in the Fragile Nucleosome series
Thank you for your interest in our seminar series! We use a variety of different session formats but, in general, trainees can plan for a ~15 min talk + 5min Q/A and PIs can plan for a ~25min talk +...
01011
Reposted by Carl de Boer
Richard Sever @richardsever.bsky.social · 06/07/2026
“move from preprint to peer-reviewed publication leaves the central claims of most abstracts intact, indicating preprints are a reliable source…papers that were never posted as preprints were retracted at roughly twice the rate of those that were…” www.biorxiv.org/content/10.6...
biorxiv.org
Tracking claim changes from preprint to publication across 72,644 biomedical studies using large language models
Preprints now disseminate a large share of biomedical research before peer review. Because they have not yet passed peer review, some scientists regard preprint claims as unverified or potentially unr...
611158
Carl de Boer @carldeboer.bsky.social · 20/06/2026
This year, mostly same content, I gave students mini Chocolate bars to complete the survey in class. No conditions for how they rated the class and honor system that they completed survey at all. Way better ratings than previous years.
110
Carl de Boer @carldeboer.bsky.social · 19/06/2026
Yeah, it is totally disheartening to see the exciting projects that keep coming back and are still not going to be funded. Doing the work and making a difference is great. Doing the work to have the grants largely not funded anyway not so much.
010
Reposted by Carl de Boer
Nature Biotechnology @natbiotech.nature.com · 11/06/2026
Large-scale, spatially resolved panoramic CRISPR screening in native tissue environments using Perturb-DBiT -@rongfan8.bsky.social @yaleschoolofmed.bsky.social www.nature.com/articles/s41...
nature.com
Large-scale, spatially resolved panoramic CRISPR screening in native tissue environments using Perturb-DBiT - Nature Biotechnology
In vivo CRISPR genetic perturbations are spatially mapped at scale.
2162
Reposted by Carl de Boer
Teif lab @teiflab.bsky.social · 12/06/2026
Great preprint by @vram142.bsky.social & Co www.biorxiv.org/content/10.6... "fiber homotypy," where chromatin fibers with similar nucleosome spacing interact more frequently in 3D --> Related to our previous "nucleosome homology recognition" in chromatin royalsocietypublishing.org/rsif/article...
0165
Carl de Boer @carldeboer.bsky.social · 12/06/2026
Brutal IP burn: "You have ordinary skill in the art."
000
Carl de Boer @carldeboer.bsky.social · 10/06/2026
We had similar findings (not emphasized in the publication) in our study of variant characterization at TNFAIP3. Different "monocyte" cell lines looked very distinct. "B cell" lines looked very distinct. Cell lines are often bad proxies for primary cells in terms of gene regulation.
050
Reposted by Carl de Boer
bioRxiv Genomics @biorxiv-genomic.bsky.social · 02/06/2026
Gigabase-scale deletion scanning of the human genome www.biorxiv.org/content/10.64898/20…
076
Carl de Boer @carldeboer.bsky.social · 30/05/2026
Understanding where genomics models work and fail will be critical to their successful deployment. Meta models can tell us this, and give some insight into failure modes. 3/3 for @muntakimrafi.bsky.social this week. CSHL AIxBIO to boot (sorry to miss!) What a week!
020
Carl de Boer @carldeboer.bsky.social · 29/05/2026
Which active learning methods are best for genomics models? See Rafi's thread on our recent preprint, below. www.biorxiv.org/content/10.6... Rafi is having a good week - he's got one more preprint to post!
biorxiv.org
052
Carl de Boer @carldeboer.bsky.social · 28/05/2026
Thanks to @muntakimrafi.bsky.social who worked tirelessly to expand the scope of the original paper, making our conclusions more nuanced and robust, co-authors @nzmyachie.bsky.social and Brett Kiyota for help making hashFrag, and you for getting this far. Link: www.biorxiv.org/content/10.1...
biorxiv.org
121
Carl de Boer @carldeboer.bsky.social · 28/05/2026
We now recommend that you *allow the leakage in the test dataset*. 😮 Although you could use hashFrag to remove it (we recommend this for validation set), doing so on the test dataset would *not prevent memorization*. Instead, it would now be harder to detect.
100
Carl de Boer @carldeboer.bsky.social · 28/05/2026
There are many important considerations for those training sequence-to-function models in the paper, most of which are not covered here, so we highly recommend you read. However, there was one critical point that we think our original preprint got wrong: what to do with the test data.⁉️
100
Carl de Boer @carldeboer.bsky.social · 28/05/2026
At the locus (not bin) scale, Enformer performed much better on sequences that were dissimilar to those it had seen during training. This showcases that the model has learned some amazingly generalizable things, but is constantly being misled by all the homology in our genomes.
100
Carl de Boer @carldeboer.bsky.social · 28/05/2026
Indeed, across all tracks, the more Enformer memorized, the worse it performed, indicating that we could likely boost performance by reducing memorization.
100
Carl de Boer @carldeboer.bsky.social · 28/05/2026
As expected with 5000+ tracks, the results were more complex here, but clear trends remained: Enformer appeared to be anchored to what it had seen during training (MSE) but performed worse as a result (Pearson's r).
100
Carl de Boer @carldeboer.bsky.social · 28/05/2026
Next we characterized Enformer, which predicts thousands of functional genomics tracks across cell types and tissues. It was a convenient choice for this because its predictions are given in 128 bp bins - long enough to identify leakage between the training and test sets.
100
Carl de Boer @carldeboer.bsky.social · 28/05/2026
This revealed that the U shape was likely a product of memorization. Importantly, this also revealed that different models memorize to different degrees, indicating that this is something we can optimize in the future 🙏
100
Carl de Boer @carldeboer.bsky.social · 28/05/2026
We thought this was due to memorization that was failing as sequences diverge - models could identify the homology but the function was no longer the same. To test this we had to come up with a memorization measure, which we again did by nearest neighbors.
100
Carl de Boer @carldeboer.bsky.social · 28/05/2026
We next checked how performance varied by sequence similarity to training data for open chromatin prediction models. The U shape that was much more subtle for MPRA models was now clear as day: models perform well on close homologs and unrelated sequences, but poorly for distant homologs.
100
Carl de Boer @carldeboer.bsky.social · 28/05/2026
When hashFrag is used to create an orthogonal validation dataset, model performance increases🎉 We think this is because an independent validation dataset can better estimate the degree of overfitting, resulting in early stopping selecting a better model.
100
Carl de Boer @carldeboer.bsky.social · 28/05/2026
We were grouping sequences into similarity bins using Smith-Waterman alignment score, but it... was... so... slow...🐌 We made hashFrag to accelerate homolog detection, which uses the sensitive homology detection of BLAST. 🚀
110
Carl de Boer @carldeboer.bsky.social · 28/05/2026
OverfitNN works surprisingly well for close homologs in an MPRA context, which makes sense as sequence encodes function, so similar sequences encode similar functions.🤓 Performance drops as sequences diverge in sequence and function.📉 Surprisingly, neural network models show a similar trend😱
110
Carl de Boer @carldeboer.bsky.social · 28/05/2026
We expand our use of OverfitNN - our cheekily-named sequence to function "model" that learns no biochemical logic at all - it explicitly relies on homology for "prediction" by simply transferring labels from nearest sequence neighbors in the training data set.
100
Carl de Boer @carldeboer.bsky.social · 28/05/2026
To show how this homology structure can be an issue for genomics models, we performed a simulation where we controlled the homology structure. We show that model performance can be inflated if evaluated on sequences that are homologous to training examples.
100
Carl de Boer @carldeboer.bsky.social · 28/05/2026
If we care about predicting phenotype from genotype, this is a major problem as our models will have seen homologs of many genotypes before. SNV doppelgangers, exact copies of fine mapped GWAS alternate alleles, are prevalent, potentially confounding variant "prediction" with memorization.😮
110