Igor Ulitsky
@igorulitsky.bsky.social
Associate Professor, Weizmann Institute of Science. RNA biologist, interested in what (long) RNA molecules do and how. Father of 4.
100%, i so missed Twitter 3-4 years ago
November 20, 2025 at 4:45 AM
100%, i so missed Twitter 3-4 years ago
The problems we identified don't invalidate the many studies that have mostly focused on lncRNA functions / KD phenotypes, but rather show that the road to truly understanding not just what lncRNAs are doing, but how they do it, remains long. Now let see what the you/ editors/reviewers think :)
November 13, 2025 at 9:26 AM
The problems we identified don't invalidate the many studies that have mostly focused on lncRNA functions / KD phenotypes, but rather show that the road to truly understanding not just what lncRNAs are doing, but how they do it, remains long. Now let see what the you/ editors/reviewers think :)
Philosophically, often, ChIRP-seq was used as "the last experiment" in a paper, aimed to produce some mechanistic insights (my guess - requested by reviewers), usually with limited follow-up and subsequent validation. The connection between the "peaks" and the lncRNA biology was loose
November 13, 2025 at 9:26 AM
Philosophically, often, ChIRP-seq was used as "the last experiment" in a paper, aimed to produce some mechanistic insights (my guess - requested by reviewers), usually with limited follow-up and subsequent validation. The connection between the "peaks" and the lncRNA biology was loose
Computationally, we provide an approach to test if "probe contamiation" is occuring to assess data quality. Filtering the peaks/reads to retain only high-quality ones remains challenging, as matches as short as 7nt appear to be sufficient for some probes to recover DNA
November 13, 2025 at 9:26 AM
Computationally, we provide an approach to test if "probe contamiation" is occuring to assess data quality. Filtering the peaks/reads to retain only high-quality ones remains challenging, as matches as short as 7nt appear to be sufficient for some probes to recover DNA
The other takehomes - what to do now? Experimentally, we suggest trying to remove the ssDNA ends after sonication. Key controls such as KO cells and antisense probes must be used.
November 13, 2025 at 9:26 AM
The other takehomes - what to do now? Experimentally, we suggest trying to remove the ssDNA ends after sonication. Key controls such as KO cells and antisense probes must be used.
This project ended up being a great collaboration stemming from a presentation at the EMBL Noncoding Genome, which, unfortunately, has been the last meeting in a now-discontinued series. So one takehome message is - talk about unpublished results at meetings!
November 13, 2025 at 9:26 AM
This project ended up being a great collaboration stemming from a presentation at the EMBL Noncoding Genome, which, unfortunately, has been the last meeting in a now-discontinued series. So one takehome message is - talk about unpublished results at meetings!
Indeed, these sites also didn't make biological sense - didn't seem to be connected to the genes regulated by NESPR (as is often the case in other studies). The NESPR results specify which controls (KO, RNAse treat) can be effectively used to understand if the lncRNA binding maps are real.
November 13, 2025 at 9:26 AM
Indeed, these sites also didn't make biological sense - didn't seem to be connected to the genes regulated by NESPR (as is often the case in other studies). The NESPR results specify which controls (KO, RNAse treat) can be effectively used to understand if the lncRNA binding maps are real.
In parallel to our meta-analysis, Louis and Pieter and collaborators studied at the NESPR lncRNA. They found with ChIRP-seq 1000s of seemingly high-confidence sites, but using proper controls, including RNase and control lines that don't express the lncRNA showed these are mostly bogus
November 13, 2025 at 9:26 AM
In parallel to our meta-analysis, Louis and Pieter and collaborators studied at the NESPR lncRNA. They found with ChIRP-seq 1000s of seemingly high-confidence sites, but using proper controls, including RNase and control lines that don't express the lncRNA showed these are mostly bogus
Specifically, RAP-seq is hard to evaluate because it was applied only to Xist, but it seems to consistently recover chrX-mapped reads. For CHART-seq, some studies, and in particular chrX-mapping reads look solid, others less so. Non-Xist studies look worse than Xist ones
November 13, 2025 at 9:26 AM
Specifically, RAP-seq is hard to evaluate because it was applied only to Xist, but it seems to consistently recover chrX-mapped reads. For CHART-seq, some studies, and in particular chrX-mapping reads look solid, others less so. Non-Xist studies look worse than Xist ones
7) CHART/RAP-seq were mostly used for Xist, and perform better, at least for the chrX-mapping reads; and there are a couple of ChIRP-seq datasets that look clean. What sets these apart is unclear. Could be target abundance, could be elusive methodological details.
November 13, 2025 at 9:26 AM
7) CHART/RAP-seq were mostly used for Xist, and perform better, at least for the chrX-mapping reads; and there are a couple of ChIRP-seq datasets that look clean. What sets these apart is unclear. Could be target abundance, could be elusive methodological details.
6) Using odd/even probe sets, which should be standard, and then intersecting helps to a limited extend, as the overlapping regions are scarce and tend to just match both even and odd probes. While some lncRNA-bound regions could be real, it seems they form a small % of the data
November 13, 2025 at 9:26 AM
6) Using odd/even probe sets, which should be standard, and then intersecting helps to a limited extend, as the overlapping regions are scarce and tend to just match both even and odd probes. While some lncRNA-bound regions could be real, it seems they form a small % of the data
5) Mechanistically, we show that the probes don't just randomly pull on DNA, but rather target the ssDNA ends formed by sonication, which offer a region that presumably hybridizes easily with the probes. This results in DNA recovered in a (lnc)RNA indepdent matter.
November 13, 2025 at 9:26 AM
5) Mechanistically, we show that the probes don't just randomly pull on DNA, but rather target the ssDNA ends formed by sonication, which offer a region that presumably hybridizes easily with the probes. This results in DNA recovered in a (lnc)RNA indepdent matter.
4) When looking at recovered peaks, there is strong enrichment (often SNR>20) of short sequence matches (~10 nt) between the probes that should pull on the RNA, and the DNA in the peaks. Hence, "motifs" in the peaks correspond to probe sequences rather than other lncRNA parts
November 13, 2025 at 9:26 AM
4) When looking at recovered peaks, there is strong enrichment (often SNR>20) of short sequence matches (~10 nt) between the probes that should pull on the RNA, and the DNA in the peaks. Hence, "motifs" in the peaks correspond to probe sequences rather than other lncRNA parts
2) Many samples show 10K-100K binding sites, which look like beautiful peaks, but are at odds with lncRNA abundance (~1/cell), and overlap between orthogonal probe sets is often <5% (but significant). 3) In contrast to ChIP-seq HOT regions recurrent accross studies are rare, most are study-specific
November 13, 2025 at 9:26 AM
2) Many samples show 10K-100K binding sites, which look like beautiful peaks, but are at odds with lncRNA abundance (~1/cell), and overlap between orthogonal probe sets is often <5% (but significant). 3) In contrast to ChIP-seq HOT regions recurrent accross studies are rare, most are study-specific
In our deep dive, we found several key insights: 1) There are very relaxed standards in the field. Replicates are rare, standards set in the original papers (2 probe sets, input control) are loosely followed, and effective controls (KO cells, RNase, antisense probes) are very rarely used
November 13, 2025 at 9:26 AM
In our deep dive, we found several key insights: 1) There are very relaxed standards in the field. Replicates are rare, standards set in the original papers (2 probe sets, input control) are loosely followed, and effective controls (KO cells, RNase, antisense probes) are very rarely used
Already ~2015 there was "corridor talk" at conferences that probes pulling on DNA rather than RNA could be a problem, but no systematic analysis was performed, and more and more studies, mostly ChIRP-seq (short and cheap probes) were performed worldwide, resulting in dozens of datasets.
November 13, 2025 at 9:26 AM
Already ~2015 there was "corridor talk" at conferences that probes pulling on DNA rather than RNA could be a problem, but no systematic analysis was performed, and more and more studies, mostly ChIRP-seq (short and cheap probes) were performed worldwide, resulting in dozens of datasets.
These methods are based on using antisense biotinylated probes to pull on RNA of interest, isolating sonicated DNA, and sequencing - analogous to ChIP-seq. The pressure in the field to find the "mechanism" for functional lncRNAs led to dozens of studies using these methods and reporting peaks
November 13, 2025 at 9:26 AM
These methods are based on using antisense biotinylated probes to pull on RNA of interest, isolating sonicated DNA, and sequencing - analogous to ChIP-seq. The pressure in the field to find the "mechanism" for functional lncRNAs led to dozens of studies using these methods and reporting peaks
About four years ago, I was on sabbatical with Martin Vingron, whose lab conducted many amazing benchmarking studies. This motivated us to look at a method that produced many datasets in the lncRNA field, ChIRP-seq (and related CHART/RAP) for mapping DNA/chromatin binding sites of lncRNAs
November 13, 2025 at 9:26 AM
About four years ago, I was on sabbatical with Martin Vingron, whose lab conducted many amazing benchmarking studies. This motivated us to look at a method that produced many datasets in the lncRNA field, ChIRP-seq (and related CHART/RAP) for mapping DNA/chromatin binding sites of lncRNAs
Potentially 24+ more years of hearing "academia is broken"
November 7, 2025 at 4:57 PM
Potentially 24+ more years of hearing "academia is broken"
I'm actually already becoming Dean on December 1st...
November 7, 2025 at 3:14 PM
I'm actually already becoming Dean on December 1st...