iiinova — where your digital self keeps evolving

Lars Barquist @lbarquist.bsky.social · 04/02/2025

This is all enabled by a robust infrastructure that can currently support ~500 genome-wide experiments, and allows for easy development of new visualization plug-ins

Schematic of the Micromix architecture, including MongoDB, a Python backend communicating through a Flask server, and a Vue.js frontend

110

Lars Barquist @lbarquist.bsky.social · 04/02/2025

This lets you easily create figures like this, showing the expression of genes in the Salmonella SPI-2 pathogenicity island across multiple conditions, as well as fitness effects in a variety of animal models (interactive server session to play with here: micromix.helmholtz-hiri.de/salmonella/?...)

110

Lars Barquist @lbarquist.bsky.social · 04/02/2025

At its core, Micromix is a graphical web interface for Python dataframes, letting you mash together and filter different datasets, then visualize the results.

figure illustrating merging, transforming/filtering, and chaining operations together to create a heatmap

110

Lars Barquist @lbarquist.bsky.social · 19/11/2024

Even more worryingly, there was one case where a mutant was predicted as cipro resistant before correction, but was actually sensitive! Our correction method recovers the correct phenotype.

100

Lars Barquist @lbarquist.bsky.social · 19/11/2024

But does it matter? Yes! In a test data set, we see a lot of mutants predicted to be cipro resistant before correction, but actually have a WT MIC.

110

Lars Barquist @lbarquist.bsky.social · 19/11/2024

So, we calculate local normalization factors, plug in our offsets, and the origin bias is gone and we still have accurate p-values based on the original counts!

An image showing differential fitness analysis before and after correcting for biases due to differences in chromosomal replication between samples

100

Lars Barquist @lbarquist.bsky.social · 19/11/2024

But this raises issues in statistical testing. The way standard analysis packages like edgeR and DEseq deal with having few replicates is by sharing information between genes: genes with similar counts are assumed to have similar variances.

A figure showing the mean-variance trend in RNA-seq data, from the limma/voom package

100

Lars Barquist @lbarquist.bsky.social · 19/11/2024

This is particularly a problem for TraDIS, as we’re using DNA copy number as a proxy for mutant abundance. For RNA-seq, this reflects real differences in RNA content, but may mask regulation or interfere with normalization. We observed this in a range of published datasets.

A figure showing bias in the chromosomal location of large fold-changes for a range of antibiotic treatments, bacteria, and functional genomics methods (e.g. RNA-seq, Tn-seq)

100

Lars Barquist @lbarquist.bsky.social · 19/11/2024

The problem for sequencing based assays occurs when you treat with something that stalls replication; the bacteria keep initiating replication, leading to a local increased chromosomal copy number in the vicinity of the ori

A figure showing how antibiotics that interfere with replication, like ciprofloxacin, can lead to more copies of the chromosome around the origin of replication

100

Lars Barquist @lbarquist.bsky.social · 19/11/2024

You see, bacteria have a problem during exponential growth: cell division takes ~20 minutes, but chromosomal replication can take up to 60 min to complete. They solve this by simultaneously producing multiple copies of their genome.

A schematic figure showing that bacteria initiate multiple rounds of chromosomal replication simultaneously during exponential growth

100

Lars Barquist @lbarquist.bsky.social · 19/11/2024

So, this all started when Amy's student Geri noticed a weird pattern in her data when applying TraDIS to look at E. coli genes involved in ciprofloxacin sensitivity: we saw a huge peak of extreme log fold-changes around the origin of replication!

A plot showing a strong peak of large log2 fold changes spatially localized in the chromosome

100

Lars Barquist @lbarquist.bsky.social · 16/11/2024

Finally, we looked at the relationship between transcript stability and RBP binding — we find that at least half of the 500 most stable transcript are bound by RBPs, and they tend to be destabilized by RBP deletion.

two plots, the first showing that transcripts with longer half-lives are more likely to be bound by RBPs, and a second showing that transcripts with longer half-lives have larger changes in stability following RBP deletion

120

Lars Barquist @lbarquist.bsky.social · 16/11/2024

… and we were able to show that this translated into a sensitivity to hydrogen peroxide stress in a ProQ deletion strain

Dilution assay showing a lack of growth of the dproQ strain in comparison to wild-type, proQ complement, and doxySR strains after exposure to 2mM hydrogen peroxide

110

Lars Barquist @lbarquist.bsky.social · 16/11/2024

We were also able to connect changes in stability to phenotypes — doing some GO term analysis of destabilized transcripts showed enrichment for transcripts involved in the oxidative stress response…

A gene set analysis plot, showing an enrichment in differential destabilization of oxidative stress-associated mRNAs in a dproQ background with no corresponding enrichment in steady-state differential expression

100

Lars Barquist @lbarquist.bsky.social · 16/11/2024

We were able to combine our stability analysis with CLIP-seq to see which RBP binding sites do lead to destabilization — for ProQ, which was known to bind 3’ UTRs, only a minority of bound transcripts are destabilized

A plot showing while ProQ binds preferentially to 3' UTRs, only a minority of these transcripts are destabilized in a dproQ background

110

Lars Barquist @lbarquist.bsky.social · 16/11/2024

Some striking findings: differences in stability are not correlated with differences detected by simple RNA-seq. We think this is because deletion of these global regulators leads to downstream effects on the transcriptome that don’t have much to do with direct regulation

Scatter plot showing a lack of correlation between steady-state differential expression and differential decay in a dproQ background compared to wild-type.

100

Lars Barquist @lbarquist.bsky.social · 16/11/2024

Now that we had a model, we could go back to our original question — what transcripts are differentially destabilized when global RBPs are deleted. Here’s an example for cspE, a known target of the RBP ProQ.

Decay curves for the cspE transcript, showing it is dramatically destabilized in a dproQ background

110

Lars Barquist @lbarquist.bsky.social · 16/11/2024

Does this make a difference in inferring half-lives? Yes! By properly fitting the data, we found that bacterial half-lives are about 3 times shorter than previously thought — on average less than 1 minute!

Figure showing comparisons between RNA decay rate distributions inferred by 3 different models, showing that a model that accounts for confounding factors leads to dramatically lower inferred half-lives

121

Lars Barquist @lbarquist.bsky.social · 16/11/2024

This wasn’t just an artifact of our sequencing protocol — you can see this in radioactive probing data as well, and it’s all over the place in the literature once you know what to look for.

An RNA decay curve showing matched RNA-seq and northern blot data, showing both curve flatten at late time points

220

Lars Barquist @lbarquist.bsky.social · 16/11/2024

To our surprise, during model development we found that transcripts don’t keep decaying indefinitely — they hit some sort of baseline, meaning that the simple linear or piece-wise linear models that have been used before don’t really fit the data very well.

A figure showing the structure of our Bayesian model, colored so parameters match the segments of a corresponding RNA decay curve that they dominate

130

Lars Barquist @lbarquist.bsky.social · 16/11/2024

To investigate this, we adapted a protocol where we treat with an RNA polymerase inhibition, then watch transcripts decay over time. To analyze the data my postdoc Laura Jenniches began building a Bayesian model in Stan, a probabilistic programming language

Experimental and computational overview for a RIF-seq experiement

110