iiinova — where your digital self keeps evolving

We then tested if this assosiation holds through our entire dataset. We used a functional assay for drug inactivation on all drugs and found the association holds up. A long-lag inhibition phenotype is a strong indicator of drug inactivation 7/9

That pushed us to ask whether cellular defenses might impact curve profiles. We cloned different resistance cassettes and measured how they altered the potency-matched growth curves. This strongly hinted that active drug inactivation underlies a long-lag inhibition profile 6/9

111

Overlaying the known mechanisms of action over the barycentric landscape ruled out this effect stem exclusively from how drug target bacteria (since drugs with the same mechanism can land in very different regions of the landscape) 5/9

Clustering drug by their impact on lag/rate/yield clearly revealed that they vary hugely in how they inhibited growth. In extreme cases, a drug solely affected only a single parameter 4/9

To compare drugs fairly, we didn’t use an arbitrary concentration. Instead, we interpolated each drug to a potency-matched condition (the concentration expected to produce the same overall level of inhibition) 3/9

So we assembled a new carefully curated dataset with growth curves across almost forty drugs, measured across multiple sub-inhibitory concentrations. For each curve, we quantified intuitive its key features: lag, growth rate, and yield 2/9

Beyond providing (to our knowledge) the first dynamical model for tumor colonization, our study matters given the fierce debate on the tumor microbiome. These statistical “fingerprints” may help distinguish genuine colonizers from technical artifacts/contamination (possibly even by microscopy) (6/7)

The surprise: lineage sizes formed a scale-free power law that matches Zipf’s law (rank–frequency slope ~−1). This signature was robust across dozens of tumors and multiple collection days post bacteria intratumor injection (5/7)

130

When we injected bacteria directly into the tumor (circumventing the bottleneck), we detected thousands of colonizing lineages, yet their sizes were still highly uneven (ruling out early tumor arrivers dominate) (4/7)

120

Since we used genetically barcoded bacteria, we could also monitor growth of individual colonizers. We found that growth was extremely uneven with a handful of lineages becoming dominant (“winner-takes-most”) (3/7)

Main takeaways: Post systemic infection, there's a tight colonization bottleneck (per-cell colonization probability ~0.005%). Yet, once colonization happens, growth is remarkably fast (~50 min generation time) and bacterial load in tumors approaches saturation within a day (2/7)

120

Amir Mitchell @amitchell.bsky.social · 17/01/2025

We just published in @molsystbiol.org with the Mugler lab (UPitt) on bacterial population dynamics during tumor colonization (mouse model). Our study was guided by a Luria–Delbrück-style idea: infer mechanism from statistics (1/7) 🧪🦠 doi.org/10.1038/s443...

21812

Interested in Antimicrobial Resistance & Quantitative Bio? Join a 3-day meeting+workshop with some of my favorite #AMR #QBio scientists this April (University of Exeter, UK). Register here: tinyurl.com/QAMR2025reg 🧪🦠

051

Amir Mitchell @amitchell.bsky.social · 07/12/2024

My only complain is that I feel a bit like Cassandra, from the greek mythology, barely anyone listens 😞

010

Amir Mitchell @amitchell.bsky.social · 07/12/2024

The ability of LMMs (chatGPT here) to *assist* in generating useful code for biologists a mind-blowing power multiplier 🤯! My advice to biologists at all career stages: Learn to code, its a superpower 🦸‍♀️🦹‍♂️ (simple example, counting fluorescent colonies isolated from mouse gut microbiome) 🧪🦠

Microscopy image of fluorescent colonies of E. coli after automatic image segmentation

1234

Careful quantification revealed that the damage decay curves are identical in contacting and non-contacting colonies. Therefore, direct contact is not only not required, but it doesn’t even increase the level of toxicity beyond what is expected by proximity (7/9)

Traces of DNA-damage reporter decay with distance from colibactin producers

100

Next, we monitored colibactin damage in separated colonies. This setup allowed us not only to validate contact independence, but also to accurately quantify how DNA-damage decays across distance (6/9)

Experiment setup for monitoring colibactin induced DNA damage in neighboring colonies.

100

Microscopy imaging revealed that DNA damage is observed even hundreds of microns (YFP halo) away from the secreter front (mCherry) suggesting that direct cell-cell is not needed for toxicity (5/9)

Diagram of experiment setup. Colony of a colibactin producing cell is placed on a lawn of DNA-damage reporter cells. The yellow halo around the colony marks DNA damage

100

We cloned a YFP DNA-damage reporter in E. coli and tested how far colibactin induced damage “travels” across a lawn of cells (we tagged secreters and responders with constitutive mCherry and CFP to tell them apart) (4/9)

Diagram showing the plasmids used for cloning the colibactin producing strain and the DNA-damage reporter strain

100

Amir Mitchell @amitchell.bsky.social · 22/10/2024

Our second paper on the bacterial toxin colibactin is now out on mBio, this time we critically evaluated the claim that colibactin toxicity necessitates cell-cell contact 🧪🦠🧵(1/9) journals.asm.org/doi/10.1128/...

1278

Can’t believe we unlocked this level of the #AI hype - huge adds for @AnthropicAI in airports (Boston & Atlanta). Wonder what % of ppl even know what’s the product (<0.1% probably)

010

As a validation of our hypothesis on self-inflicted damage, we genetically engineered a fluorescent reporter of DNA damage and found its basal activity is indeed noticeably increased in colibactin producing cells (either engineered or natural strains)

100

Exploiting the mut signature we discovered, we scanned 10K E. coli genomes in search for compatible genomic scars. Surprisingly, we found that such scars are enriched in strains that produce colibactin. This indicates that producers cannot(!) avoid some self-inflicted damage

100

Yet, although in bacteria colibactin favored binding A/T rich oligos, just like in humans, colibactin damage led to to a bacteria-specific mutation profile – mostly T->A (rather than T->C)

100

We then tested the genomic changes colibactin induces with a mutation accumulation experiment (in a 100 parallel replicates). We repetitively exposed bacteria to colibactin producing cells and then sequenced their genomes (#WGS) to infer the mutational landscape

100

We first systematically mapped all genes modulating colibactin toxicity in targeted bacteria with a loss-of-function genetic screen and found that multiple pathways, beyond nucleotide excision repair, influence colibactin sensitivity (HR, nuc synthesis, stringent)

100

Past works showed that colibactin cross-links DNA and suggested that 5-10% of colorectal cancers may arise from colibactin producing E. coli in the gut #microbiome. We set out to study how colibactin damages bacteria that may encounter it in the gut niche.

100

Amir Mitchell @amitchell.bsky.social · 03/09/2024

We found that the bacterial #toxin colibactin acts as a double-edged sword and inflicts self-damage (evident in hundreds of genomes). Our recent collaboration with @DaganLab is now out on @genomeresearch genome.cshlp.org/content/34/8/1154.…

186

Immensely proud of the latest graduate from our lab @emily_lowry23⁩ 🎊🎉🥳⁦. Dr Lowry’s second paper will be published *very* soon … (and hopefully her third paper shortly after that 🤞)

000

Amir Mitchell @amitchell.bsky.social · 29/08/2024

What makes #phd s so hard & depressing (yet also so rewarding), a thread started by @ThePhDPlace made me reflect about my own experience and my students’ journeys. Hoping this self-reflection thread will be useful for students (and PIs too!) 1/9

Amir Mitchell @amitchell.bsky.social · 07/08/2024

Results are in, mostly saying free (academic) labor is reasonable (seriously?🤦‍♂️). Going forward, I'll review 2 of every 3 such grants I'm asked about 🫡

000

Amir Mitchell @amitchell.bsky.social · 28/06/2024

Signing off a very exciting week at the co-chair of the @GordonConf on Drug Resistance. Hard, yet totally worth it! Lots to share with chairs-to-be (thread on Monday)

010

Amir Mitchell @amitchell.bsky.social · 26/06/2024

First time I'm attending a conf with 2/3 of my lab (@GordonConf / drug resistance). Strongly recommend PIs to do that, costly, *but* a really great bonding experience.

000

Amir Mitchell @amitchell.bsky.social · 16/05/2024

Deadline for @GordonConf on Drug Resistance is fast approaching (5/26). Apply by 5/21 for posters and talks. We’ll greatly appreciate help spreading the word 🙏🏼 www.grc.org/drug-resistance-confere…

Many people are behind this success, but this is primarily the triumph 🏆 of @MarianaNoto who confronted a challenge that seemed truly intractable (study the MOA of almost 200 drugs) (18/19)

Even more exciting, resistant IF2 mutants also resisted two additional non-ABX drugs that were network neighbors of the drug used for evolution (pink). This means our network can uncover shared MOA even if the MOA is unknown! (16/19)

100

Lastly, we tested our conclusions by evolving resistance against 3 non-ABX. The results supported our predictions: resistance to one non-ABX drug increased resistance against multiple ABX (by mutating regulators of pumps) (14/19)

100

We also saw increased sensitivity to both ABX and non-ABX with K/O of efflux pumps (used for cell detox). This was the 3rd key observation 🔑: despite diff. in toxicity mechanisms, non-ABX may lead to ABX resistance if they select for high-expression of efflux pumps (13/19)

When we focused exclusively on non-ABX, clusters did emerge in the drug-similarity network. This was the 2nd key observation 🔑: Many non-ABX likely share a MOA between them (12/19)

When added results from screen of non-ABX to the mix ▲◼️, almost all of them turned out to be disconnected from the ABX hubs. This was a key observation 🔑: non-ABX likely kill bacteria differently from standard ABX (10/19)

100

We then inferred a drug similarity network (drugs are nodes, connection mark drugs with similar impact on the K/O library). Using a force layout algorithm showed the ABX self-organize into hubs by their known classes 🟣🔴🟡🟢 (validating our approach worked) (9/19)

We started with simple machine-learning approaches to explore only ABX screens since their MOA is known. The results were crystal clear 💎: K/Os we found made perfect sense and ABX clustered by their known classes 🟣🔴🟡🟢 (8/19)

The idea was “straightforward”, grow all single-gene knockout (K/O) with each drug and find K/Os that are drug-sensitive/resistant to identify the killing mechanism (e.g., if K/Os of cell-wall synthesis genes are extra-sensitive, the drug likely targets the cell wall) (7/19)

100

We applied a pooled-genetic screening approach we developed to study how anti-cancer chemotherapies kill bacteria (previously published @eLife), this time around at a MUCH HIGHER throughput. These screens yielded ~ TWO million data points to analyze 🙀 (6/19)

BUT, there were dozens of toxic non-ABXs, so which ones should we focus on 🤷🤷🏽‍♀️🤷‍♂️? Being a data-hungry lab 😋, we decided to study ALL non-ABX we found to be toxic for E. coli (~200 ABX and non-ABX drugs) (5/19)

We became obsessed with figuring out HOW such non-ABX drugs kill bacteria 🧐? Or in more “sciency” terms, what is their mode-of-action (MOA) (3/19)

5-years ago, we read a landmark paper that blew our minds 🤯! Lisa Maier from the @TypasLab found that 25% of non-ABX drugs (used to treat cholesterol, anxiety, cancer, etc) can kill bacterial species from the #microbiome (2/19)