For full paper:
doi.org/10.7554/eLif...

In our work, we introduce REPOP, a Bayesian computational framework that more accurately quantifies bacterial populations from plate counts by modeling the experimental noise introduced through dilution and plating.

6/6

✅ Simulate complex, non-Markovian biological dynamics
✅ Train conditional normalizing flows to approximate intractable likelihoods
✅ Perform full #Bayesian inference on anything you can simulate.

arxiv.org/abs/2506.09374

5/6

We apply this to yeast expressing GFP under the glc3 promoter.

🌱 At first glance, high fluorescence seems like gene activation. But when you model protein inheritance across divisions...

Most cells are actually inactive — just glowing their ancestors GFP.

4/6

Despite the complexity, these dynamics are easy to simulate — protein production, cell division, fluorescence, all of it.

So we flipped the problem: We train neural networks on simulations to learn the likelihood function itself.

3/6

Because of that clock, division times aren’t memoryless -- they’re not exponential.

This breaks standard models of gene expression, that is:
NO Master Equations
NO Fokker-Planck equations

We had rethink how we do inference.

2/6

In this new preprint, we analyze #flowcytometry data of stress regulation in yeast
🧬 We indirectly observe protein levels through fluorescence.
But here's the catch:
1 - Proteins live much longer than a single cell cycle
2 - Cell division follows a biological clock

6/6
If you plate, you need REPOP.

Preprint -- doi.org/10.1101/2025...
Software -- github.com/PessoaP/REPOP

Special thanks to the Lab Members - Pedro Pessoa, Carol Lu and Stanimir Tashev
As well as Rory Kruithoff and Douglas P Shepherd
#Biophysics #QuantitativeBiology

5/6
This is why we built REPOP, an #opensource tool to REconstruct POpulations from plates.

Straightforward to use and with tutorials available on #GitHub

github.com/PessoaP/REPOP

With all the #Bayesian Rigor and #PyTorch speed

4/6
As we show in the preprint, this
- Overestimatese variability
- Can miss real structure in your population: Subpopulations and/or multimodality as biological differences across samples,

3/6
This assumes:
– No randomness in how many bacteria end up on the plate
– No randomness in the original swab

In reality, every step is noisy.

2/6
Plate counting is a simple:

You dilute a sample, plate a small volume, and count colonies.

Say you dilute by 200×, and count 50 colonies.
Easy just multiply 50 × 200 = 10k bacteria, right?

NOT QUITE...

But how do we know how accurate our estimate of π really is? 🤔

There’s a way to do it right: Combining it with Bayesian inference. Instead of just getting a rough guess, we can properly quantify uncertainty.

That is what I have written in my blog today. Check it out