I’m grateful to my co-authors and the Janes lab @lungs4living.bsky.social for their support over the last 4 years, as well as the patient donors and clinical teams on whose data this work relies.
If you have any questions about the pre-print, do reach out! doi.org/10.1101/2025...
[15/15]

Finally, this model is a rapid perturbable system for the accumulation of genomic damage in the presence of carcinogens and immune modulation. This can be applied for in silico trials of preventative therapeutics, or to other cancers with a similarly dominant causal mutagen.
[14/15]

Lots more can and should be done to further validate this picture: spatially resolved data would give us another viewpoint, and direct in vivo and in vitro interventional studies are, as ever, crucial. Co-evolution of lung and immune cells would be likely here and could be measured!
[13/15]

So what does this mean for our understanding of the upper airway? These results paint a tentative picture, in which smoking both causes mutational damage and suppresses an immune mechanism of recovery. Quiescent cells, always present, then provide a competitor to take over after smoking.
[12/15]

Adapting these same classifiers to the observed scWGS data, we found that distinct machine learning classification methodologies converged on the same answer: a quiescent subpopulation of cells, and smoking-suppressed immune predation of mutated stem cells.
[11/15]

When tested on unseen simulations, the classifiers were very accurate on some problems (identifying a quiescent subpopulation of cells) and were pretty hopeless at others (changes to the fitness landscape in the presence of smoke), giving us a good idea of what we can learn from this data.
[10/15]

An aside: these big sets of lifelong simulations rack up a long time in simulated somatic evolution - almost 70 million years. To run this in vivo and in series, we'd have to have started in the Cretaceous! A good advert for efficient code and in silico modelling...
[9/15]

Qualitative fits are nice, but with this data and this modelling approach we can go further. We created a collection of synthetic datasets from literature-informed priors, and trained machine learning classifiers to identify which hypotheses were active in a particular dataset.
[8/15]

First, we had to check that these hypotheses could reproduce the surprising findings: they could! Different combinations of hypotheses looked qualitatively similar to what we saw in the scWGS data, in quite different ways (spatial clonal patterning completely different between paradigms!)
[7/15]

We modelled this on the natural 2D spatial structure of the upper airway epithelium. We added in each mechanistic hypothesis modularly, so that we could run simulations with any combination of them.
[6/15]

To answer these questions, we implemented an established model for this stem cell population to simulate a small population of lung stem cells for each patient over their entire lifetimes, based on their smoking histories.
[5/15]

We collated a set of mechanistic hypotheses about the upper airway epithelium that could explain these dynamics. It seems reasonable that any of these could lead to the observed dynamics, but is it true? And which of them is most likely given the observed data?
[4/15]

Where it gets interesting is in the variation within an individual's lung stem cells. Some cells from lifelong smokers look like they come from a never-smoker’s lungs, and there’s more of these less-mutated cells in the lungs of ex-smokers than ongoing smokers!
[3/15]

This work started with two surprising observations in recently published single cell-derived whole genome sequencing (scWGS) data from stem cells in the upper airways of patient donors. As you'd expect, the data showed an accumulation of mutations over lifetime, accelerated by smoking.
[2/15]

Congratulations Alex!! 🥳