Doctoral application · UCL Cancer Institute

Are all regions of a tumour equally able to adapt?

A response to the PhD project on mapping spatial plasticity dynamics in colorectal cancer. This is my own reading of the question and how I would approach it — written to be argued with, not to flatter the advert. Everything here is a proposal: I have not done this work, and parts of it I would need to be taught.

I tested your premise in public data The proposal
Multi-region organoids Mass cytometry Spatial transcriptomics Four years from 2027
Before applying

I took the question as far as public data allows.

Rather than only proposing an experiment, I reanalysed a public multi-region colorectal single-cell dataset to see what could already be settled. The answer turned out to be sharper than I expected, and the part it cannot settle is the argument for the experiment.

Finding

YAP target expression varies across a tumour; the programme it drives does not

Four public cohorts, 511,077 cells. A paired per-patient core-vs-border contrast finds no significant shift (+0.0220, exact p = 0.125). Two regions of one tumour differ by 9% of the difference between two patients — less than two library preparations of the same tumour differ (56%). YAP target expression does the opposite.

Fetal-like programme9%regional, vs 56% technical floor
YAP target expression65%regional, vs 26% technical floor
Read the analysis

Why that is an argument for your experiment

Occupancy is not capacity. A snapshot cannot tell a population that cannot enter the drug-tolerant state from one that simply has not been pushed into it — and the second is what decides treatment response. No amount of further reanalysis fixes that.

The specificity control rules out the obvious objection. The method plainly can detect regional variation — it detects it in YAP signalling, in the same cells, in the same samples. What it does not detect is regional variation in the state that signalling is thought to drive. That suits a threshold relationship between pathway activity and state commitment rather than a proportional one.

Which is the testable prediction. If signalling varies regionally while resting state does not, then regional differences in capacity should appear under perturbation and in co-culture, and be largely absent at baseline. That is precisely what the design in the proposal below would test.
Why this one

The question is about capacity, not state.

Most work on treatment resistance asks what state a cell is in. This project asks something harder and more useful: what a cell is capable of becoming, and whether that capacity is distributed evenly across a tumour.

That reframing is what makes it a computational problem rather than a cataloguing exercise. A state is something you measure once. A capacity has to be inferred from how a population responds to perturbation — which means the analysis is doing real inferential work rather than describing a snapshot.

  • It is a perturbation problem. Plasticity is only visible when you push the system and watch where it goes, which is exactly what organoids plus mass cytometry are for.
  • It has a built-in control for the microenvironment. Organoids strip the stroma away. That makes cell-intrinsic and environment-induced plasticity separable by design rather than by argument.
  • It joins two data types with different geometries. Mass cytometry gives many cells and no position; spatial transcriptomics gives position and far less resolution. Making them speak to each other is the interesting part.

The biology, as I understand it

Intestinal epithelium can abandon the normal LGR5+ stem-cell programme for a fetal-like, regenerative state during wound healing — a YAP/TAZ-associated programme that is slow-cycling and tolerates cytotoxic chemotherapy, which preferentially kills what is dividing fastest.

Colorectal cancer appears to co-opt that programme. The drug-tolerant population is not necessarily a pre-existing resistant clone: cells can enter the state, persist, and re-enter proliferation afterwards. If that is right, resistance is partly a reversible phenotype, and the quantity that matters is how readily a given population makes the switch.

Which is why region matters. If the capacity to switch is uneven across a tumour, then a biopsy samples one region's capacity and generalises it to the whole. That is a sampling problem with direct clinical consequences.
My proposal

How I would try to answer it.

Written independently, before any conversation with the lab. I expect the design to change on contact with people who know the system — that is rather the point of doing a PhD — but I would rather bring something specific to be corrected than arrive with nothing.

01 Hypothesis

Plasticity — the capacity to enter the wound-healing, drug-tolerant state — varies systematically between regions of a single tumour, and that variation has two separable sources:

  • A cell-intrinsic component that is heritable through organoid derivation, and so persists when the microenvironment is removed.
  • An environment-induced component that is only recovered when the relevant stromal signals are restored in co-culture.

The prediction that makes this testable: regions whose organoids differ in plasticity in pure epithelial culture differ for intrinsic reasons, while regions that converge in pure culture but diverge again under co-culture differ because of where they sat in the tissue.

02 The separating experiment

Derive organoids from several spatially annotated regions per patient, then measure the drug-tolerant transition under a factorial perturbation:

Factor 1Region of origin
Factor 2± chemotherapy
Factor 3± stromal co-culture
ReadoutMass cytometry
Mapped back toSpatial transcriptomics

The co-culture arm is the control that makes the design work. Without it, every regional difference surviving organoid derivation is ambiguous between "intrinsic" and "the microenvironment left a mark that culture has not yet erased". With it, the interaction term between region and co-culture is the quantity of interest rather than a nuisance.

Mass cytometry is the right readout here specifically because it measures signalling and cell-state markers in the same cells at single-cell resolution. A state transition that happens in 8% of cells is invisible in a bulk lysate and obvious in a CyTOF distribution.

03 Defining a plasticity index

The project needs one number per region that means "how readily this population switches". I would define it as an inducible shift in state occupancy rather than as a marker intensity:

  1. Arcsinh-transform, bead-normalise and batch-correct the CyTOF data, with anchor samples carried across every barcoding batch.
  2. Cluster all conditions together into shared cell states, so that states are defined once rather than per sample — otherwise conditions cannot be compared.
  3. For each region and condition, compute the occupancy of the drug-tolerant state as a proportion of cells.
  4. Define the plasticity index as the shift in occupancy between the unperturbed and perturbed conditions, quantified as an earth mover's distance over the state distribution rather than a difference of means, so a change in distribution shape is not discarded.
Why a distributional measure. Mean marker intensity across a population conflates "all cells shifted slightly" with "a subpopulation shifted completely". Those are different biology, and for a question about plasticity they are the two answers you most need to tell apart.

04 Joining CyTOF to space

The two data types have incompatible geometries: dissociated cells with deep marker detail and no position; spots with position and coarser resolution. I would not try to integrate them into one embedding. I would instead use the spatial data to ask a targeted question of the regions themselves:

  • Deconvolve each spot into cell-type proportions, with the matched organoid and stromal profiles as the reference, so the reference is from the same patient rather than an atlas.
  • Characterise each sampled region's neighbourhood: which cell types co-occur around the epithelium, at what density, and how close to the invasive front.
  • Test the association between a region's measured plasticity index and its neighbourhood composition, treating the index as the outcome and the spatial features as predictors.

That keeps each modality doing what it is good at. The CyTOF measures capacity; the spatial data explains where that capacity sits and what was next to it.

05 Statistical design

The failure mode I would most want to avoid is treating wells as replicates. Regions are nested within patients, and patients are the level at which a finding generalises.

  • A mixed-effects model with patient as a random effect and region nested within it, so between-patient variation does not masquerade as a regional effect.
  • Technical replicates modelled as such, never pooled into the sample size.
  • Batch structure recorded in the design and tested against the biological factors before analysis — if batch is confounded with region, the experiment cannot answer the question and it is better to find that out at the design stage.
  • Power considered in terms of regions per patient and patients per arm, since with a handful of patients the honest limit is a well-characterised case series rather than a population claim.

06 How it could fail

The ways I would expect this to go wrong, in order of how much they would worry me:

  • Organoid drift. Derivation and passaging select. A regional difference could be a difference in how well each region's cells took to culture rather than a property of the tumour. Mitigation: low-passage work, derivation efficiency recorded as a covariate, and early-versus-late passage comparison treated as a result rather than a check.
  • Batch confounded with region. If regions are processed in the order they are dissected, batch and region are the same variable. Mitigation: randomise the barcoding layout across regions and patients.
  • The state is a continuum, not two states. If cells occupy a gradient, clustering imposes a boundary and "occupancy" becomes an artefact of where it was drawn. Mitigation: report the index across a range of boundaries and show the conclusion does not depend on one.
  • Deconvolution error. Spots contain several cells, so neighbourhood composition is inferred rather than observed. Mitigation: validate against imaging where available, and treat weakly deconvolved spots as missing rather than as zero.
  • Too few patients. Regional variation within a patient can be real while being swamped by variation between patients. That is a finding and should be reported as one, not hidden behind a pooled analysis.
Fit, honestly

What I bring, and what I would have to learn.

What I bring

High-dimensional analysis disciplineClustering, batch effects and the resolution-parameter problem, written up in the methods library Laboratory experience, in chemistryThree years of synthesis and characterisation — I know what it costs to generate data Software that other people runTwo production systems, with databases, tests and CI A habit of reporting what breaksSeveral published analyses exist mainly to document where a method fails

What I would have to learn

Organoid cultureNo experience. Derivation, passaging and co-culture would all be new Mass cytometry, at the benchI understand the analysis; I have never run a panel or barcoded a batch Spatial transcriptomics in practiceProposed in my own research plans, not yet executed Working with clinical samplesGovernance, consent and the realities of tissue from theatre

A medicinal chemistry graduate is not the obvious candidate for a cancer biology PhD. The case I would make is that the computational and experimental halves of this project are both learnable, and that I have evidence of learning hard technical things unsupervised — but the biology gap is real and I would not pretend otherwise.