Lesson 20 of 20
What you can say, and what you cannot
from one map to here
In lesson 1 we opened one map. Three colours of pixel, red for lymphocytes. Since then, over twenty lessons, we have been making one decision at a time.
| Lesson | What was decided | Decide otherwise and |
|---|---|---|
| 2 | tissue tiles as the denominator | the ranking flips |
| 3 | where to draw the threshold | the value moves up to 700× |
| 4 | 50 µm tiles | 45.7% becomes 88.8% |
| 10 | the threshold for "infiltrated" | the count of brisk patients doubles |
| 12 | one slide per patient | independence breaks |
| 15 | where to cut | a result appears where there was none |
| 17 | within which group you compare | the sign flips |
Most of these decisions never appear in the body of a paper. One line in the methods, or only in the supplement. And yet they are what the results are made of.
5,201 maps, a total tissue area of about 116 m² — we counted roughly half a tennis court of tissue.
Python
Runs in your browser
What you can say
- The distribution of infiltration differs markedly between cancer types. From 0.0% in uveal melanoma to 12.3% in stomach cancer. It matches known histology well.
- Within the same cancer type, patients with more infiltration tend to live longer. Stratified log-rank z = −3.97 (p = 7.3×10⁻⁵).
- Only two individual cancer types keep that tendency after correcting for multiple testing — melanoma (q = 1.4×10⁻⁴) and cervix (q = 0.024).
- Tissue size is not a confounder in this data. ρ = 0.047.
What you cannot say
- "This patient's TIL is 5%" — take a different slide from the same patient and the verdict flips 22.8% of the time.
- "High TIL means better prognosis" (without naming the cancer type) — measured pooled, the sign flips (z = +3.75).
- "Patients from site A have higher infiltration" — staining and scanning cannot be separated from patient mix (p = 4.6×10⁻⁷).
- "Infiltration lengthens survival" — there is no randomisation in this data. What was observed is association, not causation.
- "Lymphocytes got inside the tumour" — the map has no tumour boundary, so we can only say "inside the tissue".
What you could do next
- Cox regression — the infiltration rate uncut, as a continuous value, alongside age, stage and cancer type. The straightest road around the trap of lesson 15.
- Spatial statistics — handle the clusters and depth of lessons 8 and 9 properly and you can approach ideas like the "tumour-immune boundary". But as lesson 18 showed, they are sensitive to the model version.
- Joining other data — expression and mutation data for the same TCGA patients are public. Dr. Seok's Gene Search, First Steps and Single-Cell Immune Atlas, First Steps take that road.
- Validation — take a result from here and look again in another cohort. It is more powerful than any technique this course taught.
The last picture. The red dots are the cancer types with p < 0.05.
Python
Runs in your browser
🤔 Check yourself
What did this course teach, in one sentence?
Answer Denominator, threshold, tile size, cut point, comparison group — these decisions are what the result is made of. Not knowing them, you cannot read the number.
🤔 Check yourself
How would you honestly rewrite "patients with high TIL live longer"?
Answer Within which group (cancer type), measured how (this way), and what was observed (a tendency, not causation) — naming those three is what makes the sentence honest.
🤔 Check yourself
What did this course name as the first thing to do in the next study?
Answer Replication is more powerful than any statistical technique. The traps of lessons 15 and 16, and the batch effects of lesson 18, mostly reveal themselves when you look again in independent data.