11. Same patient, two slides — how much does this number wobble? · Dr. Seok's Tumour-Infiltrating Lymphocytes, First Steps

So far we have been pulling a number out of a map. Now we want to tie that number to patients. One thing has to be asked first.

how much can this number be trusted?

As luck would have it, this data can answer that. TCGA sometimes has two or more slides from the same patient (…-DX1, …-DX2). Same tumour, same moment, same hospital. The only difference is which piece was cut.

The DX1 / DX2 at the end of a slide name marks different sections from the same patient.

Python Runs in your browser

import til

d = til.slides()
pairs = d[d.participant.duplicated(keep=False)]
print("patients with more than one slide: %d" % pairs.participant.nunique())
print()
print(pairs[["slide", "tumor", "til_pct", "tissue_patches"]].head(6).to_string(index=False))

The ranking holds up reasonably (rho 0.73), but the binary verdict flips for about one patient in five.

Python Runs in your browser

import til
import numpy as np
from scipy import stats

d = til.slides()
two = d[d.participant.duplicated(keep=False)].sort_values(["participant", "slide"])
pair = two.groupby("participant").til_pct.agg(list)
pair = pair[pair.map(len) >= 2]

a = np.array([v[0] for v in pair])       # first slide
b = np.array([v[1] for v in pair])       # second slide

print("%d patients" % len(a))
print("median absolute difference  %.2f pp" % np.median(np.abs(a - b)))
print("rank correlation rho        %.3f" % stats.spearmanr(a, b)[0])
print()
for cut in (2, 5, 10):
    flip = np.mean((a >= cut) != (b >= cut))
    print("judged at a %2d%% threshold, %4.1f%% of patients flip" % (cut, 100*flip))

Move the threshold slider. The further a point is from the diagonal, the more that patient disagreed with themselves.

Loading…

Look at the worst case. Pancreatic cancer patient TCGA-XD-AAUI came out at 15.6% on one slide and 65.1% on the other. Same person, same tumour.

Why so different? Three things are mixed together.

The tumour is not uniform inside — this is not a fault in the data, it is biology. Tumours differ region to region (intratumour heterogeneity).
Where it was cut — did the technician section the middle of the tumour or its edge?
How the slide was made — fixation time, stain intensity, scan quality.

① is real signal and ②③ are noise. And we cannot separate the three. Since we cannot, we carry all of it as a single wobble of 1.8 pp.

The eight worst disagreements. Pancreas and lung adenocarcinoma stand out — cancer types with a lot of heterogeneity.

Python Runs in your browser

import til
import numpy as np

d = til.slides()
two = d[d.participant.duplicated(keep=False)].sort_values(["participant", "slide"])
rows = []
for pid, g in two.groupby("participant"):
    if len(g) >= 2:
        rows.append((pid, g.tumor.iloc[0], g.til_pct.iloc[0], g.til_pct.iloc[1]))

rows.sort(key=lambda r: -abs(r[2] - r[3]))
print("%-14s%6s%9s%9s%9s" % ("patient", "type", "1st", "2nd", "diff"))
for pid, t, x, y in rows[:8]:
    print("%-14s%6s%8.1f%%%8.1f%%%8.1f pp" % (pid, t, x, y, abs(x-y)))

🤔 Check yourself

Which reason for two slides of the same patient differing did this lesson not list?

🤔 Check yourself

How does the 1.8 pp measurement error get used in later lessons?