How TP53 loss can leave CDKN1A (p21) switched off, seen through five kinds of data, from a patient cohort down to single cells.
After DNA damage, the protein p53 switches on the gene CDKN1A, whose product p21 stops the cell cycle until the damage is repaired. Many tumors have lost that brake. The gene for p53, TP53, is mutated in 37% of the real tumors in View 2, the most often mutated of the known cancer genes. But a tumor can stop making p21 in more than one way, and each kind of data sees a different part of the problem. A cohort grid shows which genes are altered in which patients. A protein plot shows where on p53 the mutations fall. The promoter shows whether the start of CDKN1A is methylated. A pathway diagram shows how the signal is wired. Single cells show which cells in a tumor actually lost the transcript.
Some of what follows is real and some is simulated, and every caption says which. The TP53 mutation counts are from 10,443 real tumors. The promoter is the real human genome sequence, with its real CpGs and p53 binding sites. The pathway edges come from biochemistry and knockout experiments. The 100-patient cohort, its methylation and the single-cell experiment are simulated. The cohort is the same one that part 6 clusters and part 8 follows over time.
Each tumor carries its own set of somatic alterations: point mutations, truncations, copy-number gains and losses. The oncoprint compresses a cohort into one grid, with patients as columns, genes as rows and a mark wherever a patient's gene is altered. The bottom row adds one measurement the grid cannot infer: how much CDKN1A mRNA each tumor makes.
Figure 1. The series' shared simulated cohort: 100 patients by 10 genes, with CDKN1A mRNA (log2, colored by distance from the cohort median) underneath. Alteration frequency per gene on the right. Hover or tap a column, or focus the grid and use the arrow keys, for one patient's details. The columns start in enrolment order and sort themselves once the figure comes into view.
Sorted by TP53 status, the grid shows two things. TP53 is altered in 40 of 100 patients, and MDM2 is amplified in 12, none of whom has a TP53 alteration. MDM2 marks p53 for destruction, so amplifying it is a second way to switch p53 off, and a tumor seldom needs both. In real tumors the two rarely coincide (in one survey, 29 of 33 MDM2-amplified tumors had normal TP53); in this simulation they never do. CDKN1A itself is altered in 1 patient.
The mRNA row shows what the grid cannot explain. Of the 48 patients with neither a TP53 alteration nor an MDM2 amplification, 11 still make less CDKN1A than the cohort median. Nothing in their columns says why. View 3 finds the reason.
The oncoprint marks a patient as TP53-altered without saying where in the gene. The lollipop plot lays the 393-residue protein out as a line. The horizontal axis is the codon, shaded blocks are the protein's domains, and each stem's height is the number of mutations at that codon. These stems are real: 4,225 TP53 mutations in 3,839 of 10,443 tumors from the 32 cancer types of the TCGA PanCancer Atlas.
Figure 2. Real TP53 somatic mutations per codon in the TCGA PanCancer Atlas (cBioPortal). Domains, and the four zinc-binding residues marked under the bar, from UniProt P04637. Hotspot heads are colored by class, structural or DNA contact (Olivier and colleagues, 2010). The ticks along the bottom are the 40 simulated mutations of Figure 1's cohort. Hover, tap, or focus the plot and use the arrow keys to step through codons.
Missense mutations, which swap one amino acid, crowd into the DNA-binding domain (97% of them) and pile up at a few codons. The tallest stems are R273, R248 and R175, and R175H is the commonest single change, in 150 tumors. Those hotspots break p53 in different ways. R248 and R273 touch DNA directly, so a mutation there removes a contact while the fold can stay largely intact. R175 sits next to the zinc ion that holds the DNA-binding surface in shape (two of the zinc's four ligands are residues 176 and 179), and R175H unfolds that region. Switch to truncating mutations and the picture changes: they spread along the whole protein, because any early stop codon destroys it, and only 62% fall in the DNA-binding domain.
Mutant p53 often builds up in tumor cells, which is why strong p53 staining in a biopsy often signals a mutant rather than an active pathway. One reason is that the mutant cannot switch on MDM2, whose protein degrades p53. That reason is not enough on its own. In mice carrying the equivalent of R175H, the mutant protein stays unstable in normal tissue and builds up in some tumors, or when MDM2 or p16 is lost (Terzian and colleagues, 2008), so stress signals in the tumor do part of the stabilising.
A gene's output can also drop without any mutation. Like most human genes, CDKN1A starts inside a CpG island, a stretch of DNA unusually dense in the two-letter sequence CG. A methyl group on the C of those CGs can silence the promoter, and methylation of the CDKN1A promoter has been found together with low p21 mRNA in acute lymphoblastic leukaemia (Roman-Gomez and colleagues, 2002). The β value at a CpG is the methylated fraction of its copies in a sample, from 0 to 1.
Figure 3. The real CDKN1A promoter, hg38 chr6:36,676,114-36,680,713, numbered from the Ensembl canonical transcription start (+1 = chr6:36,678,714; no position 0). Top: the two p53 response elements at their hg38 positions, the UCSC CpG island, exon 1, and a tick for each of the 224 CpGs in the sequence. Stems: mean β at each CpG for the chosen route's patients, with the intact tumors' profile as a line. Heatmap: every cohort patient's β at every CpG, grouped by route; hover, tap or focus it and use the arrow keys. The positions are real; all β values are simulated, and drawn only where the real sequence has a CpG.
The p53 sites and the island barely meet. p53 binds 2,265 and 1,383 bases upstream of the start, in CpG-poor DNA: the 2 kb before the island holds only 16 CpGs, and the island begins 1,117 bases past the nearer site. So the routes act on different stretches of DNA. In the methylated tumors p53 still sits on its sites, but the island around the start is closed. In the TP53-mutant tumors the island is open and the sites are empty. The upstream CpGs are methylated in every patient, as CpG-poor DNA usually is, and neither p53 site contains a CpG, so that methylation is not in p53's way.
The methylated patients are not all alike. 6 of the 18 are methylated only partly, at about half the level, and each patient's methylation fades out at its own distance into the island, between +820 and +1,830. All 11 low-CDKN1A patients that Figure 1 could not explain are in this group. A patient like that looks normal in the oncoprint and still makes little p21.
The first three views are measurements. The fourth is a map: a pathway diagram whose edges come from biochemistry and knockout experiments, not from any cohort. It holds the loop that View 2 relied on. MDM2 destroys p53, and p53 switches on the MDM2 gene, so each keeps the other in check. DNA damage breaks the loop: the kinase ATM phosphorylates both proteins, MDM2 lets go, and p53 builds up and switches on its targets.
Figure 4. The core of the p53 damage response, with edges from the literature. Arrows activate; bars inhibit. Lit edges carry signal at the moment shown; dashed edges with a cross are blocked. The order in which the signal spreads comes from propagating simple rules along the edges, not from a script. The R175H switch removes p53's ability to switch genes on, MDM2 included; any new functions the mutant gains are outside this diagram. Click, tap or focus a node for a description.
With R175H, damage still stabilises p53, but none of its three targets responds: no p21 for arrest, no BAX for apoptosis, and no new MDM2 to end the response.
The pathway predicts a readout: in TP53-mutant cells, CDKN1A expression should fall. A tumor is not one kind of cell, though. A TP53 mutation in a tumor is somatic: the malignant cells carry it, while the immune cells, fibroblasts and blood vessels around them keep the patient's normal TP53. Single-cell RNA sequencing measures each cell separately, so the normal cells in the same sample become an internal control. The dot plot shows two numbers for each gene and cell type: dot size is the fraction of cells that express the gene, and color is the average expression. A middling average can mean every cell expresses the gene a little or a few cells express it strongly, and the size tells them apart.
Figure 5. A simulated single-cell experiment: two tumors, one TP53 wildtype and one R175H, eight cell types. Each cell type has a pair of dots, wildtype tumor first, R175H tumor second with an accent ring. Only the malignant epithelial cells carry the mutation; every other cell type is wildtype in both tumors. Intervals in the readout are 95% intervals for a difference of two proportions (Newcombe's hybrid score method). Hover, tap, or focus the plot and use the arrow keys.
In the malignant cells, CDKN1A is expressed in 50% of wildtype cells and 19% of mutant ones, and MDM2, another p53 target, falls with it (41% to 10%). TP53 mRNA itself does not move (68% and 68%): R175H changes the protein, not how much of the gene is transcribed. In the seven other cell types CDKN1A moves by at most 12 points, and each of those intervals includes zero. That is the internal control: the mutation acts in the cells that carry it. Bulk RNA from the whole tumor would blend the two, and the stroma would dilute the drop. Try 40 cells per type: the control intervals widen to about ±20 points, wide enough to hide a modest real change. At 400 they shrink to about ±7.
The oncoprint counts events without measuring their effect, and it missed the methylated tumors entirely. The lollipop shows where mutations fall across thousands of tumors, but not what any one of them does to p21. The promoter track shows where methylation sits, but a β value is a fraction of copies, so 0.4 could mean every copy partly methylated or a few copies fully. The pathway diagram is a map of past experiments, not a measurement of this tumor. The dot plot sees cells one at a time, but single-cell sequencing misses many of the transcripts in each cell, so "not expressing" partly means "not detected".
Where the views meet on the same patient, a claim gets stronger. Patient P062 has no alteration in any of Figure 1's genes, makes CDKN1A at 4.44 against a cohort median of 5.47, and in Figure 3 has an island β of 0.80. Three views, one answer: p53 is fine and the promoter is shut. The rest of the series takes each kind of data further, starting with reads and coordinates in the genome browser, at the same CDKN1A promoter and at TP53 codon 175.