The cohort landscape: oncoplots

Start with a cohort summary

maftools’ plotmafSummary draws six panels: variant classification, variant type (SNP, INS, DEL), the six substitution classes, variants per sample, a burden boxplot and the most frequently mutated genes. It is the sanity check before any real analysis.

TCGA-LAML is a typical low-burden tumour, with a median of about 9 non-synonymous mutations per sample. If your own data show thousands of variants in every sample, germline variants or false positives have most likely slipped through. If only a few samples sit far above the rest, they may be genuinely hypermutated (POLE mutation or mismatch-repair deficiency, for example); look at their spectrum (chapter 9) before calling them false positives.

Reading an oncoplot

In an oncoplot each row is a gene and each column a patient. A coloured tile means that patient has a non-synonymous mutation in that gene, with the colour giving the variant classification; a gene with more than one mutation in a patient is marked Multi_Hit. The bar on top is each patient’s variant count, and the number on the right is the share of patients mutated in that gene.

Genes are sorted by frequency. Patients are sorted for legibility: first by whether the top gene is mutated, then by the second gene, and so on, so the mutated tiles stack into a staircase. Two genes that rarely share a column look mutually exclusive; two that keep sharing one look co-occurring.

Hands-on with LAML: frequency, co-occurrence and exclusivity

Use the slider to change how many genes are shown. In LAML, FLT3, DNMT3A and NPM1 lead, with FLT3 and DNMT3A each mutated in about a quarter of patients, followed by IDH1/2, TET2, RUNX1, TP53, NRAS, CEBPA and others, in line with what is known about AML.

The percentages on the right use the 193 patients with a MAF record as the denominator. The clinical table holds 7 more patients who were sequenced but have no mutation record; counting them would lower every frequency slightly. When you report a frequency, say what it was divided by.

The staircase you see is only a hint. Click a gene name and the widget runs a Fisher’s exact test between it and each other gene, applies a BH correction across the genes shown, and reports q-values for exclusivity and co-occurrence. In AML, NPM1 often co-occurs with FLT3 and DNMT3A; the mutual exclusivity of IDH1/IDH2 and TET2 is a classic observation in the literature. See whether it reaches significance in these 193 patients.

Common pitfalls

• Comparing oncoplot frequencies across studies directly. Denominators (with or without samples that have no mutation), definitions of non-synonymous and coverage can all differ.
• Taking the staircase as evidence of exclusivity. It is a visual effect of the sorting; conclusions need a test and an FDR.
• Ignoring false negatives from poor coverage. Poorly covered genes look rarer than they are; CEBPA, with its very high GC content, is often thinly covered in capture sequencing.