Driver genes and hotspots
A gene that is mutated often is not necessarily a driver: it may simply be long, or sit in a region with a high mutation rate. Driver mutations are selected, and they tend to recur at the same position in the protein. A lollipop plot draws this clustering; positional clustering methods turn it into a statistic you can test.
Reading a lollipop plot
The x-axis is the protein sequence and the coloured boxes are annotated domains. Each lollipop is one amino-acid position, its height is the number of mutations seen there, and its colour is the variant classification. Switch genes in the widget and two very different patterns appear:
- Oncogene-style hotspots. IDH1 mutations concentrate at R132, IDH2 at R140 and R172, and DNMT3A has one tall lollipop at R882 in its methyltransferase domain. Such recurrent hotspots are usually gain-of-function or neomorphic changes; mutant IDH1/2, for example, make the enzyme produce 2-hydroxyglutarate.
- Tumour-suppressor-style scatter. TP53 missense mutations spread across the DNA-binding domain, mixed with truncating ones; besides R882, DNMT3A also has truncating mutations along its length, pointing to loss of function.
NPM1 is an instructive exception. Almost all its mutations are the same 4 bp insertion in exon 12, classified as a frameshift, yet not a loss of function: it rewrites the C-terminus and moves NPM1 from the nucleus to the cytoplasm. Reading “frameshift = truncation = inactivation” mechanically gets it wrong.
Positional clustering: one way to find drivers
Turning hotspots into a statistic is the idea of positional clustering. maftools’ oncodrive follows OncodriveCLUST: it finds positions where mutations cluster, computes what fraction of each gene’s mutations falls in clusters, compares this with a background distribution, and reports p-values with an FDR correction. In LAML, IDH1, IDH2 and NPM1 are the genes with the strongest positional clustering.
The method has clear blind spots:
- Truncating mutations in tumour suppressors spread along the gene and do not cluster, so they are missed. Those genes need frequency-based methods, such as MutSigCV, which accounts for gene length, background mutation rate and mutation type, or dNdScv, which compares non-synonymous with synonymous rates.
- Gene length alone is not evidence: a very long gene such as TTN collects many mutations without being selected.
- In a small cohort each gene has few mutations, which limits the power of any method.
Common pitfalls
- Ranking by frequency instead of judging drivers. Frequency depends on gene length and background mutation rate.
- Using positional clustering alone. It systematically misses tumour suppressors.
- Reading “frameshift = inactivation” mechanically. NPM1 is the counter-example.
- Ignoring assay blind spots. Long insertions such as FLT3-ITD are easily missed in short-read WES, so the frequency computed is a lower bound.