The rainfall plot: its motivation, characteristics and pitfalls

Diana Domanska, Daniel Vodák, Christin Lund-Andersen, Stefania Salvatore, Eivind Hovig, Geir Kjetil SandveBMC Bioinformatics 2017 · DOI

Summary

In cancer genomics, mutations often cluster in localized regions known as hotspots, a phenomenon termed kataegis. To visualize these patterns, researchers frequently employ "rainfall plots" (RPs). Despite their widespread use since 2012, the mathematical foundations and inherent limitations of RPs have remained largely unexplored. This paper provides a formal deep-dive into the rainfall plot, identifying its unique strengths in multi-scale pattern discovery and uncovering significant "congestion" pitfalls that can lead to biased interpretations.

The Problem: Visualizing Mutation Hotspots

Cancer mutations are rarely distributed evenly across the genome. Identifying regions of high mutational density, or hotspots, is critical for understanding disease progression. The rainfall plot was developed as an intuitive tool for this purpose: it is a scatter plot where the x-axis represents genomic position and the y-axis represents the distance to the previous mutation (typically on a log scale).

Despite its popularity, the RP lacked a formal mathematical definition. Researchers used it based on intuition, potentially overlooking systematic biases. This study addresses this gap by formally defining the RP and critically examining its performance using both real-world cancer genomics data and synthetic datasets.

Methodology

The authors approached the evaluation of rainfall plots from three perspectives:

  1. Formal Definition: Establishing a rigorous mathematical transformation of inter-event distances.
  2. Conceptual Modeling: Using stochastic processes, such as Poisson processes and Hidden Markov Models, to explain common visual patterns like horizontal bands.
  3. Simulation: Creating synthetic datasets to test the limits of the plot, particularly regarding "congestion"—where multiple data points overlap on a single pixel.
Comparison of a genomic track (a), a traditional frequency line plot (b), and a rainfall plot (c) showing genomic distance.

Strengths: Parameter-Free Discovery

A key advantage of the rainfall plot over traditional frequency-based line plots is its parameter-free nature. Frequency plots require a pre-defined "bin size" (e.g., 1MB), which can obscure features that exist at different scales. RPs, however, can simultaneously display broad genomic trends and hyper-local clusters without any manual tuning. This makes them exceptionally powerful for exploratory data analysis where the scale of interest is unknown.

Weaknesses: The Congestion Trap

The most significant finding of this study is the "congestion" pitfall. Because standard figures have a finite number of pixels, a single dot on a genome-wide RP can hide a staggering number of mutations. In dense regions, a single pixel might represent over 200,000 overlapping events.

A rainfall plot illustrating the "congestion" issue, where dense clusters of mutations at short genomic distances can saturate the visual space.

This saturation leads to two major issues:

  • Loss of Quantitative Information: It is impossible to visually count mutations in congested areas.
  • Plotting Order Bias: When using colors to represent different mutation types, the software's drawing order can cause one type to "cover up" others, creating a false visual impression of prevalence.

Practical Guidelines for Researchers

To mitigate these issues, the authors propose several guidelines:

  • Interpret "Rain" Correctly: Dense streaks at the bottom indicate high-intensity hotspots; thin clouds at the top represent low-intensity regions.
  • Use the Y-Axis for Precision: Look at the actual y-values to understand the typical distance between mutations, rather than just the density of dots.
  • Beware of Saturation: Recognize that solid-looking regions are saturated and cannot be used for counting.
  • Neutralize Color Bias: Use neutral colors for overlaps or re-plot with different drawing orders to ensure visual consistency.
  • Identify Horizontal Bands: Persistent horizontal lines suggest recurrent clustering at specific scales, which may indicate biological or technical artifacts.

Conclusion

While rainfall plots are excellent for qualitative pattern discovery across multiple scales, they are poor tools for quantitative readout. The authors suggest that future work should focus on interactive visualization tools, such as dynamic zooming or heat-map style coloring, to allow researchers to transition from global overviews to precise, quantitative counts without losing context.

Summary updated

BibTeX
@article{domanska2017rainfall,
  author = {Diana Domanska and Daniel Vodák and Christin Lund-Andersen and Stefania Salvatore and Eivind Hovig and Geir Kjetil Sandve},
  title = {{The rainfall plot: its motivation, characteristics and pitfalls}},
  journal = {BMC Bioinformatics},
  year = {2017},
  doi = {10.1186/s12859-017-1679-8},
}
Figure