Data Visualization Principles for Biologists

Navigating Complexity in High-Dimensional Multi-Omics

Data Visualization
Principles for Biologists

The core challenge in life sciences is no longer generating data, but computationally analyzing and cognitively interpreting it. Visualization is now an integrated subdiscipline of bioinformatics operating at the nexus of computational engineering, statistical modeling, and cognitive psychology.

Next-Gen Sequencing (NGS) Spatial Transcriptomics 3D Proteomics Metagenomics

The Stakes: Ineffective visualizations lead to hallucination of false biological phenomena, misallocation of funding, and direct patient harm. High-quality visuals inform therapeutic selection and predict drug responses.

Cognitive Foundations & Gestalt Principles

The human visual cortex uses innate heuristics to organize complex stimuli. Good bioinformatics design leverages these rules to minimize cognitive load.

Proximity

Spatially adjacent points are perceived as a distinct, related population (crucial for PCA/UMAP scatter plots).

Similarity

Objects sharing visual attributes (color, shape) form unified groups. Used to map categorical metadata (e.g. sequencing batches).

Connection

Lines instantly imply structural/biological relationships, forming the basis of pseudotime trajectories and protein-protein networks.

Enclosure & Interpolation

Visual cortex completes boundaries, vital for discerning continuous tissue boundaries from pixelated barcodes in spatial transcriptomics.

The Mathematics of Colorization

Color must be a quantitative encoding mechanism corresponding to mathematical reality: categorical, sequential (0 to max), or diverging (- to +).

ANTI-PATTERN: The Rainbow / Jet Palette

Perceptually non-uniform. Introduces artificial boundaries and hallucinated sharp spatial gradients in continuous data.

GOLD STANDARD: Perceptually Uniform Scales

Use Viridis, Magma, or ColorBrewer. Scales luminance linearly. Avoid red-green overlapping schemes for colorblind accessibility.

Simultaneous Contrast Illusion: The perceived hue or intensity of a data point can be distorted by its background colors. Context matters.

M.E.D.V.I.S. Evaluation Algorithm

Metrics for Evaluation and Discretization of Biomedical Visuals computationally evaluates figures using PCA, t-SNE/UMAP, and K-Means consensus clustering.

Complexity

Assessed via advanced image segmentation. Prevents excessive cognitive load from overlapping artifacts.

Color Usage

Calculates variance across RGB channels. Enforces purposeful color application over chaotic chromaticity.

Whitespace

Analyzes absolute white pixel count/ratio. Separates sub-panels and prevents severe visual clutter.

Viz Count

Quantifies distinct elements. Balances info density, indicating when multi-panels should be decoupled.

Critical Warning: The M.E.D.V.I.S framework explicitly cautions against 3D visual elements (3D pie charts, 3D scatter plots). Introducing a z-axis on a 2D medium almost universally distorts geometric relationships.

Transcriptomic Profiling: Differential Expression

Pipelines like DESeq2, edgeR, and limma-voom process counts into expression matrices. These visual architectures transform statistical tables into insight.

Volcano Plots

Balances magnitude of change (effect size) and statistical confidence (p-value). Visualizes statistical summary of repeated univariate tests.

X-Axis: log₂(Fold Change) centers data. 2-fold up = +1, 2-fold down = -1.
Y-Axis: -log₁₀(P-value) or FDR/q-value. Highly significant points project toward the apex.
Thresholding Fallacy: Rigid log₂(FC) ≥ 2.0 cutoffs aggressively discard genuine biological signals. Prefer ±0.58 (1.5-fold) with FDR < 0.05. Don't use raw FPKM over normalized counts.
π-value Score: Mathematically combines FC and P-value into a single ranking metric, yielding larger Area Under Curve (AUC) for GSEA.

Heatmaps & Hierarchical Clustering

Analyzes expression variance and co-varying dynamics via unsupervised machine learning. (Requires R packages like heatmap3 or ComplexHeatmap).

Distance Metrics: Euclidean (straight-line, sensitive to absolute magnitude) vs. Pearson Correlation (evaluates linear correlation, groups by trend).
Agglomeration (Linkage): Defines distances between newly formed clusters. Methods: Single (closest), Complete (furthest), Average (mean), Ward's (variance-minimizing). Utilizes optimized fastcluster algorithm.
Data Standardization: Feature scaling via Z-Scores (x-μ)/σ is mandatory to prevent highly expressed housekeeping genes from permanently masking biological variance.
Annotation: Margins map categorical/continuous phenotypes (age, drug dose).

MA Plots

Scatters log₂(FC) against average log₂(Expression). Highly useful for evaluating magnitude of change against mean expression, identifying expression-dependent biases, and assessing normalization quality.

Ridgeplots

Displays density distributions of log₂(FC) values for genes within targeted biological pathways. Instantly reveals if an entire pathway is systematically shifting in expression.

Dimensionality Reduction (DR) Triad

scRNA-seq (20k-30k genes across millions of cells) creates ultra-high-dimensional spaces. Frameworks like Seurat and Scanpy rely on DR to distill this down to 2D Cartesian planes.

PCA

Principal Component Analysis

Foundation: Linear algebraic transformation (Covariance matrix, Eigenvectors).
Strengths: Extremely fast, perfectly preserves global distances. Excellent for QC, outliers, and batch effects. Deterministic.
Workflow role: Rarely the final output. The top 20-50 PCs are extracted to feed non-linear algorithms.

t-SNE

Stochastic Neighbor Embedding

Foundation: Probabilistic manifold. Uses Student's t-distribution to fix "crowding", minimizes KL divergence.
Strengths: Unmatched local resolution ("island" effect), heavily correlating with known transcriptomic phenotypes.
Weaknesses: Global distances are meaningless. Aggressively inflates dense clusters. Highly sensitive to perplexity.

UMAP

Uniform Manifold Approx & Projection

Foundation: Algebraic topology (high-dimensional fuzzy simplicial complex optimized via cross-entropy).
Strengths: Superior processing speed. Balances local cluster fidelity with better global structure preservation. Discovered rare Axl-Sigma dendritic cells via PERFF-seq.
Weaknesses: Sensitive to n_neighbors and min_dist.

The "Specious Art" Warning (Chari & Pachter)

Reference mapping procedures are highly vulnerable. Transforming uniformly distributed, random noise using UMAP reference coordinates coerces that noise into adopting the distinct structural shape of the biological single-cell data. Never interpret empty geometric space between clusters as actual biological transcriptomic distance!

Algorithmic Solutions & Next-Gen Embeddings

LMC & MICS: (Landmark Mantel Correlation & Multi-resolution Cluster Supervision) jointly optimize local & global structures, reducing specious artifacts.
scDEED: Single-cell dubious embedding detector statistically verifies if pre-embedding neighbors remain neighbors.
Deep Learning (DCA, VAEs, GANs): Using Autoencoders to learn parametric latent spaces before applying UMAP anchors the projection to a more mathematically robust manifold.

Integrative Systems: Multi-Omics & Spatial

Circos Plots (Multi-Omics Frameworks)

Abandons the Cartesian plane to align linear genomics circumferentially. Mitigates genomic sparseness by locally zooming on loci. Tools: Generated programmatically via R (OmicCircos) or JavaScript for interactive web browsers (NG-Circos, BioCircos.js).

1 Segment Data (Anchor Track): Defines outermost boundary using reference genomes (e.g., hg19, mm10).
2 Mapping Data (Concentric Rings): Projects diverse plots (scatter, heatmaps) tied strictly to outer genomic coords. (e.g. CNV on Track 1, Methylation on Track 2).
3 Polygons (Core Linkages): Innermost area visualizing inter-chromosomal structural fusions, and protein-protein interactions directly through empty space.

Case Study: Tumor-Immune-Gut Axis (Oncology)

A Phase 2 clinical trial used Circos network maps to integrate bulk RNA-seq, 16S microbiome sequencing, and fecal metabolomics. The circular topology successfully linked transcriptomic shifts in the Tumor Microenvironment (e.g., Tertiary Lymphoid Structures marked by CCL19/CXCL13) to a systemic influx of T/B cells, specific fecal bacteria species, and lipid metabolites.

Next Gen: OmicsFootPrint

Deep learning framework explicitly transforms numerical multi-omics signatures into 2D circular images. Fed into DNNs coupled with SHAP algorithms, achieving 0.98 AUC for lung cancer subtype differentiation.

In Situ Histological Mapping

Spatial Transcriptomics: Mapping the Microenvironment

Standard scRNA-seq destroys physical tissue context. Spatial technologies map high-dimensional expression directly back onto rigid 2D X-Y physical coordinates, revealing localized paracrine signaling.

In Situ Hybridization (ISH) / Multiplexed FISH: Fluorescent RNA probes bind to complementary sequences. High sub-cellular resolution, lower multiplexing.
In Situ Sequencing (ISS): RNA is reverse-transcribed and sequenced directly within tissue matrix.
Spatial Barcoding (e.g. Visium): Oligo barcodes with predefined spatial addresses ligated to RNA.

Sub-cellular Tools: The expressyouRcell R package maps complex transcriptomic changes directly onto dynamic cellular pictographs with color-coded sub-compartments.

Dual-Preservation: Algorithms like SpaSNE preserve both transcriptional similarity AND physical tissue geometry, vital for calculating physical receptor-ligand communication.

Immersive VR: StarmapVis and cellexalvr let researchers physically walk through the spatial heterogeneity in fully interactive 3D spaces.

Systematic Plot Selection & Frameworks

Four primary interrogations dictate visualization choice. Bioinformaticians execute these via the "Grammar of Graphics" (using ggplot2 in R or seaborn in Python) to guarantee absolute programmatic reproducibility.

Distribution

Evaluate spread & skewness (e.g. sequencing read depth).

Histograms, Density, Violins

Rule: Pure boxplots hide bimodal multimodal distributions. Always overlay raw individual points (using jitter).

Correlation

Relationship between continuous variables (e.g. marker genes).

Scatter Plots, 2D Hex-bins

Rule: Genomics N is huge. Use alpha-blending (transparency) or hex-binning to convey true density and combat overplotting.

Composition

Part-to-whole relationships (e.g. CD4+ vs CD8+ T-cell ratios).

Treemaps, Stacked Bars

Rule: Avoid Pie Charts completely. The human visual cortex is demonstrably poor at comparing relative angles and circular areas.

Comparison & Rank

Categorical differences across cohorts.

Dot Plots, Violin Plots

Rule: Never use "Dynamite plots" (bar charts with error bar). They actively conceal the true variance and distribution.