Principal Component Analysis for Omics Data

Summarizing correlated molecular measurements with orthogonal axes of variation

Bioinformatics

Algorithms

Omics

Dimensionality Reduction

How principal component analysis projects high-dimensional omics measurements into variance-maximizing components for exploration, quality control, and covariate discovery.

Author

Jiahong Sun

Published

August 27, 2026

Modified

August 27, 2026

1 Topic at a Glance

One-sentence summary: Principal component analysis (PCA) replaces many correlated molecular features with a small number of mutually orthogonal directions that explain as much variation in the measured samples as possible.

Item	Description
Biological question	Which broad sources of variation separate samples or features, and do technical factors dominate the data?
Computational problem	Represent an $n\times p$ measurement matrix with a few informative axes without using class labels
Method type	Unsupervised linear dimensionality reduction
Input	A normalized, transformed sample-by-feature matrix plus sample and feature metadata
Output	Sample scores, feature loadings, explained-variance fractions, and diagnostic plots
Bioinformatics uses	RNA-seq and single-cell QC, population structure, proteomics, methylation, and batch-effect diagnosis
Prerequisites	Matrix multiplication, variance, centering, and a stated normalization procedure

2 The variation PCA captures

An omics experiment can measure thousands of genes, proteins, peaks, or variants for each sample. Plotting all dimensions is impossible, and many features co-vary because they share cell type, pathway activity, ancestry, sequencing depth, or batch. PCA finds linear combinations of features that describe the largest remaining variation in the matrix.

For a centered data matrix $X\in\mathbb{R}^{n\times p}$, where rows are samples and columns are features, the first loading vector $v_1$ is chosen to maximize the variance of the projected samples:

\[ v_1=\arg\max_{\lVert v\rVert=1}\lVert Xv\rVert^2. \]

Later vectors maximize the remaining variation while being orthogonal to earlier vectors. With the singular-value decomposition

\[ X=U\Sigma V^\mathsf{T}, \]

the columns of $V$ are feature loadings, $U\Sigma$ gives sample scores, and the squared singular values describe variance along each component. A PC is an axis of measured variation, not automatically a biological factor or a causal explanation.

feature matrix
    -> normalize, transform, and center features
    -> singular-value decomposition
    -> sample scores + feature loadings + variance explained
    -> QC, exploration, and covariate investigation

3 A small example

Suppose four bulk-RNA-seq samples have expression measurements for two genes after a suitable variance-stabilizing transform:

Sample	gene A	gene B
control 1	2	2
control 2	3	3
treated 1	8	8
treated 2	9	9

Both genes change together. After centering each gene, the dominant direction is approximately $(1,1)/\sqrt{2}$: PC1 separates controls from treated samples. The perpendicular direction $(1,-1)/\sqrt{2}$ has essentially no variation here. In real data, a PC can instead track library size, mitochondrial RNA fraction, donor ancestry, cell cycle, or a mixture of these effects. Metadata are necessary to interpret it.

4 Preparing omics data matters more than the eigensolver

PCA is deterministic for a fixed matrix, but the matrix is a modeling choice.

Choice	Why it matters	Typical practice
Feature filtering	Thousands of uninformative or sparse features can obscure structure	Remove features with too little expression/variation; record the rule
Normalization	Raw counts mainly reflect library size and composition	Use an assay-appropriate normalization before PCA
Transformation	Count variance increases with the mean	Use a log-like or variance-stabilizing transform where appropriate
Centering	PCA otherwise points toward the mean abundance profile	Center every feature unless a justified alternative is used
Scaling	High-variance features otherwise dominate	Scale to unit variance only when features should contribute equally

For RNA-seq counts, PCA on raw counts is usually misleading. A common workflow filters low-expression genes, accounts for sequencing depth, then uses a variance-stabilizing or regularized-log transformation. For single-cell data, sparsity, depth variation, and feature-selection choices are particularly consequential; PCA is often applied to selected highly variable genes rather than every gene.

5 Reading a PCA result

5.1 Scores: where samples lie

A PC1-versus-PC2 scatterplot places each sample by its scores. Color and shape points using pre-specified metadata such as condition, donor, library-preparation batch, or run. Separation can identify a biologically useful pattern, but it can also expose confounding: if every treated sample was sequenced in one batch, condition and batch cannot be disentangled from that plot alone.

5.2 Loadings: which features define an axis

Large positive or negative loadings identify features associated with a PC direction. Their signs are arbitrary: multiplying both a component’s scores and loadings by $-1$ produces the identical solution. Enrichment analysis of ranked loadings can help form a hypothesis, but it is exploratory and should not replace a preregistered differential-expression test.

5.3 Explained variance: how much the projection loses

The proportion for PC $k$ is $\sigma_k^2/\sum_j\sigma_j^2$. It describes reconstruction variance in the supplied matrix, not biological importance. A low-variance component may distinguish a clinically important group; a high-variance component may simply represent RNA quality.

6 Diagnostics and failure modes

Observation	Possible explanation	Response
One sample is far from all others	Mislabeling, contamination, low depth, or a real outlier	Inspect QC metrics and provenance before excluding it
PC1 follows library size or mitochondrial fraction	Incomplete normalization or damaged cells	Revisit QC and transformation; do not silently regress it away
Groups separate exactly by batch	Technical confounding	Check study design; use a justified batch-aware analysis, not PCA alone
Interpretation changes after adding samples	PCA axes are data-dependent	Refit and report the cohort and preprocessing used
A crisp two-dimensional plot misses known biology	Signal lies in later or nonlinear structure	Inspect more PCs and consider methods such as UMAP only after QC

PCA assumes linear structure and emphasizes directions of large variance. It does not supply clusters, adjusted treatment effects, or a batch correction. Removing PCs because they correlate with a variable can discard genuine biology when that variable is confounded with the signal of interest.

7 Connections

PCA is a geometric view of the same matrix that later methods may model differently. Non-negative matrix factorization learns additive, nonnegative components rather than orthogonal signed axes, which can be more interpretable for nonnegative abundance data. Upcoming neighborhood and clustering methods use low-dimensional representations to build sample graphs or groups, but should inherit PCA’s preprocessing and QC scrutiny. In population genetics, PCA of genotype-derived matrices is often used to visualize ancestry structure; the axes still reflect sampling and demographic history as well as biology.

8 Take-home Messages

PCA summarizes a centered matrix using orthogonal directions ordered by captured variance.
Scores locate samples; loadings identify the features associated with each direction; neither is self-interpreting.
Filtering, normalization, transformation, centering, and scaling define the analysis as much as the PCA calculation does.
PCA is an exploratory QC and covariate-discovery tool, not a substitute for a designed statistical comparison or causal inference.

9 References

Pearson, K. 1901. On lines and planes of closest fit to systems of points in space. Philosophical Magazine 2:559–572.
Hotelling, H. 1933. Analysis of a complex of statistical variables into principal components. Journal of Educational Psychology 24:417–441.
Patterson, N., Price, A. L., & Reich, D. 2006. Population structure and eigenanalysis. PLoS Genetics 2:e190.
Love, M. I., Huber, W., & Anders, S. 2014. Moderated estimation of fold change and dispersion for RNA-seq data with DESeq2. Genome Biology 15:550.

--- title: "Principal Component Analysis for Omics Data" subtitle: "Summarizing correlated molecular measurements with orthogonal axes of variation" description: "How principal component analysis projects high-dimensional omics measurements into variance-maximizing components for exploration, quality control, and covariate discovery." author: "Jiahong Sun" date: "2026-08-27" date-modified: "2026-08-27" categories: [Bioinformatics, Algorithms, Omics, Dimensionality Reduction] toc: true number-sections: true code-fold: show draft: false --- # Topic at a Glance > **One-sentence summary:** Principal component analysis (PCA) replaces many correlated molecular features with a small number of mutually orthogonal directions that explain as much variation in the measured samples as possible. | Item | Description | |---|---| | **Biological question** | Which broad sources of variation separate samples or features, and do technical factors dominate the data? | | **Computational problem** | Represent an $n\times p$ measurement matrix with a few informative axes without using class labels | | **Method type** | Unsupervised linear dimensionality reduction | | **Input** | A normalized, transformed sample-by-feature matrix plus sample and feature metadata | | **Output** | Sample scores, feature loadings, explained-variance fractions, and diagnostic plots | | **Bioinformatics uses** | RNA-seq and single-cell QC, population structure, proteomics, methylation, and batch-effect diagnosis | | **Prerequisites** | Matrix multiplication, variance, centering, and a stated normalization procedure | # The variation PCA captures An omics experiment can measure thousands of genes, proteins, peaks, or variants for each sample. Plotting all dimensions is impossible, and many features co-vary because they share cell type, pathway activity, ancestry, sequencing depth, or batch. PCA finds linear combinations of features that describe the largest remaining variation in the matrix. For a centered data matrix $X\in\mathbb{R}^{n\times p}$, where rows are samples and columns are features, the first loading vector $v_1$ is chosen to maximize the variance of the projected samples: $$ v_1=\arg\max_{\lVert v\rVert=1}\lVert Xv\rVert^2. $$ Later vectors maximize the remaining variation while being orthogonal to earlier vectors. With the singular-value decomposition $$ X=U\Sigma V^\mathsf{T}, $$ the columns of $V$ are **feature loadings**, $U\Sigma$ gives **sample scores**, and the squared singular values describe variance along each component. A PC is an axis of measured variation, not automatically a biological factor or a causal explanation. ```text feature matrix -> normalize, transform, and center features -> singular-value decomposition -> sample scores + feature loadings + variance explained -> QC, exploration, and covariate investigation ``` # A small example Suppose four bulk-RNA-seq samples have expression measurements for two genes after a suitable variance-stabilizing transform: | Sample | gene A | gene B | |---|---:|---:| | control 1 | 2 | 2 | | control 2 | 3 | 3 | | treated 1 | 8 | 8 | | treated 2 | 9 | 9 | Both genes change together. After centering each gene, the dominant direction is approximately $(1,1)/\sqrt{2}$: PC1 separates controls from treated samples. The perpendicular direction $(1,-1)/\sqrt{2}$ has essentially no variation here. In real data, a PC can instead track library size, mitochondrial RNA fraction, donor ancestry, cell cycle, or a mixture of these effects. Metadata are necessary to interpret it. # Preparing omics data matters more than the eigensolver PCA is deterministic for a fixed matrix, but the matrix is a modeling choice. | Choice | Why it matters | Typical practice | |---|---|---| | Feature filtering | Thousands of uninformative or sparse features can obscure structure | Remove features with too little expression/variation; record the rule | | Normalization | Raw counts mainly reflect library size and composition | Use an assay-appropriate normalization before PCA | | Transformation | Count variance increases with the mean | Use a log-like or variance-stabilizing transform where appropriate | | Centering | PCA otherwise points toward the mean abundance profile | Center every feature unless a justified alternative is used | | Scaling | High-variance features otherwise dominate | Scale to unit variance only when features should contribute equally | For RNA-seq counts, PCA on raw counts is usually misleading. A common workflow filters low-expression genes, accounts for sequencing depth, then uses a variance-stabilizing or regularized-log transformation. For single-cell data, sparsity, depth variation, and feature-selection choices are particularly consequential; PCA is often applied to selected highly variable genes rather than every gene. # Reading a PCA result ## Scores: where samples lie A PC1-versus-PC2 scatterplot places each sample by its scores. Color and shape points using *pre-specified* metadata such as condition, donor, library-preparation batch, or run. Separation can identify a biologically useful pattern, but it can also expose confounding: if every treated sample was sequenced in one batch, condition and batch cannot be disentangled from that plot alone. ## Loadings: which features define an axis Large positive or negative loadings identify features associated with a PC direction. Their signs are arbitrary: multiplying both a component's scores and loadings by $-1$ produces the identical solution. Enrichment analysis of ranked loadings can help form a hypothesis, but it is exploratory and should not replace a preregistered differential-expression test. ## Explained variance: how much the projection loses The proportion for PC $k$ is $\sigma_k^2/\sum_j\sigma_j^2$. It describes reconstruction variance in the supplied matrix, not biological importance. A low-variance component may distinguish a clinically important group; a high-variance component may simply represent RNA quality. # Diagnostics and failure modes | Observation | Possible explanation | Response | |---|---|---| | One sample is far from all others | Mislabeling, contamination, low depth, or a real outlier | Inspect QC metrics and provenance before excluding it | | PC1 follows library size or mitochondrial fraction | Incomplete normalization or damaged cells | Revisit QC and transformation; do not silently regress it away | | Groups separate exactly by batch | Technical confounding | Check study design; use a justified batch-aware analysis, not PCA alone | | Interpretation changes after adding samples | PCA axes are data-dependent | Refit and report the cohort and preprocessing used | | A crisp two-dimensional plot misses known biology | Signal lies in later or nonlinear structure | Inspect more PCs and consider methods such as UMAP only after QC | PCA assumes linear structure and emphasizes directions of large variance. It does not supply clusters, adjusted treatment effects, or a batch correction. Removing PCs because they correlate with a variable can discard genuine biology when that variable is confounded with the signal of interest. # Connections PCA is a geometric view of the same matrix that later methods may model differently. [Non-negative matrix factorization](non-negative-matrix-factorization.qmd) learns additive, nonnegative components rather than orthogonal signed axes, which can be more interpretable for nonnegative abundance data. Upcoming neighborhood and clustering methods use low-dimensional representations to build sample graphs or groups, but should inherit PCA's preprocessing and QC scrutiny. In population genetics, PCA of genotype-derived matrices is often used to visualize ancestry structure; the axes still reflect sampling and demographic history as well as biology. # Take-home Messages 1. PCA summarizes a centered matrix using orthogonal directions ordered by captured variance. 2. Scores locate samples; loadings identify the features associated with each direction; neither is self-interpreting. 3. Filtering, normalization, transformation, centering, and scaling define the analysis as much as the PCA calculation does. 4. PCA is an exploratory QC and covariate-discovery tool, not a substitute for a designed statistical comparison or causal inference. # References - Pearson, K. 1901. [On lines and planes of closest fit to systems of points in space](https://doi.org/10.1080/14786440109462720). *Philosophical Magazine* 2:559--572. - Hotelling, H. 1933. [Analysis of a complex of statistical variables into principal components](https://doi.org/10.1037/h0071325). *Journal of Educational Psychology* 24:417--441. - Patterson, N., Price, A. L., & Reich, D. 2006. [Population structure and eigenanalysis](https://doi.org/10.1371/journal.pgen.0020190). *PLoS Genetics* 2:e190. - Love, M. I., Huber, W., & Anders, S. 2014. [Moderated estimation of fold change and dispersion for RNA-seq data with DESeq2](https://doi.org/10.1186/s13059-014-0550-8). *Genome Biology* 15:550.