AI summary by TwinMind
StatQuest: PCA in Python
StatQuest with Josh Starmer
11m 37s
Introduction and Setup
- Josh Stormer presenting StatQuest tutorial on PCA in Python
- Acknowledgment to Chris Yoga Zimsky for code contribution
- Tutorial covers: data generation, PCA function usage, variation analysis, matplotlib graphing, loading scores examination
Package Imports
- pandas: data manipulation (panel data)
- numpy: random number generation, mathematical operations
- random: example dataset generation
- sklearn.decomposition.PCA: most commonly used Python PCA function
- sklearn.preprocessing: data scaling functions before PCA
- matplotlib: graph visualization
- Note: Python requires importing packages unlike R's built-in support
Sample Dataset Generation
- Created 100 gene names array (gene1, gene2, etc.)
- Sample arrays: 5 wild-type (WT) samples, 5 knockout (KO) samples
- range(1,6) generates values 1-5 (up to but not including 6)
- Pandas DataFrame creation with unpacked arrays (*WT, *KO) for column names
- Gene names used as index (row names equivalent)
- Random data from two Poisson distributions (one per sample type)
- Poisson means vary between 10-1000 per gene
- Final dimensions: 100 genes × 10 samples
- Note: Poisson distribution not ideal for RNA-seq data but sufficient for example
Data Preprocessing
- Centering: average value per gene = 0
- Scaling: standard deviation per gene = 1
- Data transposed before scaling (scale function expects samples in rows)
- Alternative method: StandardScaler.fit_transform (more common for machine learning)
- Scaling difference between scikit-learn vs R:
- Scikit-learn: (measurements - mean)² / n
- R (scale/prcomp): (measurements - mean)² / (n-1)
- Differences don't affect PCA analysis but minor effect on final graph coordinates
PCA Implementation
- Created PCA object (scikit-learn uses objects vs functions)
- Object can be trained on one dataset, applied to another
- fit() method performs PCA math: calculates loading scores, variation per component
- transform() generates coordinates for PCA graph based on loading scores and scaled data
Scree Plot Analysis
- Calculated percentage variation per principal component
- Created labels (PC1, PC2, etc.)
- Matplotlib bar plot creation
- Results: almost all variation along first principal component
- Conclusion: 2D graph using PC1 & PC2 adequate for representing original data
PCA Plot Visualization
- Coordinates from PCA.transform() placed in matrix
- Rows: sample labels, Columns: PC labels
- Scatter plot with title and axis labels
- Loop added sample names to graph
- Results:
- WT samples clustered left side (correlated with each other)
- KO samples clustered right side (correlated with each other)
- Separation along x-axis indicates WT vs KO differences
Loading Scores Analysis
- Examined PC1 loading scores to identify genes influencing cluster separation
- Created pandas Series object with loading scores
- Principal components zero-indexed (PC1 = 0)
- Sorted by magnitude (absolute value)
- Top 10 gene names and corresponding loading scores
- Results: values super similar, many genes contributed to sample separation (not just 1-2)
