Lab 4 — Visualizing single-cell genomic data

Goal. Apply PCA, t-SNE, and UMAP to a public single-cell RNA-seq dataset. Compare what each method preserves. Discuss the cost of nonlinear methods for downstream interpretation.

What you ship. Notebook with three 2-D embeddings of the same data, side-by-side, with a 200-word memo on when each is the right tool.

Setup

Install the dependencies (one-time).

# !pip install scanpy umap-learn scikit-learn matplotlib

import numpy as np
import matplotlib.pyplot as plt
import scanpy as sc
from sklearn.decomposition import PCA
from sklearn.manifold import TSNE
import umap

np.random.seed(42)

10x Genomics PBMC 3K (a canonical single-cell dataset)

adata = sc.datasets.pbmc3k()
sc.pp.normalize_total(adata, target_sum=1e4)
sc.pp.log1p(adata)
sc.pp.highly_variable_genes(adata, n_top_genes=2000)
adata = adata[:, adata.var.highly_variable]
X = adata.X.toarray()
print('X shape:', X.shape)

Exercise 1 — PCA

# YOUR TURN
# Compute first 50 PCs. Plot PC1 vs PC2.

Exercise 2 — t-SNE on PCA features

# YOUR TURN
# Run t-SNE (perplexity=30) on the first 50 PCs.

Exercise 3 — UMAP on PCA features

# YOUR TURN
# Run UMAP (n_neighbors=15, min_dist=0.1) on the first 50 PCs.

Exercise 4 — Compare visually

# YOUR TURN
# Plot all three side by side. Color by Louvain cluster (sc.tl.louvain).
# Write 200 words on which is best for which downstream task.

Done?

Submit per the cohort schedule. Peer review pairing announced the following Monday.