GTEx Tissue Classification: PCA, K-Means & Logistic Regression

Zahin Peerzade

Unsupervised and supervised ML pipeline applied to bulk RNA-seq data from the GTEx project to classify tissue types from gene expression profiles.

---

What This Project Does

Using gene expression data from 4 tissue types — Liver, Brain (Cerebellum), Skin (Suprapubic), and Muscle (Skeletal) — this notebook builds a full ML pipeline from raw counts to tissue classification:

1. PCA — Reduces 15,000-gene expression matrix to a compact representation via manual eigen-decomposition (no sklearn PCA for the unsupervised steps)
2. K-Means Clustering — Custom implementation clusters samples in PCA space; elbow method selects k=4
3. Cluster Validation — Manual NumPy-only implementations of Silhouette Score and Adjusted Rand Index
4. Logistic Regression — sklearn LR with L2 regularization trained on PCA-reduced features; regularization strength tuned across C ∈ {0.01, 0.1, 1, 10}
5. Cross-Validation — 5-fold CV on training set across accuracy, precision, recall, and F1
6. Gene Annotation — Top contributing genes identified from PCA loadings and annotated via MyGene.info API

---

Dataset

GTEx v8 (Genotype-Tissue Expression project)

• Expression matrix: GTEx_Analysis_2017-06-05_v8_RNASeQCv1.1.9_gene_tpm.gct
• Sample metadata: GTEx_Analysis.txt
• Top 15,000 genes by expression used for computational tractability

Data is not included in this repo due to size. Download from gtexportal.org.

---

Requirements

pip install pandas numpy matplotlib scikit-learn requests

---

How to Run

1. Download GTEx expression matrix and metadata (links above)
2. Place both files in the same directory as the notebook
3. Open gtex_tissue_classification.ipynb and run all cells

---

Key Results

Method	Metric	Score
K-Means (k=4)	Adjusted Rand Index	~0.9+
K-Means (k=4)	Silhouette Score	~0.4+
Logistic Regression (C=0.1)	Test Accuracy	~1.00
Logistic Regression (C=0.1)	5-fold CV F1 (macro)	~1.00 ± 0.00

The four tissues are highly separable in PCA space — muscle and liver in particular show strong clustering driven by tissue-specific expression programs.

---

Repository Structure

gtex-tissue-classification/
├── gtex_tissue_classification.ipynb   # Full pipeline notebook
└── README.md

---

References

• GTEx Consortium. The GTEx Consortium atlas of genetic regulatory effects across human tissues. Science. 2020.
• Pedregosa et al. Scikit-learn: Machine Learning in Python. JMLR. 2011.
• MyGene.info: mygene.info