Tumor Severity Stratification from Gene Expression Using Supervised Learning

In this project, we built a pipeline to classify tumor severity in lung adenocarcinoma patients using RNA-seq gene expression data from TCGA-LUAD (via cBioPortal). Patients with pathological stage > 1 were labeled as “severe,” and those with stage = 1 as “non-severe.”

510 patients — 55% severe, 45% non-severe
ML models: Logistic Regression, SVM, Random Forest, CNN
Best accuracy: 79% (CNN with 25% DEGs)

---

Step-by-Step Workflow

1️Data Acquisition

We downloaded gene expression (TPM normalized) and clinical data from cBioPortal, focusing on the TCGA-LUAD cohort.

• Merged clinical labels with gene expression matrix
• Ensured 1 row per patient and gene symbols as columns

---

Label Engineering

We created binary labels based on pathological stage:

df_clinical["label"] = df_clinical["PATHOLOGICAL_STAGE"].apply(lambda x: 1 if "Stage I" in x else 0)

• 1 = Non-severe (Stage I)
• 0 = Severe (Stage II or higher)

---

Data Preprocessing

We aligned the labels with the expression matrix and handled missing values. Then we transposed the expression matrix:

df_expr = df_expr.T
df_expr.columns = df_expr.iloc[0]
df_expr = df_expr.drop(index=df_expr.index[0])

---

Feature Selection using ANOVA F-test

We applied the SelectKBest method to identify top features:

from sklearn.feature_selection import SelectKBest, f_classif

selector = SelectKBest(f_classif, k='all')
selector.fit(X, y)
f_scores = selector.scores_

We then selected the top 5%, 15%, and 25% of features based on F-scores.

---

Model Training and Evaluation

We trained and evaluated four models using cross-validation:

Logistic Regression

from sklearn.linear_model import LogisticRegression
model = LogisticRegression()
model.fit(X_train, y_train)

Support Vector Machine

from sklearn.svm import SVC
model = SVC(kernel='rbf', C=10)
model.fit(X_train, y_train)

Random Forest

from sklearn.ensemble import RandomForestClassifier
model = RandomForestClassifier(n_estimators=100, max_depth=15)
model.fit(X_train, y_train)

Convolutional Neural Network (Keras)

model = Sequential([
    Input(shape=(X.shape[1], 1)),
    Conv1D(64, kernel_size=2, activation='relu'),
    MaxPooling1D(pool_size=2),
    Flatten(),
    Dense(128, activation='relu'),
    Dropout(0.07),
    Dense(1, activation='sigmoid')
])

---

Results

Feature Subset	Logistic	SVM	RF	CNN
5% DEGs	0.649	0.674	0.659	0.700
15% DEGs	0.767	0.703	0.662	0.730
25% DEGs	0.790	0.703	0.637	0.790

CNN achieved best performance with 25% selected features.{: .prompt-info}

---

Dockerized for Reproducibility

We created a Docker container to run this project anywhere:

FROM python:3.11-slim
WORKDIR /app
COPY requirements.txt .
RUN pip install --no-cache-dir -r requirements.txt
COPY . .
CMD ["jupyter", "notebook", "--ip=0.0.0.0", "--port=8888", "--allow-root", "--no-browser"]

Run with:

docker build -t tumor-severity-ml .
docker run -p 8888:8888 -v $(pwd):/app tumor-severity-ml

---

Future Directions

• Use DESeq2 for biologically informed feature selection
• Add SHAP for model interpretability
• Run the Nextflow pipeline on HPC to scale Optuna tuning
• Evaluate with confusion matrices and F1-score

---