PETIL: Predicting Expansion of Tumor Infiltrating Lymphocytes for the Adoptive Cell Immunotherapy in Bladder Cancers

The PETIL machine-learning model, developed using data from Moffitt Cancer Center, successfully predicts the expansion of tumor-infiltrating lymphocytes in bladder cancer patients to optimize patient selection for adoptive cell immunotherapy, achieving strong performance metrics (AUC up to 0.857) in both testing and blinded validation cohorts.

The Gist

Imagine you have a garden (your body) that has been invaded by weeds (cancer). One of the most promising ways to fight these weeds is to send in a specialized cleanup crew: your own immune cells, specifically the "Tumor-Infiltrating Lymphocytes" (TILs). These are soldiers already inside the tumor that you can pull out, multiply in a lab until you have an army, and then send back in to wipe out the cancer.

The Problem: The "Will It Grow?" Gamble
Here's the catch: Before you can send the army back, you have to take a sample of the tumor and try to grow these cells in a lab. This process takes about 4 to 6 weeks.

• The Good News: About 70% of the time, the cells grow like wildflowers, and you get a massive army to fight the cancer.
• The Bad News: About 30% of the time, the cells refuse to grow. They just sit there and die.

If you wait 6 weeks only to find out the cells won't grow, the patient has lost precious time. They might need a different treatment immediately, but they've been stuck waiting. It's also expensive and emotionally draining to wait for a result that turns out to be a "no."

The Solution: The Crystal Ball (PETIL)
The researchers at Moffitt Cancer Center wanted to stop the guessing game. They asked: "Can we look at the patient and the tumor before we even start the lab work and predict if the cells will grow?"

They built a computer program called PETIL (Predictor of Expansion of Tumor Infiltrating Lymphocytes). Think of PETIL as a highly trained weather forecaster for your immune system.

How PETIL Works (The Recipe)
Instead of using complex, expensive genetic tests that require new data, PETIL looks at information doctors already have on hand. It's like a chef who can tell you if a cake will rise just by looking at the ingredients you already have in the pantry.

PETIL looks at five specific "ingredients":

1. Age: How old is the patient?
2. BMI: What is their body mass index?
3. Tumor Weight: How heavy is the piece of tumor being tested?
4. Tumor Digest Count: How many tiny pieces was the tumor cut into?
5. Fragments Plated: How many of those tiny pieces were put into the lab dishes?

The "Smart" Part
The researchers didn't just guess which of these 5 ingredients mattered. They used a machine learning technique called Forward Feature Selection. Imagine you have a bag of 15 different spices. You want to make the perfect soup, but you don't know which ones to use.

• The computer tries adding spices one by one.
• It realizes that adding "Salt" and "Pepper" makes the soup great.
• But adding "Cinnamon" or "Nutmeg" doesn't help, or even makes it worse.
• So, it stops and says, "Okay, we only need these two spices to make a perfect soup."

PETIL did the same thing. It started with 15 different data points (age, race, cancer stage, etc.) and realized that only 5 of them were actually needed to make a good prediction.

The Results: A Winning Streak
The team tested their "weather forecast" on two groups of patients:

1. The Test Group: They looked at past data where they already knew the outcome. PETIL was right 74% of the time.
2. The Blind Test (The Real Challenge): They took data from a brand new clinical trial where they didn't know the outcome yet. They ran the prediction, and then checked the results. PETIL was right 85.7% of the time (12 out of 14 patients).

Why This Matters
Think of PETIL as a traffic light for cancer treatment:

• Green Light: The computer predicts the cells will grow. The doctor says, "Great! Let's start the 6-week lab process. We have a good chance of success."
• Red Light: The computer predicts the cells won't grow. The doctor says, "Don't waste 6 weeks waiting. Let's switch to a different treatment plan right now."

In a Nutshell
This paper is about building a smart tool that saves patients time and money. By looking at simple, everyday data points (like age and tumor size), the PETIL model can predict whether a patient's immune cells will successfully grow into an army. This helps doctors make faster, better decisions, ensuring patients get the right treatment sooner without the painful wait of a failed lab experiment.

Technical Summary

1. Problem Statement

Adoptive Cell Therapy (ACT) using Tumor-Infiltrating Lymphocytes (TILs) is a promising personalized immunotherapy for solid tumors, including bladder cancer. The process involves resecting a tumor, expanding TILs ex vivo (taking 4–6 weeks), and reinfusing them into the patient.

• The Challenge: Approximately 30% of bladder tumor specimens fail to yield successful TIL expansion.
• The Consequence: Patients whose tumors fail to expand undergo a 4–6 week delay in treatment, incurring high costs and missing the window for alternative therapies.
• The Gap: There is currently no reliable pre-treatment tool to predict whether a specific patient's tumor specimen will successfully expand TILs, leading to inefficient patient stratification.

2. Methodology

The authors developed PETIL (Predictor of Expansion of Tumor Infiltrating Lymphocytes), a machine learning (ML) pipeline optimized for medium-sized clinical datasets. The study utilized a retrospective cohort of 106 bladder cancer patients from Moffitt Cancer Center (2015–2022) and a blinded validation cohort of 14 patients from a Phase I clinical trial (2023–2025).

Data Preprocessing & Feature Engineering

• Input Data: 15 features were collected, categorized as:
  • Demographics: Age, BMI, Race, Smoker status.
  • Clinical: Clinical stage (cT), Pathological stage (pT/pN), Surgery type, Neoadjuvant chemotherapy (NAC), Histology types.
  • Specimen-based: Tumor sample weight, Number of fragments plated, Tumor digest count.
• Handling Missing Data: The dataset had ~4.4% missingness. The authors employed MIDAS (Multiple Imputation with Denoising Autoencoders), a deep learning approach, to impute missing values while preserving inter-feature correlations.
• Normalization: Data was normalized using MaxAbsScaler to ensure equal feature contribution without altering distribution shapes.
• Collinearity Reduction: Pearson correlation analysis identified collinear features. Using Mutual Information (MI) scores to rank importance, the feature "cT or pT" was removed due to high correlation with "pT" and "cT," reducing the feature space from 15 to 14.

Model Development Pipeline

1. Sample Size Adequacy Analysis: Learning curves were generated for Logistic Regression, Random Forest (RF), Gradient Boosting, and Radial Basis Function Support Vector Machines (RBF-SVM). RBF-SVM was selected as the optimal classifier because it showed high generalization and low variance on the 74-patient training set, whereas other models showed signs of overfitting or underfitting.
2. Feature Selection (FFS): A Forward Feature Selection algorithm using a Random Forest classifier was applied to identify the minimal robust subset of features.
   • Features were ranked by Feature Importance Score (FIS).
   • Features were added sequentially, and validation accuracy was monitored.
   • Result: A subset of 5 features was identified as optimal: Age at surgery, BMI, Number of fragments plated, Tumor sample weight, and Tumor digest count.
3. Hyperparameter Optimization: The RBF-SVM hyperparameters ($C$ for margin width and $\gamma$ for nonlinearity) were tuned using 5-fold cross-validation.
4. Handling Class Imbalance: The dataset was imbalanced (Yes-TIL vs. No-TIL). Instead of a standard 0.5 threshold, the authors used the Matthews Correlation Coefficient (MCC) to determine the optimal decision boundary threshold. This maximized the correlation between observed and predicted binary classifications.

3. Key Contributions

• PETIL Model: A novel, data-driven ML pipeline specifically designed for medium-sized clinical datasets (n < 150), avoiding the need for the massive datasets typically required by deep learning.
• Minimal Feature Set: Identification of a 5-feature model (2 demographic, 3 specimen-based) that is sufficient for high-accuracy prediction, simplifying clinical implementation.
• Robust Imputation Strategy: Application of MIDAS to handle missing clinical data without discarding valuable patient records.
• Blinded Validation: Successful validation on a completely independent, blinded clinical trial cohort, demonstrating real-world applicability.

4. Results

The model was evaluated on a testing cohort (32 patients) and a blinded validation cohort (14 patients).

• Testing Cohort Performance:
  • Accuracy: 0.750
  • Sensitivity (True Positive Rate): 0.762
  • Specificity (True Negative Rate): 0.727
  • AUC (Area Under Curve): 0.740
• Blinded Validation Cohort Performance:
  • The clinical trial cohort consisted of 14 patients, all of whom had successful TIL expansion (Yes-TIL).
  • PETIL correctly predicted 12 out of 14 cases.
  • Accuracy: 0.857
• Feature Importance: The model confirmed that biological specimen characteristics (weight, digest count, fragments plated) combined with patient demographics (Age, BMI) were the strongest predictors, rather than complex clinical staging variables.

5. Significance

• Clinical Decision Support: PETIL provides clinicians with a tool to stratify patients before the 4–6 week expansion process begins. Patients predicted to have "No-TIL" expansion can be immediately transitioned to alternative therapies, avoiding treatment delays.
• Cost and Resource Efficiency: By identifying non-responders early, the model prevents the high costs associated with failed TIL manufacturing processes.
• Generalizability: The pipeline is designed to be adaptable to other solid tumors. The authors note that while the current dataset is the largest single-institution collection of bladder TIL specimens, the methodology (FFS + RBF-SVM + MCC) is transferable.
• Overcoming Data Scarcity: The study demonstrates that high-performance predictive models can be built on "medium-sized" local clinical data without requiring external large-scale datasets, provided rigorous feature selection and imputation techniques are used.

In conclusion, PETIL represents a significant step toward precision medicine in bladder cancer immunotherapy, offering a computationally efficient method to optimize patient selection for ACT-TIL.