Tumor reactivity assessment using clonal expression (TRACE) reveals tumor reactive CD8+ T cell heterogeneity across solid tumors

The study introduces TRACE, a robust, open-source single-cell RNA sequencing classifier that accurately identifies tumor-reactive CD8+ T cells across diverse solid tumors without requiring dataset-specific training, thereby enabling the characterization of anti-tumor immune responses and their correlation with immunotherapy outcomes.

The Gist

Imagine your body is a bustling city under siege by a criminal gang (the tumor). The police force sent to stop them are the T cells. However, not all police officers are equally effective. Some are "Special Forces" who know exactly who the criminals are and can take them down. Others are "Bystanders"—officers who are just patrolling the area, looking busy, but have no idea who the bad guys actually are.

For years, doctors have been able to count how many police officers are in the city, but they've struggled to tell the Special Forces apart from the Bystanders. This is a big problem because if you want to boost the immune system to fight cancer, you need to know exactly how many real fighters you have.

This paper introduces a new, high-tech tool called TRACE (Tumor Reactivity Assessment using Clonal Expression) that solves this mystery. Here is how it works, broken down into simple concepts:

1. The Problem: The "Needle in a Haystack"

In a tumor, you have thousands of T cells. Some are tired and exhausted (which is actually a good sign if they are fighting cancer), and some are just resting.

• The Old Way: Scientists used to look for specific "badges" (markers) on the cells to guess if they were fighters. But it's like trying to identify a spy just because they are wearing a trench coat; many innocent people wear trench coats too.
• The Limitation: Previous computer programs tried to guess, but they were trained on very small, specific groups of people. It was like training a dog to recognize only Golden Retrievers, and then expecting it to recognize a Poodle. It often failed when faced with new data.

2. The Solution: TRACE (The Super-Intelligent Detective)

The researchers built TRACE, a machine-learning "detective" that looks at the entire "personality" of a T cell, not just one badge.

• The Training: Instead of training on just one city, they fed the detective data from six different cities (different types of cancer) and thousands of confirmed "Special Forces" clones. They also showed it lots of "Bystanders" so it knows what not to look for.
• The "Clonal" Secret: T cells that are fighting the same cancer often come from the same "family" (they share the same T-cell receptor, or TCR). TRACE is smart enough to look at the whole family. If one cousin is a fighter, TRACE checks the whole family tree to see if they are all in on the action.
• The "Binning" Trick: Imagine trying to compare the volume of music from two different speakers—one is a tiny Bluetooth speaker, the other is a massive stadium system. It's hard to compare the raw numbers. TRACE uses a trick called "expression binning." Instead of counting exact notes, it just asks: "Is the music quiet, medium, or loud?" This makes the data from different labs and machines perfectly comparable, like converting everything to a standard volume setting.

3. How They Tested It (The "Fire Drill")

To prove TRACE works, they didn't just trust the computer. They did a real-life experiment:

1. They took T cells from a melanoma patient.
2. They put them in a room with the patient's own cancer cells (the "fire drill").
3. They watched which T cells actually woke up and attacked (they lit up with a specific protein called 4-1BB).
4. The Result: TRACE looked at the T cells before the drill and correctly predicted which ones would be the attackers with 82% accuracy. It was like a weatherman predicting a storm with high precision before the first raindrop fell.

4. What TRACE Found

When they ran TRACE on hundreds of patient samples from lung, colon, and pancreatic cancers, they found some fascinating things:

• Specificity: TRACE correctly identified the "Special Forces" in the tumor, but it didn't get confused by tired police officers in healthy tissue nearby. It knows the difference between "tired from fighting cancer" and "tired from a long shift."
• The "KRAS" Connection: In lung cancer, patients with a specific mutation (KRAS) had more of these Special Forces. This helps explain why some patients respond better to treatment than others.
• The "MSI" Connection: In colon cancer, tumors with a specific genetic glitch (MSI) were full of highly reactive fighters, confirming why those patients often do well with immunotherapy.

The Bottom Line

Think of TRACE as a universal translator for the immune system.

• Before: Doctors had a blurry photo of the immune army and couldn't tell who was who.
• Now: TRACE gives them a high-definition, color-coded map that highlights exactly which cells are the tumor-fighting heroes.

Because the code and the "brain" of TRACE are open to everyone, scientists and doctors can now use this tool on their own data to figure out which patients might respond best to immunotherapy, and to design better treatments that bring in more of the "Special Forces" to the fight. It turns a guessing game into a precise science.

Technical Summary

1. Problem Statement

Tumor-infiltrating lymphocytes (TILs) are critical for the efficacy of immunotherapies like Immune Checkpoint Blockade (ICB) and Adoptive Cell Therapy (ACT). However, TILs are a heterogeneous mixture of Tumor Reactive T cells (TRTs) that recognize tumor antigens and bystander T cells that do not.

• The Challenge: Distinguishing TRTs from bystanders is difficult. Current methods rely on labor-intensive functional assays, peptide-MHC multimers, or bulk phenotypic proxies that do not scale well with single-cell RNA sequencing (scRNA-seq).
• Limitations of Existing Computational Tools: Previous machine learning approaches (e.g., TRTpred, NeoTCR8, MANAscore) suffer from:
  • Training on small, single-dataset cohorts, limiting generalizability.
  • Lack of shared model weights, hindering reproducibility.
  • Heavy reliance on dataset-specific normalization and batch correction.
  • Failure to explicitly model clonal structure (intra-clonal heterogeneity), often treating cells as independent data points rather than members of a clone.

2. Methodology: The TRACE Framework

The authors developed TRACE (Tumor Reactivity Assessment using Clonal Expression), a machine-learning framework designed to predict tumor reactivity from scRNA-seq data while accounting for clonal relationships and technical variability.

A. Data Aggregation and Labeling

• Training Data: Aggregated six scRNA/scTCR-seq datasets containing experimentally verified TRT and non-TRT labels.
• Scale: The final training set comprised 40,156 cells across 16,465 CD8+ clones, including 9,488 validated TRTs.
• Negative Class: Included verified non-TRTs, healthy PBMCs, and clones with known non-tumor antigens (from VDJdb).
• Splitting Strategy: Train/test splits were performed at the clone level to prevent data leakage (ensuring cells from the same clone do not appear in both training and testing sets).

B. Preprocessing and Feature Engineering

• Clone-Level Aggregation: Instead of analyzing single cells in isolation, TRACE aggregates cells sharing a TCR clonotype.
  • Summarization: Gene expression is summarized using the 75th percentile of expression levels within a clone (found to be optimal over 50th or 100th percentiles).
• Expression Binning: To mitigate batch effects, sequencing depth variations, and platform differences, raw counts are transformed via expression binning (discretizing expression into ordinal bins) rather than log-normalization. This reduced training time by >5x and improved robustness.
• Feature Selection: A fixed set of 50 top-ranked genes was selected based on median feature importance across cross-validation folds. These genes represent a composite phenotype of effector-memory, cytotoxicity, and chronic activation/exhaustion.

C. Model Architecture and Training

• Algorithm: XGBoost (gradient-boosted decision trees) was selected as the optimal framework over Random Forest and AdaBoost.
• Training Strategy: Nested cross-validation (50 random seeds) with systematic hyperparameter optimization (using Optuna/Bayesian optimization).
• Class Imbalance: The positive class (TRT) was balanced at a 10:90 ratio against the negative class during inner-loop training via subsampling.
• Output: A tumor reactivity score (0 to 1) for individual cells or aggregated clones.

3. Key Results

A. Performance Benchmarking

• Metrics: TRACE achieved a mean Matthews Correlation Coefficient (MCC) of 0.84 and an F1-score of 0.85 on held-out test data.
• Comparison: TRACE outperformed or matched existing methods (NeoTCR8, TRTpred, TR30, MANAscore) across multiple datasets.
  • It demonstrated superior specificity, particularly in distinguishing tumor-reactive exhausted cells from bystander exhausted cells in normal tissues.
  • It maintained a low False Positive Rate (FPR) on PBMC data compared to other methods.

B. Experimental Validation

• Co-culture Assay: The authors validated TRACE predictions using an ex vivo melanoma co-culture system (TIL + autologous tumor cells).
• Validation: TRT clones were identified by 4-1BB (TNFRSF9) expression post-co-culture.
• Concordance: TRACE predictions on the initial tumor material showed 82% accuracy for predicting experimentally validated TRTs and 80% accuracy for non-TRTs among 171 shared clones.

C. Biological Insights Across Tumor Atlases

TRACE was applied to hundreds of patient samples across lung (NSCLC), colorectal (CRC), and pancreatic (PDAC) cancer atlases:

• Exhaustion Specificity: High TRACE scores were enriched in terminally exhausted (GZMK+, HAVCR2+, CXCL13+) and proliferating CD8+ T cells within tumors. Crucially, these high scores were absent in similar exhausted states in normal adjacent tissues or healthy controls, proving specificity to tumor reactivity rather than general exhaustion.
• Clonal Expansion: TRACE+ clones were significantly more likely to be clonally expanded (>5 cells), consistent with antigen-driven expansion.
• Clinical Correlations:
  • KRAS Mutations: NSCLC samples with KRAS mutations showed higher sample-level TRACE scores (sTRACE).
  • MSI Status: Colorectal tumors with Microsatellite Instability (MSI) had significantly higher sTRACE scores than Microsatellite Stable (MSS) tumors, correlating with known higher immunogenicity.
  • PDAC: Low overall scores in primary PDAC (consistent with "cold" tumors), but higher scores in metastatic lesions.

4. Key Contributions

1. Robust, Generalizable Model: TRACE is trained on a multi-dataset, multi-indication aggregated cohort, overcoming the "single-study" bias of previous tools.
2. Clonotype-Aware Architecture: By aggregating cells at the clone level and modeling intra-clonal heterogeneity, TRACE better reflects biological reality than single-cell-only approaches.
3. Technical Robustness: The use of expression binning and the elimination of complex normalization steps allow TRACE to be applied directly to raw counts from diverse sequencing platforms (10x, BD Rhapsody, Smart-seq2, etc.) without extensive batch correction.
4. Open Science: The authors released the code, model weights, and feature sets openly, enabling the community to apply, benchmark, and update the model.

5. Significance

TRACE provides a scalable, computationally efficient, and biologically validated tool for quantifying the fraction of tumor-reactive T cells in complex immune infiltrates.

• For Biomarker Discovery: It enables the correlation of TRT frequency with clinical response to immunotherapies across diverse cancer types.
• For Therapy Development: It aids in the selection of high-quality TIL products for ACT by identifying and enriching for truly reactive clones.
• For Research: It offers a standardized method to dissect the heterogeneity of the tumor microenvironment, distinguishing between "bystander" inflammation and true anti-tumor immunity.

In summary, TRACE represents a significant advancement in computational immunology, moving from single-dataset heuristics to a robust, open-source, clone-aware framework for identifying tumor-reactive T cells.