sigNATURE maps cohort-specific T-cell states to reproducible programs of ICI response

The paper introduces sigNATURE, a reference-guided framework that maps T-cell states to reproducible atlas-aligned programs, thereby significantly improving the cross-cohort reproducibility, separation, and predictive accuracy of immune checkpoint inhibitor response biomarkers compared to traditional cohort-specific definitions.

The Gist

The Big Problem: "Lost in Translation" in Cancer Research

Imagine you are trying to find the perfect recipe for a cake that cures a specific disease. You ask three different chefs (scientists) for their recipes.

• Chef A says, "You need exactly 2 cups of flour and a pinch of salt."
• Chef B says, "You need 3 cups of flour and no salt."
• Chef C says, "It depends on the brand of flour you use."

If you try to follow these recipes, you get confused. Why? Because each chef defined "flour" and "salt" differently based on their own kitchen (their specific group of patients). In the world of cancer immunotherapy, scientists are trying to find the "recipe" for why some patients respond to treatment (the "Responders") and others don't (the "Non-Responders"). They look at T-cells (the body's immune soldiers) to find the clues.

The problem is that every study defines these T-cells differently. One study might call a cell a "Warrior," while another calls the exact same cell a "Scout." Because they use different names and definitions, the "clues" (biomarkers) they find don't match up when you try to combine the studies. It's like trying to build a house using blueprints where "wall" means something different in every single drawing.

The Solution: sigNATURE (The Universal GPS)

The authors created a new tool called sigNATURE. Think of it as a Universal GPS and Translator for T-cells.

Instead of letting every study invent its own map, sigNATURE forces every study to use the same master map (a massive "Atlas" of T-cells built from hundreds of thousands of samples).

Here is how it works, step-by-step:

1. The Master Atlas (The Reference Library):
   Imagine a giant library containing every possible type of T-cell that exists in humans, organized perfectly. This is the "Atlas." It has clear labels like "Exhausted Soldier," "Fresh Recruit," or "Special Ops."

2. The Translation (Mapping):
   When a new study comes in with their own messy data, sigNATURE doesn't let them keep their own confusing labels. Instead, it takes their cells and asks the Atlas: "Hey, which of your perfect categories does this cell look most like?" It translates the messy local data into the clean, universal language of the Atlas.

3. The Confidence Score (The "Identifiability" Check):
   This is the clever part. Sometimes, a cell is right on the border between two categories. It's a bit of a "Warrior" and a bit of a "Scout."

   • Old way: Scientists would just force a label on it and hope for the best.
   • sigNATURE way: It gives the cell a Confidence Score. If the cell is clearly a "Warrior," the score is high. If it's a confused mix, the score is low. This helps scientists know which clues are solid and which are just noise.

What Did They Find?

The researchers tested this new GPS system on two different groups of cancer patients (Lung Cancer and Skin Cancer).

• Before sigNATURE: When they looked at the data using the old, messy methods, the "Responders" and "Non-Responders" looked like a mixed-up bowl of marbles. You couldn't tell them apart. The prediction accuracy was basically a coin flip (46%).
• After sigNATURE: Once they translated the data into the Universal Atlas language, the marbles sorted themselves out! The "Responders" formed a neat pile on one side, and the "Non-Responders" on the other. The prediction accuracy jumped up to 75%.

They also discovered that some famous "clues" from previous studies were actually misleading. When mapped to the Universal Atlas, those clues turned out to be scattered across many different cell types, meaning they weren't reliable. However, they found two new, reliable clues:

1. Terminally Differentiated CD8 T-cells: These are the "Super Soldiers" that are fully trained and ready to fight.
2. Regulatory CD4 T-cells: These are the "Referees" that control the game.

Why This Matters

This paper is a game-changer because it stops scientists from arguing over definitions.

• Reproducibility: Now, if a scientist in New York finds a clue, a scientist in London can check it against the same Universal Atlas and know if it's real.
• Better Medicine: By finding the real patterns that work across all patients (not just one specific group), we can get closer to predicting who will survive cancer treatment and who needs a different approach.

In short: sigNATURE is the tool that stops scientists from speaking different languages. It gives them a shared dictionary and a shared map, so they can finally agree on what makes a cancer treatment work.

Technical Summary

1. Problem Statement

Immune checkpoint inhibitors (ICIs) have revolutionized cancer therapy, but predicting clinical response remains challenging. While single-cell RNA sequencing (scRNA-seq) has identified T-cell states associated with ICI response, these biomarkers suffer from poor reproducibility across cohorts.

• Root Cause: Most existing studies define T-cell states de novo within each specific cohort using clustering algorithms. Because cluster boundaries are highly sensitive to cohort composition, sample size, and analytic workflows, "response markers" often reflect cohort-specific artifacts rather than conserved biological programs.
• Consequence: Reported signatures (e.g., specific gene sets for exhausted or precursor T-cells) often fail to generalize across different cancer types, treatment regimens, or even different studies of the same cancer type. Furthermore, existing methods lack a mechanism to quantify the confidence of mapping a query cell to a reference state, making it difficult to distinguish robust biological states from ambiguous assignments.

2. Methodology: The sigNATURE Framework

The authors developed sigNATURE (Signature Normalization and Atlas-based T-cell Understanding for Reproducibility and Evaluation), a reference-guided framework designed to harmonize scRNA-seq data across studies.

• Reference Atlas: The framework utilizes a large, high-resolution T-cell reference atlas (TCellMap) comprising ~308,000 T cells (CD4+ and CD8+) from 324 individuals across 16 cancer types. This atlas defines 14 distinct CD8+ and multiple CD4+ T-cell states based on conserved differentiation trajectories.
• Workflow:
  1. Normalization & Preprocessing: Query datasets (e.g., from clinical trials) undergo quality control and cell-cycle regression to remove technical artifacts that distort clustering.
  2. Label Transfer & Projection: Query cells are mapped onto the reference atlas using Seurat's FindTransferAnchors and MapQuery. Cells are projected into the reference's UMAP space.
  3. k-NN Consensus Classification: Instead of relying solely on standard label transfer, sigNATURE employs a custom k-Nearest Neighbor (k-NN) consensus classifier. For each query cell, the 10 nearest neighbors in the reference are identified. A cell is assigned a label only if a majority (≥5/10) of neighbors share the same annotation. Cells failing this threshold are marked as "Not Assigned" (NA), flagging ambiguous or transitional states.
  4. Identifiability Score: A novel metric is introduced to quantify mapping confidence. It calculates the proportion of a cell's nearest neighbors that belong to the assigned cluster. High identifiability indicates a robust, well-defined state; low identifiability suggests the cell lies in a transitional or ambiguous region.
  5. Predictive Modeling: Cell state proportions (aggregated to the patient level) are used as features in logistic regression models to predict ICI response (Responder vs. Non-Responder).

3. Key Contributions

• Reference-Guided Harmonization: Moves away from de novo cohort clustering toward a unified, atlas-aligned coordinate system, enabling direct comparison of T-cell programs across independent studies.
• Confidence-Aware Mapping: Introduces the Identifiability Score, which distinguishes between robust, atlas-resolvable states and ambiguous mappings, filtering out noise caused by cohort-specific boundary definitions.
• Reproducible Biomarker Prioritization: Demonstrates that many previously published "response markers" map to heterogeneous or non-overlapping regions in the reference atlas, whereas sigNATURE isolates specific, conserved states (e.g., terminally differentiated effector CD8+ and regulatory CD4+ T cells) that are consistently associated with response.

4. Results

The framework was validated using two independent ICI scRNA-seq cohorts:

1. Liu et al. (NSCLC): 36 patients (33 pre-treatment, 14 post-treatment).
2. Yost et al. (Basal Cell Carcinoma): 11 patients.

Key Findings:

• Reproducibility of Markers: When published signatures (e.g., Sade-Feldman's TCF7/IL7R set vs. Liu's GZMK set) were projected onto the reference atlas, they mapped to distinct, non-overlapping clusters, confirming that cohort-specific definitions fragment conserved biological programs.
• Improved Resolution: sigNATURE resolved cohort-defined "exhausted" clusters into finer, functionally distinct sub-states (e.g., separating proliferative exhausted p-Tex from terminally exhausted Tex), which were previously conflated in cohort-specific clustering.
• Predictive Performance:
  • In the Liu cohort, sigNATURE-derived features improved the Area Under the Curve (AUC) for response prediction from 0.472 (cohort-derived states) to 0.528.
  • In the independent Yost validation cohort, the improvement was dramatic: AUC increased from 0.467 (cohort-derived) to 0.967 (sigNATURE-derived).
  • The predictive signal was distributed across the entire feature space (27 granular states) rather than relying on a few dominant features, indicating a structured representation of biological heterogeneity.
• Unsupervised Separation: Dimensionality reduction of sigNATURE-derived features showed clear spatial separation between responders and non-responders in both cohorts and in joint analysis, whereas cohort-derived features showed significant overlap.
• Identifiability Insights: The framework identified Terminally Differentiated Effector (t-Teff) CD8+ and Regulatory (Treg) CD4+ T cells as the most robust, high-identifiability states associated with response, filtering out markers that were enriched but mapped to ambiguous atlas regions.

5. Significance

• Paradigm Shift: sigNATURE shifts the focus from "which genes predict response in this dataset" to "which conserved cellular programs are engaged across studies."
• Clinical Utility: By improving the cross-cohort comparability of single-cell data, the framework offers a more stable basis for developing biomarkers for patient stratification and clinical trial design.
• Mechanistic Interpretability: The ability to distinguish robust atlas-resolved states from ambiguous transitions provides deeper mechanistic insights into how T-cell states drive ICI efficacy, reducing false discoveries driven by technical or cohort-specific artifacts.
• Generalizability: While tested on ICI response, the framework is applicable to any single-cell study requiring cross-dataset integration and reproducible cell state annotation.

In summary, sigNATURE provides a principled, reference-guided solution to the reproducibility crisis in single-cell immunology, transforming unstable cohort-specific findings into robust, biologically interpretable programs that can be reliably translated into clinical biomarkers.