Integrative single-cell profiling of melanoma reveals a tumor microenvironment signature predictive of immunotherapy response

This study identifies a 13-gene signature derived from integrative single-cell and bulk transcriptomic analyses of melanoma that effectively predicts immunotherapy response across multiple cancer types, demonstrating particular clinical utility in stratifying patients with low tumor mutational burden.

The Gist

The Big Picture: Why Do Some Cancer Treatments Work and Others Don't?

Imagine the immune system as a highly trained police force, and cancer cells as criminals hiding in a city (the body). In recent years, doctors have developed "immunotherapy" drugs. These drugs act like a megaphone, shouting to the police, "Wake up! Go arrest those criminals!"

However, there is a problem: The megaphone doesn't work for everyone. About half of the patients with melanoma (a type of skin cancer) or bladder cancer don't respond to this treatment. They are like a police force that hears the shout but stays asleep.

Doctors currently try to guess who will respond by looking at a few clues, like how many "typos" (mutations) are in the cancer's DNA. But this clue is often unreliable. It's like trying to predict if a car will start just by looking at the color of the paint; sometimes it works, but often it doesn't.

What This Study Did: A Detective's Deep Dive

The researchers in this paper wanted to find a better clue. Instead of just looking at the whole "city" (the tumor), they decided to zoom in and look at the individual "police officers" (immune cells) living inside the tumor before the treatment even started.

They used a high-tech microscope called single-cell RNA sequencing. Think of this as taking a photo of every single cell in the tumor and reading its "instruction manual" (its genes) to see what it was thinking and doing.

The Investigation Process:

1. The Discovery Team: They looked at data from 41 melanoma patients. They split them into two groups: those who got better (Responders) and those who didn't (Non-Responders).
2. The Comparison: They compared the "instruction manuals" of the immune cells in the "Got Better" group against the "Didn't Get Better" group.
3. The Validation: They checked if their findings held up in a second group of 19 patients, and then again in a massive group of 128 patients using a different, broader method (bulk RNA sequencing). Finally, they tested their theory on bladder cancer patients to see if it worked there too.

The Key Findings: What Was Different?

The researchers found that the immune cells in the patients who didn't respond had a very specific "personality" or "mood" before treatment even began.

1. The "Exhausted" Police Officers
In the patients who didn't respond, the T-cells (a key type of immune cell) were wearing too many "stop signs" on their uniforms. Specifically, they had high levels of genes from the TNFRSF family (like TNFRSF9, TNFRSF4, and TNFRSF18).

• The Analogy: Imagine a police officer who is so tired and worn out that they have put up "Do Not Disturb" signs on their helmet. Even if you shout at them with the megaphone (the drug), they are too exhausted to move. The study found that these "exhausted" cells were actually more common in the people who didn't get better.

2. The "Bad Neighborhood" Signs
The study also found that in non-responders, the tumor environment was full of signals that told the immune system to stand down or hide. It was like the criminals had built a high wall and put up "No Entry" signs everywhere.

3. The 13-Gene "Crystal Ball"
From all the thousands of genes they looked at, the researchers narrowed it down to just 13 specific genes.

• These 13 genes act like a 13-point checklist.
• If a patient's tumor has this specific combination of genes, it is highly likely they will respond to the immunotherapy.
• This "checklist" worked well for melanoma patients (73% accuracy) and also worked surprisingly well for bladder cancer patients (64% accuracy), even though they are different types of cancer.

The Special Case: The "Low Mutations" Group

Usually, doctors only give these powerful drugs to patients who have a lot of DNA "typos" (high Tumor Mutational Burden or TMB). They think patients with fewer typos won't respond.

However, this study found something interesting: The 13-gene checklist could spot "hidden responders."

• Even among patients with low numbers of DNA typos (who doctors usually say won't respond), this checklist successfully identified the ones who would get better.
• The Analogy: Imagine a security guard who usually only lets people in if they have a VIP badge (high mutations). This new checklist is like a smart scanner that realizes, "Hey, even without the VIP badge, this person has the right uniform and attitude to get in." It found 20 out of 21 patients who would have been turned away by the old rules.

What the Paper Does Not Say

It is important to stick to what the paper actually claims:

• It is not a new drug: This study did not create a new medicine. It created a diagnostic tool (a way to read the genes) to predict who needs the medicine.
• It is not a guarantee: The model predicts who is likely to respond, but it is not 100% perfect.
• It is not yet standard practice: The authors state this needs further validation before it can be used in hospitals to make real treatment decisions.

Summary

Think of this study as finding a new weather forecast for cancer treatment.

• Old Forecast: "If the sky is very cloudy (high mutations), it will rain (treatment works)."
• New Forecast: "Even if the sky isn't very cloudy, if you look at the specific wind patterns and humidity (the 13 genes), we can tell you exactly who will get wet (respond to treatment) and who won't."

This helps doctors avoid giving expensive, heavy treatments to people who won't benefit, and ensures that people who might benefit (even if they look like they shouldn't) get the help they need.

Technical Summary

Problem Statement
Immune checkpoint inhibitors (ICIs) have revolutionized the treatment of advanced melanoma and urothelial cancer; however, a significant proportion of patients fail to respond, with overall response rates (ORR) ranging from 17% to 58% depending on the tumor type. Current predictive biomarkers, such as tumor mutational burden (TMB) and PD-L1 expression, suffer from inconsistency across tumor types, intratumoral heterogeneity, and limited predictive efficacy as standalone tools. Furthermore, resistance often stems from pre-existing immune dysfunction and immunosuppressive mechanisms within the tumor microenvironment (TME) present before treatment initiation. While transcriptomics and proteomics have identified some molecular features associated with response, existing studies often lack detailed, cell-type-specific insights into the pre-treatment TME. There is a critical need to identify robust, pre-treatment molecular signatures that can stratify responders from non-responders, particularly in patients with low TMB who might otherwise be excluded from ICI therapy.

Methodology
The study employed a stepwise, integrative analysis of publicly available transcriptomic datasets to identify predictive features of ICI response.

• Data Sources: The analysis utilized single-cell RNA sequencing (scRNA-seq) data from 60 melanoma patients (41 in a discovery set, 19 in a validation set) and bulk RNA-seq data from 128 melanoma patients and 298 bladder cancer patients (IMvigor210 cohort). All datasets consisted of baseline (pre-treatment) samples.
• Single-Cell Processing: scRNA-seq data were processed using the Seurat package in R. Quality control filters removed low-quality cells, doublets, and erythrocytes. Batch effects were corrected using the Harmony package. Cell types were annotated using marker genes and the SingleR package, with specific attention to immune populations (T-cells, B-cells, NK cells, monocytes).
• Differential Expression Analysis: A pseudobulk approach was used to aggregate gene expression counts at the patient level to avoid false-positive inflation. Differential expression between responders (Rs) and non-responders (NRs) was assessed using DESeq2 for the discovery cohort and limma for the validation cohort, with study type included as a covariate.
• Model Development: Genes consistently differentially expressed across single-cell discovery, single-cell validation, and bulk melanoma cohorts were selected to construct a logistic regression predictive model. The model was trained on the melanoma bulk dataset using 10-fold cross-validation.
• Validation and Correlation: The model was externally validated in the independent bladder cancer cohort. Performance was evaluated using Area Under the Curve (AUC), sensitivity, and specificity. Additionally, the model's signature score was correlated with the whole transcriptome to perform Gene Ontology (GO) pathway enrichment analysis.

Key Contributions and Results

• Cell-Type Specific Signatures: Integrative analysis revealed that responders exhibit distinct molecular profiles across multiple immune cell types compared to non-responders. Notably, responders showed downregulation of TNFR superfamily genes (TNFRSF18, TNFRSF9, TNFRSF4) and other immunosuppressive genes (e.g., LGALS1, BATF, IL12RB2) in T-cells. These genes were found to be highly expressed in regulatory and exhausted T-cell states.
• The 13-Gene Signature: Thirteen transcripts were identified as consistently associated with response across all melanoma cohorts (single-cell and bulk). These genes included:
  • Upregulated in Responders: SELL, EPB41, CD96, UHRF2 (in T-cells), ALDH5A1 (B-cells), TLR10, ST6GAL1, IKZF1, MPRIP (Monocytes).
  • Downregulated in Responders: LINGO1 (T-cells), LGALS1 (T-cells), PLEC, PDGFRB (NK cells).
• Predictive Performance: A logistic regression model based on these 13 biomarkers (13BM) achieved an AUC of 0.73 in the melanoma bulk cohort. When applied to the independent bladder cancer cohort, the model maintained significant predictive power with an AUC of 0.64.
• Performance in Low-TMB Subgroups: The model demonstrated enhanced utility in patients with low tumor mutational burden (TMB < 10 mut/Mb), achieving an AUC of 0.75 in this subgroup. At a threshold optimized for high sensitivity (~90%), the model correctly identified 95.2% of responders in the low-TMB bladder cancer cohort, suggesting potential to identify candidates for ICI who would be excluded by current TMB-based criteria.
• Biological Correlates: In melanoma, the 13BM score correlated positively with immune-related pathways (T-cell differentiation, interferon response), reflecting an "immune-hot" phenotype. In bladder cancer, it correlated with metabolic processes (oxidative phosphorylation). In both tumor types, the score negatively correlated with extracellular matrix (ECM) organization, epithelial-to-mesenchymal transition (EMT), and TGF-β response, processes linked to immune evasion.

Significance and Claims
The authors claim that this study identifies pre-treatment molecular features related to immune-cancer crosstalk that are associated with ICI response. By leveraging single-cell data to derive a bulk-tissue detectable signature, the study provides a 13-gene model that offers potential added clinical value in stratifying responders. The authors emphasize that the model is particularly relevant for patients with low TMB, a group where current biomarkers often fail. The findings regarding the downregulation of TNFRSF genes in responders suggest that the overexpression of these co-stimulatory receptors may reflect dysfunctional T-cell activation and exhaustion, offering a biological rationale for the limited efficacy of agonists targeting these receptors in clinical trials. The study concludes that further validation of the model's clinical utility and spatial expression patterns is warranted.