Contrastive Alignment of Expression and Copy Number Highlights Dosage-Insensitive Genes in Cancer

This paper introduces a contrastive learning framework that aligns single-cell RNA-seq expression with copy number variations to identify dosage-insensitive genes in lung adenocarcinoma, revealing specific regulatory escape and compensation mechanisms that persist despite genomic alterations.

The Gist

Imagine a cancer cell as a chaotic construction site.

Normally, in a healthy building, if you have two blueprints (copies of a gene), you build two walls. If you lose a blueprint (a deletion), you build one wall. If you get three blueprints (an amplification), you build three walls. This is the expected rule: More blueprints = More walls. This is called "dosage."

But in cancer, the construction site gets messy. Sometimes, the workers ignore the blueprints. They might have three blueprints for a wall but only build one. Or they might have no blueprints at all but somehow build a whole tower anyway.

This paper is about finding the workers who are ignoring the blueprints.

The Problem: Noise vs. Rebellion

Scientists have known for a long time that cancer cells often have messed-up blueprints (Copy Number Variations, or CNVs). They also know that sometimes the cell's behavior (gene expression) doesn't match those blueprints.

The hard part is telling the difference between:

1. Technical Noise: The camera is blurry, so the photo looks wrong.
2. Real Rebellion: The cell is actively ignoring the blueprint because it has a secret plan (regulatory compensation).

Most methods just look at the "walls" (gene expression) and guess. This paper introduces a new way to look at both the blueprints and the walls simultaneously to see who is actually following orders and who is rebelling.

The Solution: A "Matchmaking" Algorithm

The authors built a smart computer program called CLCC (Contrastive Learning for CNV–Expression Concordance). Think of it as a high-tech dating app for cells, but with a twist.

The Analogy: The "Hard Negative" Date
Imagine you are trying to teach a robot to recognize couples.

• Normal Training: You show the robot a photo of a couple holding hands (Expression) and their matching ID cards (CNV). The robot learns: "These go together."
• The Innovation (Hard Negative Mining): The robot is also shown a tricky test. You give it a photo of a couple (Cell A) and the ID card of a different couple (Cell B) who happen to look very similar on paper (similar blueprints) but are actually total strangers in real life (different behavior).

The robot has to learn: "Wait, even though their ID cards look similar, these two people are totally different. I need to push them apart in my mind."

By forcing the AI to learn these "near-miss" mistakes, it becomes incredibly good at spotting the subtle differences between cells that should be the same but aren't.

What They Found

They tested this on lung cancer cells from 10 patients (about 80,000 cells). They sorted the cells into two groups:

1. The "Conformists": Cells where the walls match the blueprints perfectly.
2. The "Rebels" (Discordant Cells): Cells where the walls don't match the blueprints.

When they looked closely at the Rebels, they found two distinct groups of genes:

1. The "Escape Artists" (Upregulated)

These are genes that the cell is pumping out way too much, even though the blueprints say it shouldn't.

• The Metaphor: Imagine a factory that is supposed to make 100 toys, but the manager (the blueprint) says "Make 10." The workers (the cell) ignore the manager and make 1,000 toys anyway.
• The Result: The rebels were full of "Macrophage" and "Immune Evasion" genes. These are like cells putting on a disguise to hide from the body's immune system. They are acting like immune cells to blend in and avoid being attacked.

2. The "Compensators" (Downregulated)

These are genes that the cell is turning down, even though the blueprints say it should be loud.

• The Metaphor: The blueprint says "Scream at the top of your lungs!" but the cell whispers instead.
• The Result: These were genes related to T-Cells (the body's police force). The cancer cells are actively silencing the "police" genes to stop the immune system from noticing them.

Why This Matters

This is a big deal because it changes how we look at cancer tumors.

• Old View: A tumor is just a blob of cells with broken DNA.
• New View: A tumor is a mix of "Conformists" (who follow the broken DNA rules) and "Rebels" (who have found a way to ignore the rules and survive).

The "Rebels" are likely the most dangerous part of the tumor. They are the ones hiding from the immune system and resisting treatment. By finding these specific genes (like VSIG4, TREM2, and CD8A), doctors might be able to:

1. Identify which patients have these sneaky "Rebel" cells.
2. Develop new drugs that force the Rebels to listen to their blueprints again, or attack them specifically.

The Bottom Line

This paper created a new "lens" to look at cancer. Instead of just counting broken blueprints, they taught a computer to spot the cells that are ignoring the blueprints. This helps us find the "super-survivors" in the tumor that are hiding from our immune system, giving us new targets for curing cancer.

Technical Summary

1. Problem Statement

Somatic copy-number variations (CNVs) are a hallmark of cancer, often driving gene expression changes through dosage effects. However, this relationship is not strictly deterministic; tumors frequently exhibit transcriptional adaptation or buffering, where gene expression deviates from copy-number expectations.

• The Challenge: Identifying these "dosage-insensitive" genes (those that escape or compensate for CNV changes) is difficult in heterogeneous single-cell RNA-seq (scRNA-seq) data.
• The Limitation of Current Methods: Standard approaches struggle to distinguish true regulatory escape from technical noise or orthogonal biological factors (e.g., cell lineage, microenvironment) because they typically analyze expression or CNV in isolation.
• The Goal: Develop a method to quantify the concordance (or discordance) between a cell's inferred CNV profile and its gene expression profile at single-cell resolution to identify genes that systematically decouple from dosage effects.

2. Methodology: CLCC Framework

The authors introduce CLCC (Contrastive Learning for CNV–Expression Concordance), a representation-learning framework that aligns single-cell gene expression profiles with inferred CNV patterns in a shared latent space.

A. Data Preprocessing & CNV Inference

• Dataset: 10 lung adenocarcinoma (LUAD) patients from the GSE131907 atlas (~80k cells), including matched tumor and normal samples.
• CNV Inference: Used infercnvpy to infer CNV profiles by comparing tumor cell expression against normal reference cells. Profiles were smoothed using a sliding window (250 genes) across genomic coordinates.
• Clustering: Cells were clustered based on CNV profiles (not expression) to define subclusters representing distinct chromosomal alteration signatures.

B. Model Architecture

The framework employs two parallel encoders mapping high-dimensional inputs to a shared 128-dimensional latent space:

1. Expression Encoder: A feed-forward network processing normalized gene expression vectors ($G \approx 30,000$ genes).
2. CNV Encoder: A shallower network processing normalized CNV vectors ($W \approx 1,500$ genomic windows), reflecting the smoother structure of CNV data.

• Output: Both encoders produce L2-normalized unit vectors.

C. Loss Functions & Hard Negative Mining

The core innovation is the use of Hard Negative Mining to explicitly teach the model what "discordance" looks like.

1. InfoNCE Loss (Contrastive Alignment): Standard contrastive loss that pulls matched expression–CNV pairs (from the same cell) together and pushes unmatched pairs apart.
2. Hard Negative Triplet Loss:
   • Definition: For an anchor cell $i$, a "hard negative" $j$ is a cell with similar CNV profiles (high cosine similarity) but divergent expression profiles (low cosine similarity).
   • Biological Rationale: These pairs represent cells where similar genomic alterations fail to produce similar expression, suggesting regulatory escape or compensation.
   • Objective: The loss enforces that the anchor's expression embedding is closer to its own CNV embedding than to the hard negative's CNV embedding (and vice versa).
   • Formula: $L = L_{InfoNCE} + \lambda L_{HN}$, where $L_{HN}$ is a margin-based triplet loss.

D. Cell Classification

After training, a concordance score is calculated for each cell based on the cosine similarity between its expression and CNV embeddings in the latent space.

• Concordant Cells: High similarity (expression follows CNV).
• Discordant Cells: Low similarity (expression escapes CNV expectations).

3. Key Results

A. Cross-Modal Retrieval Performance

The model successfully aligned the two modalities:

• Top-1 Retrieval Accuracy: ~87.3% (Expression $\to$ CNV) and 87.1% (CNV $\to$ Expression).
• Mean Reciprocal Rank (MRR): ~0.896.
• Alignment Gap: A positive gap of ~0.108 indicates clear separation between matched pairs and the hardest negatives.

B. Identification of Escape and Compensation Genes

By performing differential expression analysis between discordant and concordant cells, the authors identified two distinct gene categories:

1. Escape Genes (Upregulated in Discordant Cells):

   • These genes are upregulated despite CNV status, suggesting active regulatory mechanisms overriding dosage.
   • Top Hits: VSIG4, FCGR1A, TREM2, MARCO, SPI1, C1QC.
   • Biological Insight: These are predominantly myeloid/macrophage markers and immune regulators. This suggests that discordant cells may represent tumor-associated macrophages or cancer cells adopting an immune-evasion phenotype.

2. Compensation Genes (Downregulated in Discordant Cells):

   • These genes are downregulated despite potential CNV-driven upregulation.
   • Top Hits: MALAT1 (lncRNA), CD8A, CCL5, GZMA, GZMH.
   • Biological Insight: These are cytotoxic T-cell and NK-cell markers. Their downregulation in discordant cells suggests a suppression of cytotoxic activity in regions of genomic instability, pointing to an immunosuppressive microenvironment.

C. Validation

• Recurrent Patterns: Key genes (e.g., XBP1, CTSB) were consistently identified across multiple patients.
• Statistical Robustness: Results held up in both pooled analysis (40,775 cells) and patient-stratified paired analysis, confirming the findings are not driven by batch effects.

4. Key Contributions

1. Novel Framework (CLCC): A contrastive learning approach specifically designed to align scRNA-seq expression with inferred CNV, moving beyond simple correlation to learn a shared latent space.
2. Hard Negative Mining Strategy: The explicit training on cell pairs with similar CNVs but divergent expression allows the model to learn the specific geometry of "regulatory escape," a task standard contrastive learning might miss.
3. Discovery of Dosage-Insensitive Programs: The study provides a systematic pipeline to identify genes that decouple from genomic dosage, revealing a new axis of tumor heterogeneity.
4. Biological Insights: The identification of myeloid enrichment in escape cells and T-cell suppression in discordant regions offers new hypotheses regarding how tumors use transcriptional adaptation to evade immune surveillance.

5. Significance and Limitations

• Significance: CLCC provides a generalizable tool for discovering regulatory mechanisms in cancer using standard scRNA-seq data. It shifts the focus from "what genes are mutated/amplified" to "how do cells transcriptionally adapt to these changes," potentially identifying new therapeutic targets (e.g., targeting the escape mechanisms of macrophage-like tumor cells).
• Limitations:
  • Inferred CNVs: CNV profiles were derived from expression data, which introduces a potential "shared-signal bias" (the model might learn artifacts of the inference method rather than true biology).
  • Causality: The framework identifies associations and prioritizes candidates but does not prove causal regulatory mechanisms.
  • Future Work: The authors plan to validate findings on datasets with orthogonal CNV measurements (e.g., scDNA-seq) and perform ablation studies on the hard-negative mining component.

In summary, this paper presents a robust machine learning framework that successfully decouples genomic dosage from transcriptional output in cancer, revealing that cells with "discordant" states often harbor specific immune-related programs that may drive tumor progression and immune evasion.