🧬 biology

Quantum AI for Cancer Diagnostic Biomarker Discovery

This study demonstrates that a two-phase quantum machine learning framework effectively identifies subtype-specific biomarkers and achieves high-precision classification for lung adenocarcinoma and squamous cell carcinoma by leveraging quantum advantages in processing multiomic data and revealing key neurotrophin-mediated oncogenic pathways.

Original authors: Mandeep Kaur Saggi, Amandeep Singh Bhatia, Humaira Gowher, Sabre Kais

Published 2026-04-22

📖 5 min read🧠 Deep dive

CC BY 4.0

Original authors: Mandeep Kaur Saggi, Amandeep Singh Bhatia, Humaira Gowher, Sabre Kais

Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). ⚕️ This is an AI-generated explanation of a preprint that has not been peer-reviewed. It is not medical advice. Do not make health decisions based on this content. Read full disclaimer

Imagine you are a detective trying to solve a very tricky case: Lung Cancer.

The problem is that there are two main suspects who look almost identical on the surface: Lung Adenocarcinoma (LUAD) and Lung Squamous Cell Carcinoma (LUSC). They are both types of non-small cell lung cancer, but they behave differently and need different treatments. If you treat them like the same criminal, the patient might not get the right cure.

For a long time, scientists have tried to tell them apart by looking at their "DNA fingerprints" (genes) and "chemical tags" (methylation). But there are so many fingerprints and tags—millions of them—that it's like trying to find a specific needle in a haystack the size of a mountain using a regular flashlight.

This paper introduces a new, super-powered flashlight: Quantum Machine Learning (QML).

Here is a simple breakdown of how the researchers solved the case:

1. The Crime Scene (The Data)

The researchers gathered evidence from a massive database called TCGA (The Cancer Genome Atlas). They looked at two types of clues for hundreds of patients:

The "Volume" Clues (RNA): How loud are the genes shouting? (Are they working hard or sleeping?)
The "Lock" Clues (DNA Methylation): Are the genes locked down with chemical tags? (If a gene is "hyper-methylated," it's locked tight and silenced. If it's "hypo-methylated," the lock is broken and the gene is running wild.)

2. The Investigation (Phase 1 & 2)

First, the team used old-school detective work (classical math) to filter the noise. They asked:

"Which genes are screaming in LUAD but whispering in normal lungs?"
"Which genes are screaming in LUSC but whispering in normal lungs?"
"Which genes scream differently in LUAD compared to LUSC?"

They found three special groups of clues:

Sample 1: Genes that were locked down (methylated) and silent (downregulated). Think of these as "brakes" that were accidentally slammed on.
Sample 2: Genes that were unlocked (unmethylated) and screaming (upregulated). Think of these as "gas pedals" that were stuck to the floor.
Sample 3: A super-combo of both the brakes and the gas pedals.

3. The Quantum Leap (The New Tool)

This is where the magic happens. Instead of using a normal computer to sort through these clues, they built a Quantum Neural Network (QNN).

The Analogy:
Imagine you have a giant library of books (the genes).

A Classical Computer is like a librarian who reads one book, checks the index, puts it back, and moves to the next. It's fast, but it has to do it one by one.
A Quantum Computer is like a librarian who can read every book in the library at the exact same time, while also seeing how the books relate to each other in a 3D web of connections. It uses "superposition" (being in many states at once) to find patterns that a normal computer might miss.

They fed their three samples (Sample 1, 2, and 3) into this Quantum AI. They tested it with different sizes of "clue sets" (from 64 clues up to 256 clues) to see how well the AI could distinguish between the two cancer types.

4. The Verdict (The Results)

The Quantum AI was incredibly accurate.

The Winner: Sample 3 (the combo of locked-down/silent genes AND unlocked/loud genes) was the best detective. It gave the AI the most complete picture.
The Score: The AI achieved 96% to 99% accuracy. It was almost perfect at telling LUAD from LUSC.
The Efficiency: Even though the Quantum model was looking at huge amounts of data, it needed far fewer "brain cells" (parameters) to learn than a standard computer model. It was like solving a complex puzzle with fewer moves.

5. The "Why" (The Biological Story)

When the researchers looked at which genes the AI cared about most, they found something fascinating. The top clues weren't just random; they were heavily involved in neuroscience and brain signaling.

The Metaphor:
You might think, "Wait, this is lung cancer, why are we talking about brains?"
The researchers found that the cancer cells were hijacking neurotransmitter pathways (the chemical messengers used by neurons). It's as if the lung cancer cells started acting like confused neurons, using brain signals to grow and spread.

Genes like NTRK2 and NGFR (which are usually involved in nerve growth) were acting up.
Pathways like MAPK and PI3K-Akt (the cell's internal "growth engines") were revving too high.

This suggests that these cancers might be more "neuro-like" than we thought, opening up new doors for treatment.

The Big Takeaway

This paper is a proof-of-concept. It says: "Hey, we can use the weird, powerful laws of quantum physics to help doctors diagnose cancer more accurately and faster."

While we don't have quantum computers in every hospital yet, this study shows that when we combine Quantum AI with multi-layered biological data (genes + chemical tags), we can find the "smoking gun" biomarkers that separate different types of cancer. This could lead to personalized treatments where a patient gets the exact drug their specific cancer type needs, rather than a one-size-fits-all approach.

In short: They used a quantum super-computer to find the perfect mix of genetic clues that tell two look-alike lung cancers apart, discovering that these cancers might be secretly using "brain signals" to grow.

1. Problem Statement

Lung cancer, specifically Non-Small Cell Lung Cancer (NSCLC), is the leading cause of cancer-related mortality. The two predominant subtypes, Lung Adenocarcinoma (LUAD) and Lung Squamous Cell Carcinoma (LUSC), share overlapping clinical features, making accurate subtype classification difficult yet critical for targeted therapy and patient stratification.

While classical machine learning (ML) has been applied to multi-omics data (DNA methylation and RNA-seq) for biomarker discovery, it faces significant challenges:

High Dimensionality: Omics datasets contain tens of thousands of features (genes/probes) but relatively few samples, increasing the risk of overfitting.
Data Complexity: Integrating heterogeneous data types (epigenomics and transcriptomics) with different scales and structures is computationally demanding.
Interpretability: Deep learning models often act as "black boxes," lacking biological interpretability regarding the identified features.

The authors propose a Quantum Machine Learning (QML) approach to overcome these limitations, leveraging quantum principles (superposition, entanglement) to process high-dimensional multi-omic data more efficiently and accurately.

2. Methodology: The QUBID Framework

The study introduces QUBID (Quantum-based biomarker Identification and Diagnostic), a hybrid classical-quantum framework consisting of five phases:

Phase 1 & 2: Differential Analysis & Feature Selection (Classical)

Data Source: TCGA (The Cancer Genome Atlas) datasets comprising DNA methylation (485,578 CpG probes) and RNA-seq (20,531 genes) for LUAD and LUSC.
Comparisons:
- Tumor vs. Normal: Identified differentially expressed genes (DEGs) and differentially methylated regions (DMRs) for LUAD and LUSC separately.
- Tumor vs. Tumor: Compared LUAD tumors directly against LUSC tumors to find subtype-specific markers.
Filtering: Applied statistical thresholds (adjusted p-value $\le$ 0.05, Log Fold Change $> 0.5$ or $< -0.5$ , and methylation delta $\beta$ thresholds) to isolate significant features.

Phase 3: Multi-Omic Integration & Categorization

The authors categorized genes based on the correlation between methylation and expression to create three distinct feature sets (Samples):

Sample 1: Genes that are Hypermethylated (silenced) and Downregulated.
Sample 2: Genes that are Hypomethylated (activated) and Upregulated.
Sample 3: A combined set of Sample 1 and Sample 2, representing a comprehensive view of epigenetic and transcriptional dysregulation.

Phase 4: Quantum Neural Network (QNN) Classification

Model Architecture: A Variational Quantum Circuit (VQC) embedded within a classical neural network.
- Input Encoding: Amplitude encoding maps classical feature vectors into quantum states using $n = \lceil \log_2(d) \rceil$ qubits.
- Circuit Structure: 6–8 qubits with 6 layers of parameterized rotations ( $R_y, R_x$ ) and all-to-all CNOT entanglement gates.
- Measurement: Expectation value of the Pauli-Z operator on the final qubit, passed through a sigmoid function for binary classification.
Training: Optimized using the Adam optimizer with varying learning rates, batch sizes (32, 64), and epochs (32–50).
Feature Importance: A hybrid quantum-gradient method was used to rank features by combining quantum weight contributions and classical gradient sensitivity.

Phase 5: Functional Enrichment

Top-ranked biomarkers were analyzed using Gene Ontology (GO) and KEGG pathway enrichment to determine biological relevance.

3. Key Contributions

Novel Hybrid Framework: First application of a hybrid classical-quantum framework specifically for distinguishing LUAD and LUSC subtypes using integrated DNA methylation and RNA-seq data.
Multi-Omic Integration Strategy: Demonstrated that combining epigenetic (methylation) and transcriptomic (expression) data yields superior classification performance compared to single-omics approaches.
Quantum Advantage in Parameter Efficiency: Showed that QNN models achieve high accuracy with significantly fewer trainable parameters compared to classical Deep Neural Networks (DNNs) and Support Vector Machines (SVMs).
Biological Discovery: Identified specific neurotrophin-mediated oncogenic pathways (e.g., NTRK2, NGFR, NTF3) as central to subtype differentiation, validated by both QML and classical enrichment analysis.

4. Results & Performance

The study evaluated performance across three sample sets and varying feature dimensions (64, 128, 256 features).

Best Performing Sample: Sample 3 (Combined Hypermethylated/Downregulated + Hypomethylated/Upregulated) consistently outperformed Sample 1 and Sample 2.
- Test Accuracy: ~96.97% (at 64 features) to 96.36% (at 256 features).
- AUC (Area Under Curve): Reached 0.9960 with 256 features, indicating near-perfect class separation.
- F1-Score: Consistently above 0.96.
Comparison with Classical Models:
- QNN vs. Classical: The QNN achieved comparable or superior accuracy and AUC to Random Forest (RF), SVM, and Neural Networks (NN) while using orders of magnitude fewer parameters.
  - Example: For 256 features, QNN used only 96 parameters, whereas a classical NN required 263,169 parameters.
- Stability: QNN showed robust performance across different feature sizes, whereas SVM performance degraded significantly at higher dimensions.
Biological Insights:
- KEGG Pathways: Enrichment in Neurotrophin signaling, MAPK, Ras, and PI3K-Akt pathways.
- Key Biomarkers: NTRK2, NGFR, NTF3, CACNA2D3, and ADAM23.
- GO Terms: Significant enrichment in synaptic signaling, ion channel regulation, and axon guidance, suggesting a strong link between neuronal development pathways and lung cancer subtypes.

5. Significance & Conclusion

This study demonstrates that Quantum Machine Learning is a viable and powerful tool for precision oncology.

Scalability: The QNN framework effectively handles high-dimensional multi-omic data without the computational explosion typical of classical deep learning.
Clinical Relevance: The identified biomarkers (e.g., NTRK2) offer potential targets for diagnostic stratification and therapeutic intervention in NSCLC.
Future Outlook: While current quantum hardware is in the NISQ (Noisy Intermediate-Scale Quantum) era, this work proves that hybrid quantum-classical algorithms can provide meaningful insights and superior feature selection capabilities today. It paves the way for next-generation biomedical analytics that leverage quantum computing for complex disease classification.

In summary, the paper establishes that integrating multi-omic data via a quantum classifier not only improves the accuracy of lung cancer subtype diagnosis but also reveals biologically significant pathways that might be missed by classical methods, all while maintaining computational efficiency.