⚛️ quantum physics

Hybrid quantum-classical attention for histopathology-based molecular profiling in data-limited cancers

This paper presents a hybrid quantum-classical strategy that replaces standard softmax attention with a quantum-derived doubly stochastic matrix in histopathology transformers, demonstrating selective improvements in gene expression prediction for data-limited cancers and biologically relevant targets while validating the underlying mechanism on IBM quantum processors.

Original authors: Kahn Rhrissorrakrai, Aritra Bose, Aldo Guzman-Saenz, Filippo Utro, Laxmi Pardia

Published 2026-09-21

📖 5 min read🧠 Deep dive

CC BY 4.0

Original authors: Kahn Rhrissorrakrai, Aritra Bose, Aldo Guzman-Saenz, Filippo Utro, Laxmi Pardia

Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

In the fight against cancer, doctors often rely on two distinct types of information to understand a tumor. One comes from looking at the tissue under a microscope, a routine process where a pathologist examines a thin slice of a biopsy to see how the cells are arranged and what they look like. The other comes from reading the chemical instructions inside those cells, known as genetic or molecular profiling, which reveals which genes are active and how the cancer might behave. While the microscopic view is standard in hospitals worldwide, the genetic reading requires expensive equipment and a large amount of tissue, making it difficult to perform for small biopsies, rare cancers, or patients in places where advanced labs are scarce. This gap leaves many patients without a complete picture of their disease. Scientists have been trying to bridge this divide by teaching computers to look at routine microscope images and predict the genetic activity hidden within, hoping to extract molecular insights from the visual patterns of the tissue alone.

A team of researchers at IBM Research has taken a new approach to this challenge by introducing a tool from the emerging field of quantum computing into the software that analyzes these medical images. They developed a hybrid system that combines standard computer processing with a specific type of quantum calculation designed to help the software pay attention to the right parts of an image. In their study, they tested this system on digital images of tissue samples from 29 different types of cancer, ranging from common forms like breast and lung cancer to rare and difficult-to-study varieties. The goal was to see if this quantum-inspired method could help the computer make more accurate guesses about the genetic makeup of a tumor when the actual genetic data is missing or when there are very few patient samples to learn from.

The researchers found that the hybrid system did not simply make the computer better at guessing every single gene. Instead, it acted more like a selective filter, improving the accuracy for some specific genes while making the predictions for others slightly worse. This trade-off was not random; the genes that became easier to predict often belonged to biological groups that are known to be important for that specific type of cancer. For instance, in a rare and aggressive cancer of the adrenal gland called adrenocortical carcinoma, the hybrid system significantly improved the prediction of genes involved in cell division and DNA repair. Crucially, the genes that saw the biggest improvement were also the ones linked to a poorer outlook for patients. This suggests that the new method was successfully picking up on the visual signs of the most dangerous aspects of the disease, even in a setting where data is scarce.

The study also explored whether this method could work across different groups of patients. When the researchers trained the system on one large collection of pancreatic cancer data and then tested it on a completely separate, independent collection, the results were mixed. The system improved predictions for a specific set of genes related to metabolism and the cell's lineage, but it struggled with many others. This indicates that the technology is not a universal fix that works perfectly for every gene in every situation. Instead, it appears to be a specialized tool that can be tuned to focus on particular biological pathways. The researchers also tested the quantum part of their system on actual quantum processors, the specialized hardware designed for these calculations. They confirmed that the core mathematical operation used by the system could be successfully performed on these machines, even though the full image analysis was still done on standard computers.

One of the most significant findings was that the hybrid system seemed to work best when there was very little data to begin with. In the smaller groups of patients, where traditional computer models often struggle to learn, the quantum-inspired attention mechanism provided the most noticeable gains. This is a promising sign for rare cancers, where gathering large numbers of samples is nearly impossible. The researchers also discovered that they did not need to use the quantum calculation for the entire training process to get benefits. Using it only for the first few steps of the learning process was enough to guide the system toward better performance, which could make the technology more practical to use in the future.

Ultimately, this work suggests a new way to think about using advanced computing in medicine. Rather than trying to replace standard methods entirely, the hybrid approach offers a way to enhance specific parts of the analysis. It shows that by changing how a computer decides which parts of an image to focus on, using a structure derived from quantum physics, it is possible to uncover hidden molecular signals in routine tissue slides. While the technology is still in the early stages and requires further testing to be used in hospitals, it offers a potential path forward for giving patients with rare or data-limited cancers a more complete understanding of their disease, using the images that are already available in their medical records.

Technical Summary: Hybrid Quantum-Classical Attention for Histopathology-Based Molecular Profiling

Problem Statement

Molecular profiling is critical for cancer diagnosis, prognosis, and treatment selection, yet access to transcriptomic assays (e.g., RNA-seq) is uneven due to cost, tissue limitations, and the difficulty of obtaining samples from small biopsies, rare cancers, or retrospective cohorts. While routine hematoxylin and eosin (H&E) whole-slide images (WSIs) are widely available, predicting gene expression from these images remains challenging. Existing deep learning models, particularly those based on transformer architectures like SEQUOIA, struggle with generalizability, target specificity, and performance in data-limited settings. Standard attention mechanisms (softmax) may lead to imbalanced token interactions or "token collapse," potentially limiting the model's ability to extract molecular signals from complex histopathology data, especially when training cohorts are small.

Methodology

The authors propose a hybrid quantum-classical attention strategy integrated into the SEQUOIA framework, a transformer-based model for predicting gene expression from histopathology images.

Core Architecture: QDSFormer in SEQUOIA

Integration: The study replaces the standard row-wise softmax normalization in the SEQUOIA transformer's attention module with a Quantum-Derived Doubly Stochastic Matrix (QDSM).
Mechanism:
- Standard Attention: Computes similarity scores ( $S = QK^T/\sqrt{d}$ ) and applies row-wise softmax, resulting in a right-stochastic matrix (rows sum to 1, columns do not).
- QDSM Attention: Replaces softmax with a quantum-derived doubly stochastic matrix where both rows and columns sum to 1. This is generated using a quantum conditional optimal transport procedure (QDSFormer).
- Constraint: The doubly stochastic constraint imposes a more balanced assignment structure, theoretically stabilizing attention and reducing token collapse.
Implementation Details:
- Training: The QDSM attention was evaluated via classical simulation of the corresponding quantum circuit during model training. The rest of the pipeline (ResNet-50 feature extraction, MLP regression, optimization) remained classical.
- Matrix Size: For computational tractability in the 29 TCGA cancer cohorts, the attention matrix was restricted to 8 × 8 (generated by a 14-qubit circuit).
- Training Strategies: Two strategies were compared:
  1. Full-QDSM: QDSM attention applied for all 200 training epochs.
  2. Warm-Start: QDSM attention applied only for the initial epochs (e.g., 5) before reverting to standard softmax.

Experimental Design

Datasets:
- Primary: 29 cancer types from The Cancer Genome Atlas (TCGA) with matched WSIs and RNA-seq data (n ≥ 50 samples per type).
- External Validation: An independent pancreatic cancer cohort from the Clinical Proteomic Tumor Analysis Consortium (CPTAC).
- Hardware Validation: Separate experiments on IBM quantum processors (Eagle R3 and Heron R1) to recover doubly stochastic matrices of sizes 8×8, 64×64, and 128×128.
Evaluation Metrics:
- Gene-level prediction performance measured by Pearson correlation between predicted and measured expression.
- Statistical significance determined via two-sided t-tests across 10 train-test splits.
- Gene-level Benefit Modeling: A mixed-effects regression model (using leave-one-cancer-out cross-validation) to predict the improvement ( $\Delta r$ ) based on baseline gene features (expression abundance, variance, co-expression degree).
- Survival Analysis: Cox proportional hazards models applied to genes improved by QDSM in Adrenocortical Carcinoma (ACC).

Key Results

1. Selective Gains in Data-Limited Cancers

The hybrid attention mechanism did not uniformly improve prediction across the entire transcriptome. Instead, it redistributed predictive performance:

Net Improvement: In 16 of 29 cancer types, more genes showed significant improvement than worsening.
Data-Limited Advantage: The largest relative improvements were observed in smaller, data-limited cohorts. A negative correlation was found between cohort size and the ratio of improved-to-worsened genes ( $r = -0.35$ for genes; $r = -0.42$ for enriched gene sets).
Top Performers: Adrenocortical carcinoma (ACC), Uveal Melanoma (UVM), Pheochromocytoma (PCPG), and Pancreatic Adenocarcinoma (PAAD) showed the strongest net benefits.

2. Biological Relevance in Adrenocortical Carcinoma (ACC)

ACC served as a detailed case study for a rare, molecularly heterogeneous cancer:

Gene-Level Improvement: 1,020 genes showed significantly improved prediction under QDSM, compared to only 130 worsened genes.
Pathway Enrichment: Improved genes were enriched for biologically coherent processes relevant to ACC, including DNA damage/repair, transcriptional/epigenetic regulation, and protein ubiquitination.
Prognostic Value: The genes preferentially improved by QDSM were significantly associated with worse overall survival (Hazard Ratios > 1). Top genes included LYG1, KIF18A, and HDAC4, linking the model's technical gain to clinically meaningful prognostic information.

3. Cross-Cohort Transferability (Pancreatic Cancer)

In the transfer experiment from TCGA-PAAD to CPTAC:

Selective Transfer: QDSM improved prediction for approximately 2,000 genes but worsened prediction for approximately 6,200.
Target Specificity: Improved genes included biologically relevant targets such as tumor suppressors (ARID1A, STK11), metabolic regulators (G6PD, ALDOB), and lineage-associated transcription factors (GATA6, FOXA2).
Limitation: The model did not provide uniform improvement under domain shift, reinforcing that the benefit is target-specific rather than a general transcriptome-wide enhancement.

4. Predictability of Benefits

Gene-Level Modeling: A mixed-effects model using baseline gene features (predictivity, abundance, variance, co-expression) explained a meaningful fraction of the variation in QDSM benefit (LOCO correlation $\approx 0.52$ ).
Residual Analysis: Genes that improved more than expected were enriched for developmental and morphogenetic pathways. Genes that improved less than expected were enriched for ribosomal metabolism, DNA replication, and adaptive immune regulation. This suggests the benefit is partly predictable but also context-dependent.

5. Hardware Feasibility

Separate experiments on IBM quantum processors successfully recovered doubly stochastic matrices (8×8, 64×64, 128×128) with Spearman correlations $\ge 0.7$ against exact classical solutions within 100,000 shots.
This validates that the quantum primitive underlying the attention mechanism is implementable on current hardware, even if end-to-end training on QPUs is not yet feasible.

Significance and Claims

The paper positions hybrid quantum-classical attention not as a universal replacement for standard softmax attention, but as a targeted, context-dependent modeling strategy for digital pathology.

Molecular Triage: The approach is most valuable for "molecular triage" in settings where direct sequencing is unavailable, incomplete, or impractical (e.g., rare cancers, small biopsies). It offers a way to recover specific, biologically relevant gene programs from routine histology.
Data Efficiency: The results suggest that the doubly stochastic constraint acts as a useful inductive bias in data-limited settings, regularizing token interactions to improve the inference of specific molecular targets that are morphologically encoded.
Clinical Linkage: The study demonstrates that technical improvements in prediction accuracy can be directly linked to clinically relevant outcomes (e.g., survival in ACC), validating the utility of the method for precision oncology applications where data is scarce.
Modular Hybrid Workflow: The work supports a modular architecture where classical systems handle feature extraction and regression, while quantum subroutines (or their classical simulations) provide structured attention normalization.

The authors conclude that while the method does not uniformly boost transcriptome-wide accuracy, its ability to selectively improve the prediction of prognostically significant genes in data-limited cancers makes it a promising strategy for image-based molecular profiling.