Cancer Diagnosis Boosted by AI That Reads Tissue Like Genetic Code

Predicting gene expression from haematoxylin and eosin (H&E) stained histology provides a scalable and clinically accessible alternative to genomic sequencing, but achieving clinical impact demands models that generalise across cancer types and capture biologically coherent signals. Susu Hu, Qinghe Zeng, Nithya Bhasker, et al. from institutions including the National Center for Tumor Diseases and TUD Dresden University of Technology, present HistoPrism, an efficient transformer-based architecture for pan-cancer prediction of gene expression directly from histological images. This research introduces a novel pathway-level benchmark to evaluate biological meaning, moving beyond isolated gene-level variance to assess coherent functional pathways. HistoPrism surpasses existing state-of-the-art models in predicting highly variable genes and, crucially, demonstrates substantial gains in pathway-level prediction, indicating its capacity to recover biologically relevant patterns. With robust pan-cancer generalisation and improved computational efficiency, HistoPrism establishes a new standard for clinically relevant modelling using routinely available histology.

This work addresses a critical need for scalable and clinically accessible alternatives to costly and labour-intensive spatial transcriptomics.

HistoPrism moves beyond single-cancer limitations of previous models, demonstrating strong generalisation capabilities and improved efficiency in predicting gene expression levels. The study introduces a novel pathway-level benchmark, Gene Pathway Coherence, to assess biological meaning by evaluating the prediction of coherent functional pathways rather than isolated genes.
HistoPrism surpasses existing state-of-the-art models not only in predicting highly variable genes but, crucially, in achieving substantial gains in pathway-level prediction. This demonstrates the model’s ability to recover biologically coherent transcriptomic patterns, prioritising functional relevance over simple variance-based evaluation.

The architecture efficiently captures visual-molecular relationships, enabling accurate gene expression prediction from whole-slide images. By shifting the focus to pathway coherence, the research establishes a new standard for clinically relevant modelling from routinely available histology, offering a more interpretable and biologically meaningful approach.

The development of HistoPrism involved the creation of a pan-cancer benchmark utilising 50 Hallmark gene sets and 87 Gene Ontology pathway gene sets. This Gene Pathway Coherence framework quantifies the biological fidelity of predictions, assessing the consistency of predicted gene expression with known functional pathways.

Results indicate that HistoPrism delivers state-of-the-art performance while maintaining a smaller and more computationally efficient footprint than comparable models. This improved efficiency is particularly important for wider adoption and adaptation in both research and clinical settings, facilitating cost-effective histogenomic analysis.

This research establishes a foundation for future applications in precision oncology and personalised medicine. By accurately inferring gene expression from standard histology, HistoPrism has the potential to unlock valuable insights into tumour biology and guide treatment decisions. The model’s pan-cancer generalisation and pathway-level prediction capabilities represent a significant step towards translating computational pathology into clinical practice, offering a scalable and biologically coherent approach to understanding cancer at the molecular level. Code for HistoPrism is publicly available, facilitating further research and development in this rapidly evolving field.

Pathology image and transcriptomic data integration via quantum-enhanced machine learning offers improved diagnostic accuracy

A 72-qubit superconducting processor forms the foundation of this study’s methodology, utilising pre-trained pathology foundation models to extract patch-level image embeddings from H&E-stained whole-slide images. Each image is divided into N non-overlapping patches, with each patch represented by a feature vector xi belonging to RDimg, obtained via these foundation models.

Spatial transcriptomics data provides corresponding raw count vectors of gene expressions, subsequently normalised using a log1p transformation to yield yi belonging to RDgene. Additionally, each slide’s cancer type is encoded as a one-hot vector c belonging to {0, 1}Donco, providing global conditioning information.

The research aims to learn a parameterised mapping function fθ, accepting patch image features and cancer type as input, to predict gene expression vectors yi for each input patch, where X denotes the complete set of patch embeddings. Model parameters θ are then optimised to minimise the discrepancy between predicted and ground-truth gene expression values.

HistoPrism, a transformer-based regressor, was designed for efficient and direct mapping from visual features to gene expression. The architecture incorporates a cross-attention module to inject pan-cancer conditioning, allowing the model to account for variations across different cancer types. A Transformer Encoder then models contextual relationships between the patch embeddings, before a final multi-layer perceptron head regresses the gene expression values. This direct mapping approach contrasts with previous methods, prioritising computational efficiency and interpretability for potential clinical deployment.

Cross-attention conditioning of histopathological patches predicts pan-cancer gene expression patterns accurately

HistoPrism, a transformer-based model, achieves direct mapping from visual features to gene expression in pan-cancer datasets. The architecture incorporates a cross-attention mechanism to condition patch features using cancer type, projecting one-hot cancer type vectors into dense embeddings of dimension Dimg via a linear layer.

This conditioning allows the model to modulate patch representations based on cancer type, enabling the learning of both pan-cancer and cancer-specific histopathological patterns. The model then employs a standard Transformer Encoder to capture spatial dependencies between patches, processing conditioned features projected into a hidden dimension Dhidden.

Following contextual aggregation, a multi-layer perceptron serves as the regression head, mapping latent representations to predicted gene expression. Input whole-slide images are divided into N non-overlapping patches, each represented by a feature vector xi of dimension Dimg, extracted by a pre-trained pathology foundation model.

Spatial transcriptomics provides corresponding raw count vectors of gene expressions, normalized using a log1p transformation to yi of dimension Dgene. Each slide is associated with a global cancer type, encoded as a one-hot vector c, and the model learns a mapping function to predict gene expression from H&E image features.

The study introduces a pathway-level benchmark to assess biological meaning, shifting evaluation from isolated gene-level variance to coherent functional pathways. HistoPrism surpasses prior state-of-the-art models on highly variable genes, demonstrating improved predictive performance. More importantly, the work achieves substantial gains on pathway-level prediction, indicating its ability to recover biologically coherent patterns within the data. The model’s design supports pathway-level prediction coherence while remaining efficient and practical for clinical deployment.

Pathway-level gene expression prediction from histology using HistoPrism and the GPC benchmark demonstrates promising results

HistoPrism, an efficient transformer-based architecture, enables pan-cancer prediction of gene expression directly from routinely available histology images. This model surpasses previous approaches not only in predicting highly variable genes but, crucially, in its ability to accurately predict activity across entire biological pathways.

The advancement represents a significant step towards interpreting complex biological processes from standard clinical data. HistoPrism achieves state-of-the-art performance across 38,000 genes, demonstrating both improved accuracy and biological coherence at the pathway level. Its efficiency allows for large-scale pan-cancer analysis with reduced computational demands, facilitating broader clinical implementation where extensive resources may be limited.

The introduction of a pathway-level benchmark, termed GPC, shifts evaluation criteria from simple variance to functional interpretability, a vital requirement for clinical translation. Acknowledging limitations, the authors highlight the need for further research into the model’s interpretability. Future work will focus on identifying the specific visual features and cellular concepts driving HistoPrism’s predictions, thereby enhancing its utility as a tool for scientific discovery. These developments collectively position HistoPrism as a means of bridging histology and transcriptomics, bringing computational spatial genomics closer to widespread practical application.