Glioblastoma is an aggressive form of brain cancer known for its biological complexity. Even tumors that receive the same diagnosis can contain very different cells and molecular characteristics, which can influence how the disease progresses and how patients respond to treatment.

Researchers from Adelaide University have developed a computational approach that combines information from tumor tissue images with RNA sequencing data to look for previously unrecognized groups of glioblastoma patients.

Combining two views of glioblastoma

Pathologists routinely examine tumor samples stained with hematoxylin and eosin, or H&E, under a microscope. These images reveal features such as the appearance and organization of tumor cells and surrounding tissue.

RNA sequencing provides a different type of information. Instead of showing what a tumor looks like, RNA-seq measures gene activity, providing a molecular profile of what is happening inside the tumor cells.

Each approach captures different aspects of tumor biology. The researchers investigated whether combining them could reveal clinically meaningful patterns that might be missed when either type of information is analyzed alone.

Multimodal workflow integrating histology and RNA-seq for patient stratification

(a) Whole-slide image (WSI) acquisition. (b) WSI tiling into non-overlapping patches (20×). (c) Tissue-containing tiles retained after quality control. (d) Patch-level CNN feature extraction with a pretrained ResNet-50 (2048-D) and per-patient aggregation. (e) Patient-level histology feature vector (2048-D). (f) Autoencoder (AE) compression of histology features to a 30-D latent embedding. (g) RNA-seq preprocessing with low-variance gene cut-off and normalisation. (h) Normalised gene-expression matrix across patients. (i) AE compression of RNA-seq (~48 k genes) to a 30-D latent embedding. (j) Late fusion of the two feature representations obtained from histology (30-D) and RNA (30-D). (k) Final 60-D multimodal feature vector per patient. (l) Clustering followed by survival analyses (Kaplan Meier and Cox) identify prognostic subgroups. (m) Discovery of high-importance genes from RNA representations (30-D) contributing to cluster separation and survival differences.

Using deep learning to analyze tumor data

The researchers developed a workflow that processed whole-slide pathology images and RNA sequencing data separately before bringing the information together.

For the pathology images, tumor slides were divided into smaller sections called tiles. A pretrained deep learning model extracted thousands of image features describing characteristics of the tissue. An autoencoder, a type of neural network used to compress complex data, then reduced these features into a smaller representation for each patient.

A similar process was applied to the RNA sequencing data. Starting with expression measurements for approximately 48,000 genes, the researchers filtered and normalized the data and used another autoencoder to produce a condensed molecular representation of each tumor.

The image and RNA sequencing features were then combined to create a single profile representing both the tumor’s physical appearance and its gene expression.

Identifying glioblastoma subgroups

The researchers used unsupervised machine learning to search these combined profiles for patterns. Unlike methods that are trained to recognize predefined categories, unsupervised learning attempts to discover groups within data without being told in advance what those groups should look like.

Several clustering methods were evaluated. The strongest approach separated the patients into two groups, one containing 150 patients and a much smaller group containing eight patients.

Importantly, the groups differed substantially in survival. Among patients with survival information available, median survival was 454 days in one group compared with 138 days in the other.

The poorer-prognosis group also remained associated with increased mortality risk after researchers adjusted their analysis for patient age.

Connecting patient groups with tumor biology

The researchers then examined which genes contributed most strongly to the differences between the two groups.

Their analysis produced relatively small sets of genes associated with each cluster, including genes connected with NOTCH and gamma-secretase signaling and oxidative phosphorylation. These biological processes have roles in cell signaling, metabolism, and cancer development.

Finding interpretable gene signatures is important because it can help researchers move beyond simply identifying statistical groups and begin investigating the biology that may explain why those groups behave differently.

A multimodal approach to glioblastoma research

The results demonstrate the potential value of combining RNA sequencing with information already available from routine pathology images. Instead of treating tumor appearance and gene expression as separate sources of information, multimodal analysis can examine how the two relate to one another.

The researchers emphasize that the identified patient groups are not ready to guide treatment decisions. The smaller cluster included only eight patients, and the findings will require validation using larger, independent patient populations from multiple medical centers.

Still, the approach provides a framework for investigating glioblastoma heterogeneity. Combining RNA sequencing, digital pathology, and machine learning could help researchers identify biologically distinct tumor subgroups, investigate the mechanisms behind differences in patient outcomes, and generate new hypotheses for future glioblastoma research.