by Christos Evangelou, MSc, PhD – Medical Writer and Editor

In a new study, researchers at the Université de Moncton in Canada developed a hybrid deep learning architecture called EMGP-Net that can predict gene expression patterns directly from routine histopathological images, potentially eliminating the need for expensive spatial transcriptomics technologies. EMGP-Net combines two cutting-edge AI models to achieve high accuracy in gene expression prediction, offering new avenues for personalized cancer treatment in resource-limited settings. The study was published in Computers.

Addressing the Cost Barrier of Spatial Transcriptomics

Spatial transcriptomics tools enable high-resolution analysis of gene expression across tissue sections while preserving spatial context. However, their high cost has limited their clinical adoption, especially in community hospitals and resource-limited settings.

Dr. Moulay Akhloufi, corresponding author of the study, explained that with this research, they wanted to address the unmet need of making molecular-level tumor analysis accessible to clinicians worldwide without requiring specialized, expensive equipment.

“With the recent advancements in deep learning, such as transformer models and, more recently, SSM-based architectures like MambaVision, we now have the opportunity to apply these innovations to critical domains like healthcare,” Dr. Akhloufi stated. “This approach lays a strong foundation for future studies and opens new avenues for researchers to build upon. Ultimately, it brings us closer to real-world impact, especially in contexts where spatial transcriptomics remains prohibitively expensive to implement.”

Methodology

EMGP-Net is a fusion of two state-of-the-art deep learning models: MambaVision, a recently developed hybrid architecture that combines efficiency with global pattern

recognition, and EfficientFormer, a transformer model previously validated for predicting gene expression in tissues.

Dr. Akhloufi explained that rather than simply combining features from MambaVision and EfficientFormer, the architecture of EMGP-Net uses multi-head attention to weight and integrate information from both models. This fusion approach enables the model to focus on the most relevant features from each component, thereby improving prediction accuracy.

The research team trained the model on two datasets: a HER2+ dataset containing 36 tissue sections from 8 patients, and the ST-Net dataset containing 68 sections from 23 patients. Both datasets included paired histopathological images and spatial transcriptomics data, allowing the model to learn relationships between tissue morphology and gene expression patterns.

The researchers used leave-one-patient-out cross-validation for internal testing and external validation across both datasets to ensure that the model can generalize to new patient populations.

Clinical Performance

In internal validation using the HER2+ dataset, EMGP-Net achieved the highest Pearson correlation coefficient (PCC) of 0.7903 for the PTMA gene, with 14 genes showing PCC values higher than 0.7. These included breast cancer biomarkers such as GNAS (PCC: 0.7843), B2M (PCC: 0.7777), and HNRNPA2B1 (PCC: 0.7532).

When trained on the HER2+ dataset and tested on the ST-Net dataset, EMGP-Net achieved the highest PCC of 0.7145 for the ERBB2 gene and outperformed state-of-the-art models, including GeNetFormer and ST-Net, across most genes.

The model successfully predicted expression levels for all 250 targeted genes with positive correlation coefficients, demonstrating robust performance across the entire gene panel. For 145 of the genes in the panel, the model achieved PCC values greater than 0.5, outperforming other models.

Potential Clinical Implications

Dr. Akhloufi emphasized that by enabling gene expression analysis using standard H&E-stained tissue slides, which are routinely prepared in pathology labs worldwide, EMGP-Net could democratize access to molecular tumor profiling.

“The strength of computer vision models lies in their ability to extract meaningful patterns from images; in our case, the model also leverages the additional spatial transcriptomics information,” Dr. Akhloufi noted. “The overarching goal of this research is to simplify the diagnostic process while achieving strong performance, comparable to results obtained using costly spatial transcriptomics platforms.”

He added that EMGP-Net could enhance treatment selection for patients with breast cancer, helping clinicians identify candidates for targeted therapies. The ability of the model to predict the expression of genes like ERBB2 (HER2), GNAS, and other established biomarkers could inform decisions about HER2-targeted treatments, hormone therapy, and immunotherapy.

Addressing Model Limitations and Future Directions

The researchers acknowledge that several limitations must be addressed before clinical implementation of the model. The relatively small size of available datasets may limit generalizability across diverse patient populations, and differences in tissue preparation, scanning equipment, and patient characteristics could affect model performance on new data.

“We recognize that integrating AI models like EMGP-Net into clinical workflows presents both logistical and technical challenges, including data standardization, model interpretability, and alignment with existing diagnostic protocols,” Dr. Akhloufi stated. “To address these, we plan to collaborate with colleagues in oncology and with a cancer research institute in Atlantic Canada.”

The team is also interested in incorporating explainable AI techniques to enhance model interpretability.

“While our current model does not yet include explainable AI components, we recognize the critical importance of interpretability,” Dr. Akhloufi emphasized. “Understanding which regions the model focuses on can reveal its strengths and limitations, guiding further improvements.”

Dr. Akhloufi envisions that the model could have applications beyond breast cancer.

“EMGP-Net has the potential to be adapted to other cancer types or tissue pathologies,” he noted. “However, successful generalization depends heavily on the availability of high-quality data; specifically, spatial transcriptomics and histopathological images with sufficient sample diversity for each condition.”

Dr. Akhloufi concluded by predicting that over the next 5–10 years, machine learning will play an increasingly transformative role in spatial biology and digital pathology.

“Rather than replacing clinicians, AI will serve as a powerful tool to augment their expertise, enabling faster, more accurate, and more informed decision-making,” he added.

References

1. Thâalbi O, Akhloufi MA. EMGP-Net: A Hybrid Deep Learning Architecture for Breast Cancer Gene Expression Prediction. Computers. 2025; 14(7):253. https://doi.org/10.3390/computers14070253