Breast cancer is one of the most common malignancies worldwide, and mutations in the PI3K/AKT/mTOR (PAM) signaling pathway are prevalent in its development. Among these, PIK3CA mutations play a pivotal role in guiding treatment with PI3K inhibitors, which have shown promising anti-tumor effects. However, conventional molecular assays like polymerase chain reaction (PCR) and next-generation sequencing (NGS) require expensive infrastructure and are not always feasible in routine clinical practice. Deep learning models have emerged as a cost-effective solution, predicting key mutations from digital pathology images. Despite this, most existing models rely on single-modal data, often lacking the complementary insights that structured clinical data can provide. These challenges highlight the need for improved prediction models.
In a study published (DOI: 10.20892/j.issn.2095-3941.2025.0771) in Cancer Biology & Medicine in February 2026, a team of researchers from Hebei Medical University Fourth Hospital developed a novel multimodal artificial intelligence (AI) model for predicting PIK3CA mutations in breast cancer. This model integrates deep learning-based analysis of whole-slide pathology images with structured clinical data, including age, molecular subtype, and lymph node status. The research utilized data from The Cancer Genome Atlas (TCGA) and three external clinical cohorts, demonstrating the model's robustness and its potential as an accessible alternative to molecular testing in diverse clinical settings.
The study's multimodal framework, known as the Multimodal PIK3CA Model (MPM), combines two components: a histopathology model and a clinical model. The histopathology model processes high-resolution whole-slide images using a transformer-based pretrained encoder (H-optimus-0) and a clustering-constrained attention multiple instance learning classifier (CLAM-SB). This model identifies morphological features associated with PIK3CA mutations. The clinical model, based on XGBoost, analyzes structured clinical data to predict mutation status. Both models generate independent probability predictions, which are fused using a decision-level late fusion strategy to produce a final mutation status prediction. The MPM outperformed single-modality models, achieving an area under the curve (AUC) of 0.745 in internal testing, with stable performance across external validation datasets (0.695 to 0.680 AUC). The inclusion of clinical variables, such as molecular subtype and lymph node status, enhanced the model's predictive accuracy, highlighting the importance of combining morphological and clinical data. The study also demonstrated the model’s ability to generalize across diverse cohorts, making it a promising tool for real-world clinical application.
Dr. Yueping Liu, the lead corresponding author of the study,remarked, “This multimodal AI framework represents a significant advancement in computational pathology. By integrating complementary clinical and morphological data, our model not only enhances the prediction of PIK3CA mutations but also offers a scalable, cost-effective solution for clinical practice. With its strong generalization across diverse cohorts, it has the potential to improve personalized treatment decisions for breast cancer patients, bridging the gap between advanced molecular testing and routine clinical workflows.”
The MPM's robust performance and ability to incorporate both digital pathology and clinical data make it a valuable tool for clinical decision support. The model provides a practical, cost-effective alternative to traditional molecular testing, which is often inaccessible in resource-limited settings. With its strong generalizability across different medical centers and patient cohorts, the MPM could be deployed in routine clinical practice to predict PIK3CA mutations in breast cancer, thus guiding the use of PI3K-targeted therapies. Future research may focus on refining the model for other mutations and cancers, expanding its applicability in precision oncology.
###
References
DOI
10.20892/j.issn.2095-3941.2025.0771
Original Source URL
https://doi.org/10.20892/j.issn.2095-3941.2025.0771
Funding information
This study was financially supported by the Hebei Natural Science Foundation (Grant No. H2024206504), the Medical Science Research Project of Hebei (Grant No. 20260484, 20260530), and the Fundamental Research Funds for the Central Universities (Grant No. 20822041J4123).
About Cancer Biology & Medicine
Cancer Biology & Medicine (CBM) is a peer-reviewed open-access journal sponsored by China Anti-cancer Association (CACA) and Tianjin Medical University Cancer Institute & Hospital. The journal monthly provides innovative and significant information on biological basis of cancer, cancer microenvironment, translational cancer research, and all aspects of clinical cancer research. The journal also publishes significant perspectives on indigenous cancer types in China. The journal is indexed in SCOPUS, MEDLINE and SCI (IF 8.4, 5-year IF 6.7), with all full texts freely visible to clinicians and researchers all over the world (http://www.ncbi.nlm.nih.gov/pmc/journals/2000/).