Artificial intelligence and digital pathology for personalised cancer medicine

Presenter: Dr Danh-Tai Hoang 

Precision oncology is increasingly becoming integral to clinical practice, showing notable improvements in treatment outcomes. While molecular data such as mRNA gene expression and DNA methylation profiles offer comprehensive information, obtaining this data in clinical laboratories can take several weeks. This delay poses a challenge for patients with high-grade tumours who often require immediate treatment. Additionally, the costs and limited availability of sequencing, especially in developing countries, present further challenges. In contrast, histopathology images are routinely available, cost-effective, and timely.

In this talk, I will first present DEPLOY, an integrated deep learning framework that predicts DNA methylation from histopathology images, and subsequently classifies brain tumours into 10 major subtypes. The model was trained on a dataset of 1,796 patients, and subsequently tested on three independent datasets comprising 2,156 patients. Remarkably, DEPLOY achieved an accuracy of 95% in highly confident predicted samples [1]. Second, I will present DeepPT, another deep learning framework that predicts mRNA gene expression from histopathology images across 16 cancer types. DeepPT significantly outperformed other existing approaches, including HE2RNA, SEQUOIA, and tRNAformer. More importantly, we found that the predicted gene expression can be used to successfully predict patient response to cancer therapies across multiple cancer types [2]. Lastly, I will briefly present our ongoing efforts in predicting spatial transcriptomics and genetic mutations from histopathology images.

References

[1]. D-T. Hoang et al. (2024), “Prediction of DNA methylation-based tumor types from histopathology in central nervous system tumors with deep learning”, Nature Medicine.

[2]. D-T. Hoang et al. (2024), “A deep-learning framework to predict cancer treatment response from histopathology images through imputed transcriptomics”, Nature Cancer (accepted).