Bi-level graph learning unveils prognosis-relevant tumor microenvironment patterns in breast multiplexed digital...

Area of Science:

• Computational biology
• Cancer research
• Artificial intelligence in medicine

Background:

• The tumor microenvironment (TME) is crucial for cancer progression and patient outcomes.
• Deep learning methods are increasingly used to characterize the TME.
• Interpretability challenges in deep learning limit the identification of generalizable biomarkers.

Purpose of the Study:

• To develop a data-driven and interpretable approach for identifying TME cellular patterns associated with patient prognosis.
• To create a risk-stratification system for breast cancer patients based on TME cellular patterns.

Main Methods:

• Construction of a bi-level graph model, including a cellular graph for TME modeling and a population graph for inter-patient similarity.
• Application of the method to breast cancer data.
• Validation of identified patterns in two independent patient cohorts.

Main Results:

• Identified prognostic cellular patterns within the TME.
• Developed a risk-stratification system that provides complementary information to standard clinical subtypes.
• Demonstrated the generalizability of the approach through validation in independent cohorts.

Conclusions:

• The proposed interpretable deep learning approach effectively identifies TME cellular patterns linked to patient prognosis.
• This methodology offers a valuable tool for risk stratification in breast cancer and has potential applications in other cancer types.
• The findings highlight the importance of spatial cellular patterns in the TME for understanding and predicting patient outcomes.