Deciphering spatial-temporal mechanisms of PD-1 blockade resistance via biologically informed machine learning

Area of Science:

• Oncology
• Computational Biology
• Immunotherapy

Background:

• Immune checkpoint inhibitors (ICIs) have revolutionized cancer treatment, but response rates are limited, and resistance is common.
• The rapid expansion of multi-omics and spatial profiling data creates a bottleneck in extracting mechanistic insights for immuno-oncology.
• Current biomarkers lack spatiotemporal resolution, and AI models can be uninterpretable.

Purpose of the Study:

• To introduce Biologically Informed Machine Learning (BIML) as a novel paradigm to overcome limitations in current cancer therapy research.
• To integrate biophysical principles and biological priors into predictive models for enhanced mechanistic understanding.
• To bridge the gap between computational inference and clinical application through integrated validation strategies.

Main Methods:

• Embedding biophysical principles (e.g., pharmacokinetic ODEs) and biological priors (e.g., PPI networks) into machine learning models.
• Applying BIML for integrative decoding of tumor immune microenvironment (TME) heterogeneity.
• Utilizing orthogonal ex vivo validation methods like patient-derived organoids and microphysiological systems.

Main Results:

• BIML enables quantitative characterization of T-cell exhaustion dynamics.
• Identification of spatial barriers to immunotherapy, such as fibroblast-mediated immune exclusion.
• Transformation of static spatial data into dynamic, testable trajectories for deeper biological insight.

Conclusions:

• BIML offers a powerful approach to generate actionable insights in immuno-oncology.
• This computation-experiment closed-loop facilitates the prioritization of rational combination strategies.
• Integrating BIML with experimental validation is crucial for advancing cancer therapy and overcoming resistance.