Correlation does not equal causation: the imperative of causal inference in machine learning models for immunotherapy

Area of Science:

• Immunology
• Computational Biology
• Biostatistics

Background:

• Machine learning (ML) is vital for precision immunotherapy, integrating multi-omics data for biomarker discovery and response prediction.
• Immunological studies often over-rely on correlation, neglecting causal inference, limiting traditional ML models' ability to capture immune dynamics.
• A systematic review found no ML/deep learning studies on immune checkpoint inhibitors or melanoma modeling incorporated causal inference.

Purpose of the Study:

• To highlight the knowledge-practice gap in immunotherapy research regarding causal inference in ML.
• To introduce recent advances in causal ML as potential solutions for identifying genuine causal relationships.
• To discuss challenges and propose future directions for implementing causal ML in clinical practice.

Main Methods:

• Systematic review of 90 studies on immune checkpoint inhibitors and 36 melanoma modeling studies.
• Review of recent causal ML advancements (e.g., Targeted-BEHRT, CIMLA, CURE).
• Analysis of challenges in practical implementation of causal ML.

Main Results:

• Despite ML use, no reviewed studies incorporated causal inference, indicating a significant gap.
• Causal ML models can differentiate causation from correlation, integrate multimodal data, and control for confounders.
• Key implementation challenges include poor data quality, algorithmic opacity, methodological complexity, and communication barriers.

Conclusions:

• Bridging the knowledge-practice gap requires advancing causal ML research and developing new platforms.
• Interdisciplinary training programs are crucial for translating causal ML from theory to clinical application.
• Causal ML represents a paradigm shift for immunotherapy research and clinical decision-making within 5-10 years.