An algorithm for direct causal learning of influences on patient outcomes

Area of Science:

• Causal inference and machine learning
• Computational biology and bioinformatics
• Genetics and genomics

Background:

• Learning direct causal influences is crucial for understanding complex biological systems and diseases.
• Existing causal learning algorithms have limitations in accuracy and efficiency when applied to real-world data.
• Genome-wide association studies (GWAS) and clinical datasets offer valuable insights but require robust causal discovery methods.

Purpose of the Study:

• To introduce and evaluate a novel algorithm, Direct Causal Learner (DCL), for identifying direct causal relationships.
• To compare DCL's performance against established causal learning algorithms using simulated and real-world datasets.
• To assess DCL's utility in identifying causal factors for late-onset Alzheimer's disease (LOAD) and breast cancer survival.

Main Methods:

• Developed the Direct Causal Learner (DCL) algorithm, utilizing Bayesian scoring and a novel deletion approach.
• Generated 14,400 simulated datasets to rigorously test DCL against Path Consistency (PC), Conservative PC (CPC), Fast Greedy Search (FGS), and Fast Causal Inference (FCI) algorithms.
• Applied DCL and other algorithms to a real GWAS dataset for LOAD and breast cancer datasets (Metabric) to evaluate predictive accuracy and identify clinical risk factors.

Main Results:

• DCL significantly outperformed FGS, PC, CPC, and FCI in correctly predicting direct causes in simulated datasets (McNemar's test: p<<0.0001).
• DCL demonstrated superior performance in partially predicting direct causes for complex networks and exhibited faster runtimes compared to other algorithms.
• In real-world applications, DCL identified known causal genetic variants for LOAD (rs6784615, rs10824310) and significant predictors for breast cancer mortality (ER, HER2 status).

Conclusions:

• The Direct Causal Learner (DCL) algorithm represents a significant advancement in causal learning, consistently outperforming existing methods.
• DCL's effectiveness in identifying direct causes in both simulated and real clinical GWAS data highlights its potential for advancing biomedical research and clinical applications.
• The algorithm's ability to uncover known causal relationships validates its predictive power and suggests broad applicability in disease etiology research.