Causal network inference based on cross-validation predictability

Yuelei Zhang, Qingcui Li, Xiao Chang, Luonan Chen, Xiaoping LiubioRxiv 2022 · DOI

TL;DR

This paper introduces Cross-Validated Predictability (CVP), a novel data-driven algorithm for inferring causal relationships in biological networks from any type of observed data, not just time series. CVP demonstrates superior accuracy and robustness compared to existing methods, particularly in identifying crucial "driver genes" involved in cancer progression, which were experimentally validated.

What Problem Does This Paper Address?

Understanding the causal relationships between molecules (like genes) is crucial for unraveling biological mechanisms and developing treatments, especially in complex diseases like cancer. While many methods exist for inferring these relationships, most rely on time-series data, which is often unavailable for biological systems. Existing methods for non-time-series data often struggle with complex network structures like feedback loops or require extensive prior knowledge. This leaves a significant gap in our ability to accurately infer causal networks from diverse biological datasets.

Methods

The core of this paper is the Cross-Validated Predictability (CVP) algorithm. CVP operates on any observed data (not just time series) and works by assessing how well one variable can predict another after accounting for other variables.

Here's a simplified breakdown:

  1. Hypothesis Testing: For any two variables (say, gene X and gene Y), CVP sets up two competing hypotheses:
  2. H₀ (Null Hypothesis): X does not cause Y. Y can be predicted using other related variables (Z) but not by X.
  3. H₁ (Alternative Hypothesis): X does cause Y. Y can be better predicted when X is included along with other variables (Z).
  4. Cross-Validation: To test these hypotheses, the data is split into training and testing sets multiple times (k-fold cross-validation). Models are trained on the training sets and tested on the unseen testing sets.
  5. Predictive Error: The algorithm calculates the prediction error (Sum of Squared Errors, SSE) for both hypotheses on the testing sets. If the error is significantly lower when X is included (H₁), it suggests X causes Y.
  6. Causal Strength: A "causal strength" metric (calculated as the log of the ratio of errors) quantifies the degree of causality.

This approach is "data-driven" and "model-free," meaning it doesn't assume a specific underlying model and can handle complex network structures like feedback loops, unlike some other methods.

Datasets

The CVP algorithm was validated on a wide range of datasets:

  • Simulated Data: Two sets of simulated causal networks were generated to test the algorithm's accuracy in controlled environments.
  • Benchmark Datasets:
    • DREAM Challenges (DREAM3, DREAM4): These are widely used benchmark datasets for gene network inference, including both simulated and real data.
  • IRMA dataset: A synthetic yeast network used for benchmarking.
  • Real Biological Datasets:
    • SOS DNA repair network (E. coli): A well-studied gene regulatory network.
  • Yeast cell cycle data: Time-series gene expression data.
  • Human HeLa cell cycle data: Gene expression data.
  • BLCA (Bladder Urothelial Carcinoma) dataset from TCGA: Cancer gene expression data.
  • Large-scale E. coli dataset (GSE20305): Time-series data for a complex system.
  • Plankton food chain data: Time-series data on species abundance.
  • Air pollution and respiratory disease data (Hong Kong): Environmental and health data.
  • COVID-19 transmission data (Japan): Epidemiological data.
  • Gastric and Lung Cancer datasets (TCGA): Gene expression data for cancer progression studies.
  • Glioblastoma Multiforme (GBM) data: Cancer gene expression data.

Key Insights and Ideas

  • Causality from Any Data: The most significant insight is that CVP can infer causal relationships from any observed data, not just time series. This dramatically expands the applicability of causal inference in biology.
  • Robustness to Network Complexity: CVP can handle complex biological networks, including those with feedback loops, which are common but challenging for many existing methods.
  • Identifying "Driver Genes": CVP excels at identifying "functional driver genes" – genes that play a critical role in disease progression, particularly cancer. This is a crucial step towards targeted therapies.
  • Experimental Validation: The paper provides strong experimental validation for CVP's predictions, particularly in liver cancer, where knocking out predicted driver genes significantly inhibited cancer cell growth.
  • Data-Driven Simplicity: The algorithm requires minimal prior knowledge or specific hypotheses, making it broadly applicable across different biological systems.

Key Results

  • Superior Performance: CVP consistently outperformed eight other popular algorithms across various benchmark and real-world datasets, achieving higher AUROC (Area Under the Receiver Operating Characteristic Curve) and AUPR (Area Under the Precision-Recall Curve) values. For instance, in DREAM challenges, CVP achieved AUROC values up to 0.86, significantly higher than other methods.
  • Accurate Network Reconstruction: CVP accurately reconstructed known causal networks, often with fewer false positives and negatives than competing methods. For example, in the SOS DNA repair network, CVP inferred 16 edges with only 3 false positives, achieving 81% accuracy.
  • Identification of Novel Cancer Genes: CVP identified 15 functional driver genes in late-stage gastric cancer, 14 of which were previously associated with the disease. Similarly, it identified 15 genes for lung cancer progression, with 13 validated by literature.
  • Experimental Validation of Driver Genes: Knockout experiments in liver cancer cells using CRISPR-Cas9 confirmed that the top 10 predicted driver genes significantly inhibited cell proliferation and colony formation. Four specific genes (SNRNP200, XRN2, RALGAPB, SRSF10) induced about 50% growth inhibition.
  • Validation of Regulatory Relationships: Eight specific regulatory relationships predicted by CVP in liver cancer were experimentally validated through gene knockout experiments, showing strong downregulation of target genes.

Limitations and Future Work

  • Computational Cost for Large Networks: While CVP is efficient, the paper notes that statistical tests for causality can be time-consuming for very large-scale networks, suggesting the causal strength metric is a more scalable approach.
  • Potential for "False Positives" in Incomplete Pathways: In cases where the known biological pathways are incomplete (like in the BLCA dataset), CVP might identify "false positives" that are actually real regulatory edges not yet documented. This highlights the need for comprehensive biological knowledge bases.
  • Future Work: The authors suggest that CVP has significant implications for understanding disease mechanisms and biological functions, implying its application to a wider range of network problems across various fields. The experimental validation of driver genes also opens doors for further wet-lab investigations.

Summary updated

Abstract

Abstract Identifying causal relations or causal networks among molecules/genes, rather than just their correlations, is of great importance but challenging in biology and medical field, which is essential for unravelling molecular mechanisms of disease progression and developing effective therapies for disease treatment. However, there is still a lack of high-quality causal inference algorithms for any observed data in contrast to time series data. In this study, we developed a new causal concept for any observed data and its causal inference algorithm built on cross-validated predictability (CVP) can quantify the causal effects among molecules/genes in the whole system. The causality was extensively validated by combining a large variety of statistical simulation experiments and available benchmark data (simulated data and various real data). Combining the predicted causal network and the real benchmark network, the CVP algorithm demonstrates high accuracy and strong robustness in comparison with the mainstream algorithms. In particular, the CVP algorithm is robust in identifying reliable driver genes and network biomarkers from the perspective of network biology, with the prediction results outperforming the mainstream conventional methods for predicting driver genes. CRISPR-Cas9 knockdown experiments in the liver cancer have validated that the functional driver genes identified by the CVP algorithm effectively inhibit the growth and colony formation of liver cancer cells. By knockdown experiments, we demonstrated the accuracy and significance of the causality predicted by CVP and identified the novel regulatory targets of functional driver genes SNRNP200 and RALGAPB in the liver cancer. These inferred causal networks explain regulatory patterns in different biological contexts in a clear sense and provide biological insights into molecular mechanisms of disease progression from a causality perspective.
BibTeX
@misc{zhang2022causal,
  author = {Yuelei Zhang and Qingcui Li and Xiao Chang and Luonan Chen and Xiaoping Liu},
  title = {{Causal network inference based on cross-validation predictability}},
  year = {2022},
  doi = {10.1101/2022.12.11.519942},
  howpublished = {bioRxiv},
}
Figure