Abstract
Identifying the underlying causes of cancer survival is a critical challenge in precision medicine. While Causal Discovery (CD) algorithms offer a path toward uncovering these mechanisms from observational data, the lack of ground-truth validation often necessitates labor-intensive manual literature reviews. This paper presents a novel framework that automates this process by employing Large Language Models (LLMs) to validate causal claims derived from multi-omics breast cancer data. Our results demonstrate that LLMs can effectively act as automated knowledge proxies, significantly accelerating the discovery of biologically relevant biomarkers.
Introduction
In the clinical domain, distinguishing between a correlation and a causal driver is paramount. While traditional machine learning models excel at predicting patient outcomes, they often fail to provide the "why" behind a prediction. Causal Discovery (CD) aims to address this by constructing Directed Acyclic Graphs (DAGs) that represent cause-and-effect relationships. However, the complexity of biological systems makes validating these graphs difficult. This study proposes an end-to-end pipeline that utilizes the vast knowledge encoded in biomedical LLMs to verify the causal links identified by CD algorithms.
Methodology
The proposed pipeline follows a structured approach to transition from raw genomic data to validated causal insights.
1. Feature Selection and Data Preparation
Using the TCGA Breast Cancer dataset (705 patients), the researchers initially faced nearly 2,000 variables across somatic mutations, copy number variations, gene expression, and protein levels. To manage this high dimensionality, Mutual Information (MI) and Max-Min Markov Blanket (MMMB) techniques were employed to select the 10–20 most influential features relative to patient survival.
2. Causal Discovery
Several algorithms were applied to the filtered data to build the causal graphs:
- Constraint-based (PC, FCI): Utilizing statistical independence tests.
- Score-based (FGES): Searching for the most probable graph structure.
- Generalized Precision Matrix (GPM): A specialized approach for mixed discrete and continuous data.
3. LLM-Based Validation
The core innovation lies in using LLMs (BlueBERT and SciBERT) as automated validators. Two primary metrics were used to determine if the discovered links matched existing medical literature: Perplexity Scores (measuring the model's "surprise" at a claim) and Masked Language Modeling (testing the model's ability to predict missing biological terms within a causal statement).
Results and Discussion
The integration of the Generalized Precision Matrix (GPM) with the PC algorithm yielded the most biologically plausible results. The study identified several key genetic factors, including MLL3 and TNXB, which are known biomarkers for breast cancer progression and metastasis.
As illustrated in the figure above, the distribution of copy number variations in genes like IDO1 shows distinct patterns between surviving and deceased patient cohorts, providing empirical support for the features selected by the causal model.
In terms of validation, BlueBERT emerged as the superior validator, achieving an accuracy of 81.25% in the MLM task. Furthermore, perplexity analysis showed that validated causal claims had significantly lower scores (31.01) compared to their negations (2502.9), indicating that the LLM's internal knowledge aligns strongly with the discovered causal structures.
Conclusion
This research demonstrates a scalable method for validating AI-driven medical insights. By treating LLMs as "automated peer reviewers," researchers can bypass the manual validation bottleneck. Future work will focus on addressing the "staleness" of LLM training data and extending the framework to handle cyclic biological relationships, which are common in feedback-driven genetic pathways.
Summary updated