🤖 AI Summary
This study addresses the susceptibility of gene biomarkers extracted by deep sequential models to confounding effects from tissue composition, which degrades classification performance. For the first time, it integrates chain-of-thought reasoning from large language models (LLMs) into gene selection, leveraging the Mamba state space model to process TCGA-BRCA RNA-seq data and applying causal feature refinement to gradient-based salient genes for deconfounding. The work introduces the concept of “selective credibility,” demonstrating that precise deconfounding substantially enhances performance—even without full recall of known biomarkers. The LLM-filtered 17-gene signature achieves an AUC of 0.927, significantly outperforming both the original 50-gene set (AUC: 0.832) and a 5,000-gene variance-based baseline (AUC: 0.903), while reducing feature dimensionality by 294-fold.
📝 Abstract
Gradient saliency from deep sequence models surfaces candidate biomarkers efficiently, but the resulting gene lists are contaminated by tissue-composition confounders that degrade downstream classifiers. We study whether LLM chain-of-thought (CoT) reasoning can faithfully filter these confounders, and whether reasoning quality drives downstream performance. We train a Mamba SSM on TCGA-BRCA RNA-seq and extract the top-50 genes by gradient saliency; DeepSeek-R1 evaluates every candidate with structured CoT to produce a final 17-gene set. The raw 50-gene saliency set (no LLM) performs worse than a 5,000-gene variance baseline (AUC 0.832 vs. 0.903), while the LLM-filtered set surpasses it (AUC 0.927), using 294x fewer features. A faithfulness audit (COSMIC CGC, OncoKB, PAM50) reveals only 6 of 17 selected genes (35.3%) are validated BRCA biomarkers, yet 10 of 16 known BRCA genes in the input were missed - including FOXA1. This gap between downstream performance and reasoning faithfulness suggests selective faithfulness: targeted confounder removal is sufficient for performance gains even without comprehensive recall.