A Reproducible RAG-LLM Framework for Literature-Derived Feature Extraction in Parkinson's Disease Blood Transcriptomic

Parkinson's disease (PD) remains difficult to diagnose early, and blood-based transcriptomic biomarkers often fail to generalize across cohorts due to feature instability. Large Language Models (LLMs) offer a way to mine biomedical literature for biological features, but their predictive utility in clinical transcriptomics remains largely unexamined. We present a fully reproducible Retrieval-Augmented Generation (RAG) and LLM pipeline applied to PD blood transcriptomics. Instead of arbitrary probe selection, we curated a pathway-guided set of 48 genes (NK-cell cytotoxicity, Lysosome, and Autophagy). We performed a systematic prompt ablation across three strategies: Simple (0–5), Chain-of-Thought (CoT), and Few-Shot. The Few-Shot prompting strategy produced richer and more interpretable graded evidence than simple prompting. We then integrated these literature-derived features as a deterministic linear combination using Logistic Regression on the GSE99039 discovery cohort (n=438). The baseline model achieved an AUC of 0.8846. The integrated model achieved an identical AUC (ΔAUC = 0.00), demonstrating that this specific literature-weighted linear feature did not provide additive predictive power over transcriptomic expression. External validation on GSE6613 (n=72) showed a consistent null integration effect (AUC = 0.5036), despite 47/48 target genes being successfully recovered across different microarray platforms. SHAP analysis of the baseline model highlighted NK-cell-mediated cytotoxicity as the dominant biological signal. Rather than a 'positive' breakthrough, this study provides a rigorous, open-source benchmark demonstrating that the specific literature-weighted linear feature used here did not provide additive predictive power when added to the same transcriptomic features. We will release the exact prompts, the 48-gene feature matrix, and the pipeline to support future reproducibility and non-linear integration strategies. For Logistic Regression, the null result is algebraically expected. For Random Forest and XGBoost, the near-zero ΔAUC is an empirical observation within this feature set, sample size, and tuning protocol. Control priors (uniform, permuted, and random weights) similarly yielded near-zero ΔAUC, indicating that the null outcome is largely a property of appending a deterministic scalar projection.