| 摘要: | 多囊性卵巢症候群 (PCOS) 是一種常見的內分泌疾病，嚴重影響女性生殖健康，表現為月經不規則、卵巢功能障礙、排卵稀少和生長激素 (GH) 水平降低。生長激素是女性生殖系統中最重要的荷爾蒙之一，有助於卵巢應對壓力。在育齡期，卵巢體細胞，例如顆粒細胞 (GC) 和卵丘細胞 (CC)，長期受到氧化應激，可能因其在激素失調和卵泡發生中的作用而有助於理解激素代謝紊亂的複雜性。透過利用 Vision-Language Tokenizer 下的人工智慧驅動的通路推斷，結合轉錄組分析和 STRING 等精選資料庫得出的差異表達基因 (DEG)，該通路發現有望揭示 PCOS 患者和非 PCOS 患者中促性腺激素誘導 CC 和 GC 細胞表達的生物學機制。
目的：本研究提出了一種綜合系統生物學方法，該方法整合了轉錄組分析、蛋白質-蛋白質相互作用網絡以及基於視覺方法的大型語言模型 (LLM) 的指導，以研究促性腺激素在多囊卵巢綜合徵 (PCOS) 和非多囊卵巢綜合徵 (PCOS) 患者中引發的失調。
結果：透過分析來自合作社 (CC) 的兩個合併的 GEO 微陣列資料集，鑑定出 584 個高置信度的失調 DEG，生物醫學領域 PubMedBERT 的「基因-疾病」和「基因-基因」聚類模型優先考慮了其中 86 個基因。隨後，在 LLM 輔助下，輸入 18 個聚類候選蛋白和 GO 術語，揭示了 mRNA 加工和 NAD⁺ 生物合成中存在重大干擾。具體而言，NMNAT3、CPSF4 和 NUP210 的下調表明，在氧化壓力背景下，與伴侶蛋白 CCT4 介導的上調相關的線粒體 NAD⁺ 生物合成可能減少，從而導致 β 細胞基因表達受抑制的轉錄機制可能存在缺陷。
方法：本研究利用視覺高級法學碩士 (LLM)，包括 Scholar GPT-4o、Claude Sonnet-4、Gemini 2.5 Flash 和 GitHub Copilot，構建了一條推斷通路，並透過視覺語言任務提示，將觀察到的轉錄組變化與潛在的代謝功能障礙聯繫起來。 PubMedBERT 篩選顯示，基因間語意相似度得分較高，STRING 資料庫也支持了蛋白質-蛋白質交互作用。在 LLM 提示之前檢索 GO 術語和通路，為重新組裝 PCOS 患者的分子通路建立了一個強有力的框架，並強調了發炎信號傳導。;Polycystic Ovary Syndrome (PCOS) is a common endocrine disorder that significantly affects female reproductive health, manifesting as menstrual irregularities, ovarian dysfunction, oligo-anovulation and lowered growth hormone (GH) level. GH, one of the most significant hormones in female reproductive system, helps the ovaries respond to distress. During reproductive ages, the prolonged induced oxidative stress ovarian somatic cells, such as Granulosa cells (GCs) and Cumulus cells (CCs), may provide insights into the complexities of hormonally metabolic disorders, due to their role in hormonal dysregulation and folliculogenesis. By utilizing AI-driven pathway inferences under Vision-Language Tokenizer with DEGs from transcriptomic analysis and curated databases like STRING, the pathway discovery can potentially comprehend biologically plausible mechanisms related to gonadotropin induction on CCs and GCs between PCOS and non-PCOS patients.
Aim: This study presents a comprehensive systems biological approach that integrates results of transcriptomic analysis, protein-protein interaction networks, and the vision-approached large language models (LLMs) guidance to investigate the dysregulations triggered by gonadotropins in both PCOS and non-PCOS patients.
Results: Analysis of two merged GEO microarray datasets from CCs identified 584 high-confidence dysregulated DEGs that biomedical-domain PubMedBERT clustering Genes-Diseases and Gene-Gene prioritizes 86 genes. Subsequently, LLMs-assisted input with 18 clustered candidate proteins and GO terms revealed major disturbances in mRNA processing and NAD⁺ biosynthesis. Particularly, downregulation of NMNAT3, CPSF4 and NUP210 points to putative deficiencies in the transcriptional machinery involved in suppressed β-cell gene expressions under a potential depletion of mitochondrial NAD⁺ biosynthesis related to chaperonin CCT4-mediated upregulation in the context of oxidative stress.
Methodology: The study utilizes the vision-advanced LLMs, including Scholar GPT-4o, Claude Sonnet-4, Gemini 2.5 Flash, and GitHub Copilot, to construct an inferred pathway prompting with vision-language tasks that connects the observed transcriptomic changes to underlying metabolic dysfunctions. PubMedBERT filtering show the high semantic similarity score across the genes along with STRING database, reinforcing the protein-protein interactions. The retrievals of GO terms and pathways prior to LLMs prompting establish a strong framework for reassembling the molecular pathways in PCOS patients, emphasizing inflammatory signaling. |
