基於轉錄體學關鍵特徵法之生殖毒性篩選策略

Abstract

近年來工業化驅動化學物質多樣性與使用量快速擴增，然而大多數物質缺乏全面的毒理學資料，其中生殖與發育毒性相關資料的匱乏尤為嚴重（僅約2%化學物質具備相關測試資料），凸顯了高效生殖毒性篩選的迫切需求。生殖與發育毒性一直是非動物替代測試方法（New Approach Methodologies, NAMs）中最難處理的終點之一，因為它牽涉內分泌調控、配子形成、著床、胚胎發育等複雜且跨尺度的生物學過程，難以用單一替代性方法完整涵蓋。為填補此缺口，本研究開發一套「關鍵轉錄體特徵（Key Transcriptomic Characteristics, KTC）」架構，整合高通量轉錄體反應與既有的生殖毒性關鍵特徵法（Key Characteristics, KCs），用於化學安全性評估中的化學物質篩選與優先排序。 本研究的轉錄體數據取自基因表現綜合資料庫（GEO, GSE272548），涵蓋1,750種化學物質於MCF-7細胞中經6小時暴露後的反應。整合六項資料庫與清單後，共分類出549種生殖毒性物質與1,056種非生殖毒性物質；再扣除4種缺乏可用轉錄體路徑資料而視為生物不活性的物質後，建立了包含1,601種化學物質的參考資料集（包含546種生殖毒性物質、1,055種非生殖毒性物質）。透過差異基因表現分析、路徑富集分析，並以逐步優化的方式將富集倍數閾值訂為2.11倍，並建構出涵蓋112條生物學預警路徑的清單。這些路徑主要集中在發炎反應（KC7）與訊號傳導改變（KC9）兩大機制領域，兩者合計超過清單半數，且經邏輯迴歸證實各自獨立與生殖毒性物質狀態顯著相關（勝算比分別為1.91與1.44）。 比較三個模型的表現後發現，基於生物學效應的KTC模型與兩種傳統模型（VEGA CAESAR model、DART決策樹）呈現不同、但彼此互補的表現特性。DART決策樹的特異度（81.7%）與平衡準確率（68.2%）為三者最高；VEGA CAESAR模型敏感度最高（74.1%），但特異度最低（37.4%）；KTC模型的表現介於兩者之間（平衡準確率63.0%），並達到100%的化學物質涵蓋率，能評估因缺乏明確化學結構而無法處理的物質。McNemar's檢定證實三模型的分類錯誤模式彼此系統性不同（p<0.0001），因此本研究進一步建立共識分類架構，將陽性預測值與陰性預測值分別提升至67.0%與88.9%。 KTC模型應用於篩選我國化學物質優先排序計畫之案例研究，不僅成功辨識已有充分生殖毒性證據的物質（三模型一致判定陽性者代表與GHS分類100%相符），且具有篩選出結構模型未能標記的候選物質，可信度較高（KTC獨有陽性物質有75%獲GHS支持，相對於VEGA獨有陽性物質僅25%）。進一步將此模型應用於更廣泛的化學物質清單後，在145種目前完全缺乏GHS或其他法規生殖毒性分類的物質中，成功標記出68種（46.9%）具備可偵測生物學警訊的候選物質。這些結果支持以HTTr為基礎的KTC模型，作為一套具擴充性且機制可解釋的生殖毒性篩選與優先排序策略，有助於協助法規單位優先排序高風險候選物質以進行後續測試。

Rapid industrialization has exponentially increased chemical diversity and usage, yet comprehensive toxicological data remain scarce for most substances—reproductive and developmental toxicity data are particularly limited, with only about 2% of chemicals having relevant testing data—highlighting the urgent need for efficient reproductive toxicity screening. Reproductive and developmental toxicity remains one of the most difficult endpoints for new approach methodologies (NAMs), because it involves complex, multiscale biological processes such as endocrine regulation, gametogenesis, implantation, and embryofetal development, which are difficult to capture using any single alternative assay. To help address this gap, this study developed and benchmarked a Key Transcriptomic Characteristics (KTC) model that integrates high-throughput transcriptomic responses with the established Key Characteristics (KCs) of reproductive toxicants for chemical screening and prioritization in chemical safety evaluation. Using transcriptomic data from the Gene Expression Omnibus (GSE272548), which evaluated 1,750 chemicals following a 6-hour exposure in MCF-7 cells, this study classified 549 reproductive toxicants and 1,056 non-reproductive toxicants using six authoritative databases and lists; after excluding 4 chemicals lacking usable transcriptomic pathway data (bio-inactive), a reference dataset of 1,601 chemicals (546 reproductive toxicants, 1,055 non-reproductive toxicants) was curated. Differential gene expression and pathway enrichment analyses, combined with stepwise threshold optimization (RF ratio = 2.11), identified 112 biological warning pathways associated with reproductive toxicants. These pathways were dominated by two mechanistic domains, inflammation (KC7) and altered signal transduction (KC9), which together accounted for more than half of the list and were independently associated with reproductive toxicant status (OR = 1.91 and 1.44, respectively). Performance benchmarking showed that the biology-based KTC model and two structure-based models, VEGA CAESAR and the DART decision tree, had genuinely different and complementary performance profiles. The DART decision tree achieved the highest specificity (81.7%) and balanced accuracy (68.2%), while VEGA CAESAR achieved the highest sensitivity (74.1%) but the lowest specificity (37.4%); the KTC model fell between these extremes (balanced accuracy 63.0%) while achieving complete chemical coverage (100%), including chemicals that lack the defined structures required by structure-based models. McNemar's tests confirmed that the three models made systematically different classification errors (all p < 0.0001), motivating a consensus classification scheme that raised positive and negative predictive values to 67.0% and 88.9%, respectively. Applied to a regulatory case study of Taiwan's Priority Existing Chemicals, the KTC model recovered chemicals with well-established reproductive toxicity evidence (100% of three-model consensus positives were GHS-confirmed) while also identifying candidates missed by structure-based models with substantially higher reliability (75% GHS-supported for KTC-unique flags versus 25% for VEGA-unique flags). Applied more broadly, the model flagged 68 of 145 chemicals (46.9%) that currently carry no GHS or other regulatory reproductive hazard classification. These findings support the HTTr-informed KTC model as a scalable and mechanistically interpretable approach for reproductive toxicant screening and prioritization, helping regulatory agencies prioritize high-risk candidates for further testing.