非靶向代謝體學 (Untargeted metabolomics) 高度仰賴來自液相層析串聯質譜 (LC-MS/MS) 資料的準確代謝物鑑定。然而，由於缺乏全面且具實用性的評估，要選擇最可靠的軟體工具仍是一大挑戰。為了彌補此一缺口，我們系統化評估了四種廣泛使用的開源軟體工具（MS-DIAL 5.3.240719、MZmine 4.2.0、SIRIUS 6.0.3 與 BUDDY 1.7），並探討將其輸出結果結合為集成方法 (ensemble approach) 是否能減輕個別軟體的偏差，並將鑑定準確率最大化。 本研究評估了傳統的參考光譜庫搜尋 (spectral library searches) 以及進階的運算型（從頭/由下而上，de novo/bottom-up）分子式註解策略。評估使用了八個公開的標準代謝物資料集（包含 endogenous、CASMI 2022、plant、urinary、amino acids、primary endogenous、IROA 與 endo-lipids），這些資料集是透過高解析度質譜儀在數據依賴型擷取 (DDA) 模式下所取得，並以北美質譜庫 (MassBank of North America， MoNA) 作為參考標準。效能評估涵蓋了多項實務指標：真實鑑定率 (identification rates， IR)、分數的鑑別力、分子質量範圍的影響、加合物 (adduct) 偵測準確率以及運算時間。為了評估實務上的下游效用，本研究在一個包含 480 個小鼠大腦樣本的真實實驗資料集上，利用鑑定出的代謝物訓練隨機森林分類器 (Random Forest classifiers)，以對大腦區域、年齡與性別等生物特徵進行分類。 結果顯示，MS-DIAL 是在光譜庫搜尋方面最穩健的單一工具，達到了最高的平均鑑定率 (87%)。SIRIUS 與 MZmine 表現相當（皆為 80% IR），而 BUDDY 表現較差（65% IR）。在光譜庫搜尋的運算時間方面，MZmine 展現出最高的效率（總計 117 分鐘），其次為 BUDDY（234 分鐘）與 MS-DIAL（256 分鐘）；而 SIRIUS 由於採用了複雜的碎片樹 (fragmentation tree) 演算法，因此需要最長的運算時間（1，034 分鐘）。儘管 SIRIUS 與 BUDDY 的運算方法成功鑑定出了 158 個未收錄於參考光譜庫中的新穎「暗物質 (dark matter)」分子式，但其偽陽性風險明顯較高。本研究亦證實了準確的加合物偵測與成功的代謝物鑑定之間，存在著極顯著的統計相關性。關鍵的是，將四款工具整合後，不僅將光譜庫搜尋的綜合鑑定率提升至 93%，更在下游的預測建模中成功減輕了個別工具的偏差，為複雜的生物分類提供了高度穩健且可靠的生物標記。;Untargeted metabolomics relies heavily on accurate metabolite identification from liquid chromatography-tandem mass spectrometry (LC-MS/MS) data， yet selecting the most reliable software remains challenging due to a lack of comprehensive， practical evaluations. To address this gap， we systematically evaluated four widely used open-source tools (MS-DIAL 5.3.240719， MZmine 4.2.0， SIRIUS 6.0.3， and BUDDY 1.7) and investigated whether an integrated ensemble approach combining their outputs could mitigate individual software biases and maximize identification accuracy. We assessed both traditional spectral library searches and advanced computational (de novo/bottom-up) formula annotation strategies. The evaluation used eight public standard metabolite datasets acquired via high-resolution mass spectrometers in data-dependent acquisition (DDA) mode， referenced against the MassBank of North America (MoNA) library. Performance was measured across multiple practical metrics: identification rates (IR)， score discriminatory power， molecular mass range impacts， adduct detection accuracy， and computation time. To evaluate practical downstream utility， Random Forest classifiers were trained on the identified metabolites from a real experimental dataset of 480 mouse brain samples to classify biological traits， including brain region， age， and sex. Results indicate that MS-DIAL is the most robust individual tool for spectral library search， achieving the highest average identification rate (87%). SIRIUS and MZmine performed comparably (80% IR)， while BUDDY underperformed (65% IR). Regarding computation time for spectral library searches， MZmine was the most efficient (117 minutes total)， followed by BUDDY (234 minutes) and MS-DIAL (256 minutes)， whereas SIRIUS required the longest time (1，034 minutes) due to its complex fragmentation tree algorithms. While computational (de novo/bottom-up) approaches in SIRIUS and BUDDY successfully identified 158 novel "dark matter" formulas absent from reference libraries， they exhibited a significantly higher risk of false positives. We also confirmed a highly significant statistical correlation between accurate adduct detection and successful metabolite identification. Crucially， integrating all four tools boosted the combined identification rate in spectral library search to 93% and successfully mitigated individual tool biases in downstream predictive modeling， yielding highly robust and reliable biomarkers for complex biological classifications.
