從單細胞轉錄體定序資料重建T細胞受體庫之工具比較與效能分析

Abstract

隨著次世代定序技術與單細胞分析方法的快速發展，T細胞受體 (T Cell Receptor, TCR) 成為探討免疫反應與疾病機轉的重要工具。近年來，有一些工具宣稱可以從單細胞轉錄體定序資料 (Single cell transcriptome data) 重建T細胞受體庫，然而，目前針對不同平台與同一患者間不同階段的T細胞受體重建比較研究仍相對有限，尤其是Parse Biosciences平台之應用評估更顯稀少。本研究針對同一葛瑞夫兹氏病 (Graves' diseases)患者於用藥不良反應之急性期 (The acute phase of TiA patient) 與恢復期 (Recovery phase) 兩階段之樣本，採用10x Genomics與Parse Biosciences兩種單細胞平台，系統性比較MiXCR與TRUST4兩套T細胞受體重建工具之表現。研究中透過精確率 (Precision)、召回率 (Recall) 與F1分數 (F1-score) 等分類指標，評估兩工具在不同比對容錯條件：完全比對 (Exact-match)、一處不匹配 (1-mismatch)、二處不匹配 (2-mismatch) 下的序列預測準確性。並進一步分析CDR3序列之序列特徵 (motif) 相似度、重建序列長度與工具運行時間。結果顯示，在多數情境下，TRUST4於barcode+ref模式具備較高的召回率與整體穩定性，尤其在10x Genomics平台之用藥不良反應檢體的α鏈CDR3氨基酸序列中，召回率相較MiXCR高出約20%。F1分數亦顯示，當引入容錯比對條件後，TRUST4的整體平衡性表現較佳；MiXCR則是在β鏈方面擁有較高精確率。此外，序列特徵視覺化分析指出，不同工具與平台產生的CDR3序列在結構特徵上雖有分歧，但仍有一定程度之重合。總結而言，TRUST4展現出良好的跨平台適應能力與運行效率，而MiXCR則於特定資料類型中具備較高的序列精確度。本研究提供實證數據作為未來從轉錄體定序資料重建T細胞受體庫時之工具與平台選擇的參考依據，並建議未來可擴充至更多疾病模型與整合多模態資料，以強化T細胞受體重建在臨床與研究場景中的應用潛力。

With the rapid advancement of next-generation sequencing technologies and single-cell analysis methods, single-cell transcriptome sequencing has emerged as a promising approach for reconstructing T Cell Receptor (TCR) repertoires, offering new insights into immune responses and disease mechanisms. Recently, several tools have been developed to reconstruct TCR repertoires from single-cell transcriptome data, including TRUST4 and MiXCR. However, comparative studies on TCR reconstruction across different platforms and between various stages of the same patient's disease remain relatively limited, particularly with respect to the evaluation of the Parse Biosciences platform's application. This study aims to systematically compare the performance of two TCR reconstruction tools, MiXCR and TRUST4, on single-cell data obtained from two phases of a Graves' disease patient's acute phase (the acute phase of TiA patient) and recovery phase. We utilize two single-cell platforms, 10x Genomics and Parse Biosciences, for this comparison. The study evaluates the sequence prediction accuracy of both tools using classification metrics such as Precision, Recall, and F1-score, under different alignment tolerance conditions: exact match, one mismatch, and two mismatches. In addition, we analyze the sequence feature (motif) similarity of CDR3 sequences, the reconstructed sequence lengths, and the runtime performance of the tools. The results show that, under most conditions, TRUST4 in barcode+ref mode exhibits higher recall rates and overall stability, particularly for the α-chain CDR3 amino acid sequences in the drug reaction samples of the 10x Genomics platform, where the recall rate is approximately 20% higher than that of MiXCR. The F1-score further demonstrates that, when introducing tolerance in sequence matching, TRUST4 performs better in terms of overall balance. MiXCR, on the other hand, shows higher precision in β-chain reconstructions. Additionally, sequence feature visualization analysis reveals that while there are differences in the structural features of the CDR3 sequences generated by different tools and platforms, a certain degree of overlap is still present. In summary, TRUST4 exhibits good cross-platform adaptability and operational efficiency, while MiXCR offers higher sequence accuracy in certain data types. This study provides empirical data to inform future tool and platform selection for reconstructing TCR repertoires from transcriptome sequencing data. Furthermore, we recommend extending future research to include more disease models and multimodal data integration to enhance the clinical and research applications of TCR reconstruction.