scMAC：用大语言模型增强单细胞功能先验，补全bulk TCR库的功能语境
scMAC: large language model-enhanced single-cell functional priors for completing functional context in bulk T cell receptor repertoires
scMAC 是一个以单细胞数据为引导的功能语境补全框架，用于从 bulk TCR 测序数据推断反映转录组状态的 T 细胞功能表征。该方法先用多尺度表达建模结合大语言模型对基因和蛋白描述的编码构建去噪的功能参考，再学习原型引导的 TCR 到功能映射，把配对细胞水平的潜变量对齐与功能原型正则化结合，从而在组库水平估计功能组成。
来源：Briefings in Bioinformatics · doi.org
Bulk T cell receptor (TCR) sequencing enables scalable immune repertoire profiling but does not measure the transcriptional states required to infer T cell functional programs. Paired single-cell RNA/TCR sequencing offers a single-cell functional prior by jointly observing TCR identity and transcriptomic states. However, transferring this prior to bulk repertoires is challenging: sparse and noisy single-cell profiles can distort the functional reference, and the probabilistic relationship between TCR identity and transcriptome-inferred functional state makes direct cell-level alignment unstable, potentially biasing repertoire-level functional composition estimation. We present scMAC, a single-cell-guided framework for functional context completion that infers TCR functional representations reflecting transcriptome-derived T-cell states. scMAC builds a denoised functional reference by integrating multi-scale expression modeling with gene-semantic priors derived from large language model encodings of gene and protein descriptions, improving gene-level functional coherence. It then learns a prototype-guided TCR-to-function mapping that combines paired cell-level latent alignment with functional prototype regularization, enabling robust repertoire-level functional composition estimation. Across real single-cell and paired RNA/TCR datasets, scMAC outperformed existing single-cell representation methods in fine-grained cell-state representation, gene functional consistency, expression recovery, and disease-relevant pathway analysis, achieving an adjusted Rand index of 0.656 on a colorectal cancer dataset with 83 cell subtypes. For repertoire-level functional composition estimation, scMAC closely matched single-cell-derived functional proportions, with an R2 of 0.979 and a Bray–Curtis distance of 0.075. These results support scMAC as a practical route for transferring single-cell-derived functional knowledge to bulk TCR repertoire analysis beyond clonotype identity and clonal abundance. scMAC is publicly available at https://github.com/Bioinformatics7181/scMAC.