半乳糖凝集素-7標記口腔鱗狀細胞癌的惡性角質細胞狀態：三個公開隊列的單細胞整合分析 Galectin-7 marks malignant keratinocyte states in oral squamous cell carcinoma: a three-cohort single-cell meta-analysis

摘要Abstract

半乳糖凝集素-7（galectin-7，由 LGALS7 編碼）是角質細胞的標誌蛋白，但其在口腔鱗狀細胞癌（OSCC）單細胞層級的細胞分佈，以及它是否隨癌細胞分化而改變，尚未在多個獨立隊列中以同一套方法檢驗。本研究以 Seurat 重新分析三個公開 OSCC 單細胞轉錄組：Choi 等人的 GSE181919（10x，口腔子集）、Puram 等人的 GSE103322（Smart-seq2）與 Peng 等人的 GSE172577（10x）。細胞類型差異以樣本為單位進行 Wilcoxon 檢定。腫瘤上皮細胞內，LGALS7 與基底、分化、黏附、增殖、上皮–間質轉化（EMT）、侵襲及致癌訊息基因模組做 Spearman 相關，並以 slingshot 在惡性或腫瘤上皮細胞內建立由基底群集指向分化群集的擬時序。

三個隊列一致顯示，LGALS7 與同源基因 LGALS7B 集中在惡性或上皮細胞。原發腫瘤中，LGALS7 檢出率在 GSE181919 惡性細胞為 53.3%（1,376 細胞），GSE103322 為 78.2%（1,427 細胞），GSE172577 標記定義之上皮細胞為 36.9%（41,187 細胞）；纖維母細胞、T 細胞與巨噬細胞多低於 3%。LGALS1 與 LGALS3 廣泛表現，LGALS2 與 LGALS9 則偏向髓系與內皮。GSE181919 中，OSCC 相對正常上皮的檢出率差異達樣本層級顯著（p = 0.022），但正常上皮僅 3 個樣本、199 個細胞；白斑上皮（63.9%）與癌症相近（p = 0.94）。腫瘤上皮內，LGALS7 與角質細胞身分相關，而與增殖或典型 EMT 無關：GSE181919 偏向基底角蛋白與上皮黏附（rho = 0.35 與 0.38），GSE103322 同時相關於基底（rho = 0.50）、分化（rho = 0.30）與以 LAMC2 為主的侵襲模組（rho = 0.40），GSE172577 則相關於上層分化角蛋白（rho = 0.35）。擬時序僅在 GSE172577 呈現預期的基底走向分化，且 LGALS7 沿此軸上升（rho = 0.42）。GSE181919 與 GSE103322 的曲線並未構成同一條分化路徑。因此，galectin-7 是 OSCC 惡性上皮的特徵，並與角質細胞狀態共變；它與分化的關係具隊列特異性，不能概括為單一的分化軌跡。

Galectin-7, encoded by LGALS7, is a keratinocyte lectin. Whether it is restricted to malignant oral squamous cell carcinoma (OSCC) cells, and whether its level changes with cancer-cell differentiation, has not been tested with one analysis plan across independent single-cell cohorts. We reanalyzed three public OSCC scRNA-seq datasets in Seurat: GSE181919 (Choi et al.; 10x; oral-cavity subset), GSE103322 (Puram et al.; Smart-seq2), and GSE172577 (Peng et al.; 10x). Cell-type differences were tested at the sample level. Within tumor epithelium, LGALS7 was correlated with basal, differentiation, adhesion, proliferation, epithelial–mesenchymal transition (EMT), invasion, and oncogenic-signaling programs, and slingshot pseudotime was fit from the most basal cluster toward the most differentiated cluster.

Across cohorts, LGALS7 and its paralog LGALS7B were concentrated in malignant or epithelial cells. In primary tumors, LGALS7 was detected in 53.3% of malignant cells in GSE181919 (1,376 cells), 78.2% in GSE103322 (1,427 cells), and 36.9% of marker-defined epithelial cells in GSE172577 (41,187 cells). Fibroblasts, T cells, and macrophages were mostly below 3%. LGALS1 and LGALS3 were broad, whereas LGALS2 and LGALS9 favored myeloid and endothelial cells. In GSE181919, OSCC exceeded normal epithelium in percent detection (p = 0.022), but the normal comparison rests on three samples and 199 epithelial cells. Leukoplakia epithelium was similar to carcinoma (63.9%; p = 0.94). Inside tumor epithelium, LGALS7 tracked keratinocyte identity rather than proliferation or canonical EMT: basal keratins and adhesion in GSE181919 (rho = 0.35 and 0.38), basal, differentiated, and LAMC2-driven invasion programs in GSE103322 (rho = 0.50, 0.30, and 0.40), and suprabasal keratins in GSE172577 (rho = 0.35). Pseudotime followed the intended basal-to-differentiated direction only in GSE172577, where LGALS7 increased along the curve (rho = 0.42). The curves in GSE181919 and GSE103322 did not recover that same path. Galectin-7 is therefore a malignant epithelial feature of OSCC and co-varies with keratinocyte state. Its link to differentiation is cohort-specific and is not a single shared trajectory.

簡介Introduction

早期口腔鱗狀細胞癌手術後仍有相當比例復發，而現有分期不足以辨識高風險病人。Näsiaho 等人以成像質譜流式（imaging mass cytometry）分析超過 70 萬個細胞、25 種蛋白質，指出浸潤緣增殖中的淋巴內皮（KI67+ PROX1+）是獨立的不良預後指標。該研究也重新使用三個公開單細胞轉錄組來描述淋巴內皮，但蛋白質面板不含半乳糖凝集素，因此無法從該空間蛋白體資料回答 galectin-7 的問題。

A substantial fraction of early-stage oral squamous cell carcinomas relapse after surgery, and stage alone does not identify the patients at risk. Näsiaho and colleagues profiled more than 700,000 cells for 25 proteins by imaging mass cytometry and reported proliferating lymphatic endothelium (KI67+ PROX1+) at the invasive margin as an independent adverse biomarker. Their study also reanalyzed three public single-cell transcriptomes to characterize lymphatic endothelial cells. The protein panel does not include galectins, so that spatial proteome cannot address galectin-7.

Galectin-7 是角質細胞特異的原型半乳糖凝集素，LGALS7B 為其緊鄰同源基因。Choi 等人在頭頸鱗狀細胞癌的逐步癌化單細胞圖譜中，已描述 LGALS7B 陽性惡性細胞，並將其豐度與較差預後相連。Puram 等人提供帶有作者癌症細胞註解的 Smart-seq2 原發與淋巴結轉移資料，Peng 等人則提供六例 OSCC 的 10x 資料。這三個隊列技術不同、細胞註解深度不同，適合作為平行的 meta-analysis，而不適合把表現量直接合併。

Galectin-7 is a keratinocyte-restricted prototype galectin, and LGALS7B is its neighboring paralog. In a stepwise single-cell atlas of head and neck squamous cell carcinoma, Choi and colleagues described LGALS7B-positive malignant cells and linked their abundance to unfavorable outcome. Puram and colleagues released Smart-seq2 profiles of primary and lymph-node tumors with author cancer-cell labels, and Peng and colleagues released 10x profiles of six OSCC tumors. The three cohorts differ in technology and in the depth of cell annotation. They support a parallel meta-analysis. They do not support pooling expression values into one matrix.

本研究因此只使用這三個轉錄組，問兩個問題：galectin 家族在 OSCC 微環境中如何分工，以及 galectin-7 在惡性上皮內是否對應角質細胞分化，而不是增殖或 EMT。

This report uses only those three transcriptomes. It asks how the galectin family is partitioned across the OSCC microenvironment, and whether galectin-7 within malignant epithelium corresponds to keratinocyte differentiation rather than to proliferation or EMT.

假說與分析流程Hypothesis and analysis plan

分析前設定兩項假說。

Two hypotheses were specified before interpreting the results.

1. 假說一。LGALS7 與 LGALS7B 在 OSCC 惡性或上皮細胞的檢出率高於基質與免疫細胞；其他 galectin 佔有不同的細胞區室。 Hypothesis 1. LGALS7 and LGALS7B are detected more often in OSCC malignant or epithelial cells than in stromal and immune cells, while the other galectins occupy different compartments.
2. 假說二。在腫瘤上皮細胞內部，LGALS7 隨基底走向分化的角質細胞狀態上升，並與增殖及 EMT 呈負相關或無相關。 Hypothesis 2. Within tumor epithelial cells, LGALS7 rises along a basal-to-differentiated keratinocyte state and is negatively related, or unrelated, to proliferation and EMT.

擬時序是腫瘤上皮內部的表現狀態排序，不是正常黏膜走向癌症的時間軸，也不是實驗量到的時鐘。 Pseudotime is an ordering of expression states inside tumor epithelium. It is not a normal-to-cancer timeline and it is not a measured clock.

結果Results

三個隊列的組成Cohort composition

GSE181919 取口腔（subsite OC）：10 個原發癌（10,397 細胞，其中作者標註惡性細胞 1,376）、3 個正常黏膜（7,232 細胞，上皮 199）與 4 個白斑（6,527 細胞，上皮 183），合計 24,156 細胞。GSE103322 使用作者已過濾的 5,902 細胞，其中原發惡性細胞 1,427、淋巴結惡性細胞 788。GSE172577 六個腫瘤在細胞過濾後保留 57,907 / 65,255 細胞，標記定義之上皮細胞 41,187。三個物件分開分析，因為 Smart-seq2 的 log2(TPM/10+1) 與 10x 的 log1p 正規化計數不能比較絕對表現量。跨隊列比較使用檢出率。

The oral-cavity subset of GSE181919 comprised 10 primary tumors (10,397 cells, 1,376 author-labeled malignant cells), 3 normal mucosae (7,232 cells, 199 epithelial), and 4 leukoplakias (6,527 cells, 183 epithelial), 24,156 cells in total. GSE103322 contributed the author-filtered 5,902 cells, including 1,427 primary and 788 lymph-node malignant cells. GSE172577 retained 57,907 of 65,255 cells after filtering, of which 41,187 were marker-defined epithelium. The three objects were analyzed separately. Absolute expression cannot be compared between Smart-seq2 log2(TPM/10+1) and 10x log1p-normalized counts. Cross-cohort comparisons use detection rate.

Galectin 家族佔據不同的細胞區室Galectin genes occupy different compartments

原發腫瘤中，LGALS1 與 LGALS3 幾乎遍布基質、免疫與上皮。GSE181919 纖維母細胞的 LGALS1 檢出率為 100%，巨噬細胞的 LGALS3 為 94%。LGALS2 與 LGALS9 偏向髓系與內皮：同一隊列中，LGALS2 在巨噬細胞為 61%、在惡性細胞為 5%；LGALS9 在巨噬細胞為 70%、內皮為 63%、惡性細胞為 18%。LGALS4 很少被檢出。LGALS7 與 LGALS7B 呈現相反的模式，集中在惡性或上皮細胞（GSE181919 惡性細胞分別為 53% 與 72%；GSE103322 為 78% 與 66%；GSE172577 上皮為 37% 與 65%）。假說一得到支持。

In primary tumors, LGALS1 and LGALS3 were detected across stroma, immune cells, and epithelium. Fibroblast LGALS1 detection was 100% in GSE181919, and macrophage LGALS3 detection was 94%. LGALS2 and LGALS9 favored myeloid and endothelial cells: in the same cohort, LGALS2 was detected in 61% of macrophages and 5% of malignant cells, and LGALS9 in 70% of macrophages, 63% of endothelial cells, and 18% of malignant cells. LGALS4 was rarely detected. LGALS7 and LGALS7B showed the opposite pattern and were concentrated in malignant or epithelial cells (53% and 72% of GSE181919 malignant cells; 78% and 66% in GSE103322; 37% and 65% of GSE172577 epithelial cells). Hypothesis 1 is supported.

Galectin-7 是惡性上皮的標記Galectin-7 is a marker of malignant epithelium

在每一個原發腫瘤隊列，LGALS7 的檢出率高峰都落在惡性或上皮細胞，基質與淋巴細胞多在 3% 以下（圖 2）。GSE103322 未分類細胞有 27.9% 檢出，這組細胞沒有作者的明確類型，不納入細胞類型推論。GSE172577 內皮細胞為 8.7%，仍明顯低於上皮的 36.9%。

In every primary-tumor cohort, LGALS7 detection peaked in malignant or epithelial cells. Stromal and lymphoid detection was mostly below 3% (Figure 2). Unassigned cells in GSE103322 reached 27.9%; they lack an author cell type and were not used for cell-type inference. Endothelial detection in GSE172577 was 8.7%, still well below the epithelial rate of 36.9%.

樣本層級檢定把每個樣本先彙總成一個數值，避免把同一腫瘤的細胞當成獨立觀察。GSE181919 配對比較中，惡性細胞的檢出率與平均表現高於 T 細胞（10 對，p = 0.0059）、B/漿細胞（8 對，p = 0.014）、纖維母細胞（8 對，p = 0.014）、巨噬細胞（9 對，p = 0.009）與樹突細胞（6 對，p = 0.036）。相對內皮細胞的差異未達 0.05（5 對，p = 0.059）。GSE103322 原發惡性細胞高於纖維母細胞（11 對，p = 0.0039）與 T 細胞（6 對，p = 0.036）。原發與淋巴結惡性細胞的樣本中位數檢出率分別約 86% 與 84%（p = 0.94），平均表現亦無差異（p = 1.0）。GSE172577 上皮高於纖維母細胞、T 細胞與髓系（各 6 對，p = 0.036）。相對肥大細胞（p = 0.10）、內皮（p = 0.059）、周細胞（p = 0.059）與 B/漿細胞（p = 0.18）未達 0.05，其中 B/漿細胞只有 3 對。

Sample-level tests collapse each sample to one value, so cells from the same tumor are not treated as independent. In paired GSE181919 tumors, malignant cells exceeded T cells (10 pairs, p = 0.0059), B/plasma cells (8 pairs, p = 0.014), fibroblasts (8 pairs, p = 0.014), macrophages (9 pairs, p = 0.009), and dendritic cells (6 pairs, p = 0.036) in both detection and mean expression. The endothelial comparison did not reach 0.05 (5 pairs, p = 0.059). In GSE103322, primary malignant cells exceeded fibroblasts (11 pairs, p = 0.0039) and T cells (6 pairs, p = 0.036). Primary and lymph-node malignant cells had similar sample-median detection (about 86% and 84%, p = 0.94) and similar mean expression (p = 1.0). In GSE172577, epithelium exceeded fibroblasts, T cells, and myeloid cells (6 pairs each, p = 0.036). Comparisons with mast cells (p = 0.10), endothelium (p = 0.059), pericytes (p = 0.059), and B/plasma cells (p = 0.18) did not reach 0.05; the B/plasma test had only three pairs.

GSE181919 是唯一同時有正常黏膜與白斑的隊列（圖 3）。OSCC 惡性細胞檢出率 53.3%，正常上皮 5.5%，樣本中位數分別為 70% 與 0%（檢出率 p = 0.022；平均表現 p = 0.11）。正常組只有 3 個樣本、199 個上皮細胞，這個差異不能解讀成正常黏膜沒有 galectin-7。白斑上皮為 63.9%，與 OSCC 無差異（檢出率 p = 0.94，平均表現 p = 0.62）。Galectin-7 在此隊列標示的是已改變的鱗狀上皮，包含白斑，而不只是浸潤癌相對正常的對比。

GSE181919 is the only cohort with matched normal mucosa and leukoplakia (Figure 3). Detection was 53.3% in OSCC malignant cells and 5.5% in normal epithelium, with sample medians of 70% and 0% (detection p = 0.022; mean expression p = 0.11). The normal group has three samples and 199 epithelial cells, so this difference should not be read as absence of galectin-7 from normal mucosa. Leukoplakia epithelium was 63.9% and did not differ from OSCC (detection p = 0.94; mean expression p = 0.62). In this cohort, galectin-7 marks altered squamous epithelium, including leukoplakia, rather than a contrast that is unique to invasive carcinoma versus normal tissue.

與癌症相關基因的關聯Association with cancer-related genes

相關分析限定在腫瘤上皮或惡性細胞，使用 data layer 的 Spearman rho，並在每個隊列與子集內做 Benjamini–Hochberg 校正。細胞數很大時 p 值會極小，因此解讀以 rho 的大小與跨隊列是否重複為準（圖 4）。

Correlations were restricted to tumor epithelial or malignant cells, used Spearman rho on the data layer, and were Benjamini–Hochberg adjusted within each cohort and subset. With large cell numbers, p values become very small. Interpretation therefore follows the magnitude of rho and whether the sign recurs across cohorts (Figure 4).

GSE181919 惡性細胞（n = 1,376）中，LGALS7 與基底角蛋白模組相關（rho = 0.35；KRT5 0.30、KRT14 0.24、TP63 0.30），也與上皮黏附模組相關（rho = 0.38；PKP1 0.45、JUP 0.41、DSG3 0.27）。上層分化模組只有 rho = 0.07。增殖為負相關（rho = −0.13），在 LGALS7 已檢出的 733 個細胞中更明確（rho = −0.30）。EMT 模組接近 0（rho = 0.01），侵襲模組略負（rho = −0.08）。訊息基因中，CCND1 為正（rho = 0.33），CDKN2A 與 MYC 為負（rho = −0.24 與 −0.12）。EPCAM 為負（rho = −0.27），所以黏附訊號來自橋粒與密連蛋白，而不是 EPCAM 本身。

In GSE181919 malignant cells (n = 1,376), LGALS7 correlated with the basal keratin program (rho = 0.35; KRT5 0.30, KRT14 0.24, TP63 0.30) and with epithelial adhesion (rho = 0.38; PKP1 0.45, JUP 0.41, DSG3 0.27). The suprabasal differentiation program was rho = 0.07. Proliferation was negative (rho = −0.13), and more clearly so among the 733 LGALS7-detected cells (rho = −0.30). The EMT module was near zero (rho = 0.01) and the invasion module was weakly negative (rho = −0.08). Among signaling genes, CCND1 was positive (rho = 0.33), whereas CDKN2A and MYC were negative (rho = −0.24 and −0.12). EPCAM was negative (rho = −0.27), so the adhesion signal comes from desmosomal and plaque genes rather than from EPCAM itself.

GSE103322 惡性細胞（n = 2,215，含淋巴結）的基底相關更強（rho = 0.50；KRT14 0.63、COL17A1 0.50、KRT5 0.48），分化模組亦為正（rho = 0.30；KRT13 0.31）。侵襲模組 rho = 0.40，主要由 LAMC2（0.41）、PTHLH（0.21）與 MMP10（0.17）貢獻。EGFR rho = 0.25。增殖模組接近 0（rho = −0.02）。黏附模組整體接近 0，因為 EPCAM 為 −0.44 而 DSG3 為 +0.40。EMT 模組僅 rho = 0.13，而且內部方向不一致：SNAI2 為正，ZEB1 與 FN1 為負。

In GSE103322 malignant cells (n = 2,215, including lymph-node tumor cells), the basal association was stronger (rho = 0.50; KRT14 0.63, COL17A1 0.50, KRT5 0.48) and the differentiation program was also positive (rho = 0.30; KRT13 0.31). The invasion module was rho = 0.40, led by LAMC2 (0.41), PTHLH (0.21), and MMP10 (0.17). EGFR was rho = 0.25. Proliferation was near zero (rho = −0.02). The adhesion module as a whole was near zero because EPCAM was −0.44 while DSG3 was +0.40. The EMT module was only rho = 0.13 and was internally inconsistent: SNAI2 was positive, whereas ZEB1 and FN1 were negative.

GSE172577 標記定義之上皮（n = 41,187）呈現另一種格局：分化模組 rho = 0.35（KRT13 0.44、SPINK5 0.38、KRT4 0.31、IVL 0.29），基底模組整體只有 rho = 0.03，雖然 KRT5 個別為 0.32。侵襲模組為負（rho = −0.28），EMT 略負（rho = −0.08），致癌訊息模組略正（rho = 0.16；MYC 0.26、FOS 0.23）。在 LGALS7 已檢出的 15,187 個細胞中，分化相關降到 rho = 0.03，表示全細胞的正相關有一部分來自 LGALS7 與分化基因一起被檢出。LGALS7B 仍是最接近的同源基因（全細胞 rho = 0.51，LGALS7 陽性細胞 rho = 0.67；另外兩個隊列的惡性細胞為 0.22 與 0.32）。

Marker-defined epithelium in GSE172577 (n = 41,187) showed a different pattern: the differentiation program was rho = 0.35 (KRT13 0.44, SPINK5 0.38, KRT4 0.31, IVL 0.29), whereas the basal module as a whole was rho = 0.03 even though KRT5 alone was 0.32. Invasion was negative (rho = −0.28), EMT weakly negative (rho = −0.08), and oncogenic signaling weakly positive (rho = 0.16; MYC 0.26, FOS 0.23). Among 15,187 LGALS7-detected cells, the differentiation correlation fell to rho = 0.03, so part of the all-cell association reflects joint detection of LGALS7 and differentiation genes. LGALS7B remained the closest paralog (rho = 0.51 in all epithelial cells and 0.67 in LGALS7-detected cells; rho = 0.22 and 0.32 in the malignant cells of the other two cohorts).

假說二只得到部分支持。三個隊列都否定「LGALS7 是增殖或典型 EMT 標記」。與分化的方向則不一致：Choi 偏向基底與黏附，Puram 同時具有基底、分化與侵襲角蛋白，Peng 偏向層分化。沒有一個癌症程式在三個隊列以相同方向且 |rho| ≥ 0.3 重複，除了 LGALS7 本身的上皮富集，以及與 KRT5 的正向秩相關（0.30、0.48、0.32）。

Hypothesis 2 is only partly supported. All three cohorts argue against LGALS7 as a proliferation or canonical EMT marker. The differentiation result is not shared: Choi aligns with basal and adhesion programs, Puram with basal, differentiated, and invasive keratin programs, and Peng with suprabasal differentiation. No cancer program recurs in the same direction with |rho| at least 0.3 in all three cohorts, apart from the epithelial enrichment of LGALS7 itself and a positive rank correlation with KRT5 (0.30, 0.48, and 0.32).

擬時序與角質細胞分化Pseudotime and keratinocyte differentiation

Slingshot 只套在腫瘤上皮或惡性細胞。每個隊列先重新分群，再把基底分數最高的群集當起點、分化分數最高的群集當終點。曲線是否真的由基底走向分化，用擬時序與兩個分數的相關事後檢查，而不是用起點標籤本身認定。

Slingshot was fit only inside tumor epithelial or malignant cells. Each cohort was reclustered. The cluster with the highest basal-minus-differentiated score was the start, and the cluster with the lowest score was the end. Whether the curve actually ran from basal toward differentiated cells was checked afterwards by correlating pseudotime with the two scores. The start label alone was not taken as proof of direction.

GSE172577 在降採樣後有 3,580 個細胞落在所選曲線上。基底分數隨擬時序下降（rho = −0.46），分化分數上升（rho = 0.58），方向符合假說。LGALS7 沿曲線上升（rho = 0.42），與分化分數只有中等相關（rho = 0.18），與基底分數無關（rho = 0.02）。圖 5 的黑色折線是分箱平均，右端被大量 LGALS7 未檢出的細胞拉低，因此折線不是單調上升；秩相關仍然為正。

After downsampling, 3,580 GSE172577 cells lay on the selected curve. The basal score fell along pseudotime (rho = −0.46) and the differentiated score rose (rho = 0.58), which matches the intended direction. LGALS7 increased along the curve (rho = 0.42), with a modest correlation to the differentiated score (rho = 0.18) and none to the basal score (rho = 0.02). The black line in Figure 5 is a binned mean. Its right-hand end is pulled down by many LGALS7-undetected cells, so the line is not monotonic even though the rank correlation is positive.

GSE181919 有 663 / 1,376 個惡性細胞落在曲線上，而且 slingshot 找到 6 條曲線，採用涵蓋細胞最多的一條。這條曲線的方向朝向基底：基底分數 rho = 0.47，分化分數 rho = −0.41。LGALS7 隨之上升（rho = 0.59），並與基底分數相關（rho = 0.47）。GSE103322 有 1,443 個細胞落在曲線上，基底分數（rho = −0.71）與分化分數（rho = −0.34）都沿擬時序下降，因此它不是一條乾淨的基底到上層路徑。LGALS7 沿曲線下降（rho = −0.59），但仍與基底分數正相關（rho = 0.51）、與分化分數中等正相關（rho = 0.36）。

In GSE181919, 663 of 1,376 malignant cells lay on the curve. Slingshot returned six curves, and the curve covering the most cells was kept. That curve ran toward the basal program (basal rho = 0.47; differentiated rho = −0.41). LGALS7 rose with it (rho = 0.59) and with the basal score (rho = 0.47). In GSE103322, 1,443 cells lay on the curve. Both the basal score (rho = −0.71) and the differentiated score (rho = −0.34) declined along pseudotime, so the path is not a clean basal-to-suprabasal axis. LGALS7 declined along the curve (rho = −0.59) while remaining positively correlated with the basal score (rho = 0.51) and, more moderately, with the differentiated score (rho = 0.36).

因此，擬時序支持「LGALS7 隨基底走向分化而上升」只發生在 Peng 的標記定義上皮。Choi 的曲線把 LGALS7 放在較基底的一端，Puram 的曲線則沒有分離出單一分化軸。這與相關分析的隊列差異一致，也說明不能把三個資料集的擬時序平均成一個分化結論。

Pseudotime therefore supports a rise of LGALS7 from basal toward differentiated states only in the marker-defined Peng epithelium. The Choi curve places LGALS7 toward the basal end, and the Puram curve does not isolate a single differentiation axis. This matches the cohort differences in the correlation analysis. The three pseudotime results should not be averaged into one differentiation claim.

材料與方法Materials and methods

資料Data

GSE181919 使用 GEO 的 UMI 計數矩陣與條碼後設資料。分析保留 subsite 為 OC 的樣本，並保留白斑。條碼在計數矩陣使用句點、在後設資料使用連字號，合併前先對齊。組織類型沿用作者欄位：CA 為原發腫瘤、NL 為正常、LP 為白斑。惡性細胞只出現在原發腫瘤，正常與白斑的上皮使用作者的 Epithelial 標籤。GSE103322 使用 GSE103322_HNSCC_all_data.txt.gz。檔案前五列是後設資料而不是基因；基因符號帶有單引號。數值已是 log2(TPM/10+1)，因此不再做 NormalizeData。counts slot 以 (2^x − 1) × 10 還原非零項，零保持為零，檢出定義為 counts > 0。作者癌症標記為 1 的細胞視為惡性。非癌類型沿用矩陣中的粗分類，並去掉前置連字號。GSE172577 由六個 10x 樣本（SYSMH1–SYSMH6）的 barcodes、features 與 matrix 讀入。細胞過濾只做在這個隊列：每個細胞 200–7,500 個基因，粒線體比例 < 20%，保留 57,907 / 65,255 細胞。此隊列沒有作者細胞類型。

GSE181919 used the GEO UMI count matrix and barcode metadata. Samples with subsite OC were kept, including leukoplakia. Count barcodes use dots and metadata barcodes use dashes; the two were aligned before joining. Author tissue labels were kept: CA primary tumor, NL normal, and LP leukoplakia. Malignant labels occur only in primary tumor. Normal and leukoplakia epithelium uses the author Epithelial label. GSE103322 used GSE103322_HNSCC_all_data.txt.gz. The first five data rows are metadata rather than genes, and gene symbols are single-quoted. Values are already log2(TPM/10+1), so NormalizeData was not run again. The counts slot stores (2^x − 1) × 10 for nonzero entries and leaves zeros at zero. Detection is counts > 0. Cells with the author cancer flag equal to 1 were called malignant. Non-cancer labels use the coarse classes in the matrix, with a leading hyphen removed. GSE172577 was read from the barcodes, features, and matrix files of six 10x tumors (SYSMH1–SYSMH6). Cell filtering was applied only here: 200–7,500 genes per cell and mitochondrial fraction below 20%, retaining 57,907 of 65,255 cells. This cohort has no author cell types.

專案資料夾中的成像質譜流式矩陣（約 708,261 細胞、27 個通道）沒有 galectin 蛋白，不納入本分析。decontX 環境 RNA 校正在本次執行中關閉。三個隊列沒有整合成單一物件；harmony 未使用，降維為各資料集內部的 PCA。

The imaging-mass-cytometry matrix in the project folder, about 708,261 cells by 27 channels, contains no galectin protein and was not used. Ambient-RNA correction with decontX was left off. The cohorts were not integrated into one object. Harmony was not used. Dimension reduction was PCA within each dataset.

細胞類型Cell types

GSE181919 與 GSE103322 的主要結論使用作者細胞類型。GSE172577 以模組分數指定譜系，上皮標記為 KRT5、KRT14、KRT6A、KRT17 與 EPCAM，故意不放入 LGALS7，避免用目標基因定義它所屬的細胞。分數低於 0 的細胞標為 Unassigned。Peng 的上皮狀態記為「標記定義之腫瘤上皮」，不是拷貝數確認的惡性細胞。

The main conclusions for GSE181919 and GSE103322 use author cell types. GSE172577 lineages were assigned by module score. The epithelial module was KRT5, KRT14, KRT6A, KRT17, and EPCAM. LGALS7 was deliberately omitted so that the gene under study did not define its own cell class. Cells with a score below zero were labeled Unassigned. The Peng epithelial state is recorded as marker-defined tumor epithelium, not as copy-number-confirmed malignancy.

Galectin 表現與樣本層級檢定Galectin expression and sample-level tests

檢視的基因為 LGALS1、LGALS2、LGALS3、LGALS4、LGALS7、LGALS7B、LGALS8、LGALS9、LGALS9B、LGALS9C、LGALS12、LGALS13、LGALS14 與 LGALS16。報告圖集中在有穩定檢出的前八個。檢出定義為 counts slot 大於 0。平均表現取 data layer。每個樣本、組織與細胞類型先彙總，細胞類型比較要求該類型在樣本中至少 20 個細胞，且至少 3 個樣本。未配對比較用 Wilcoxon rank-sum，同一原發腫瘤內的惡性或上皮細胞對其他類型用 Wilcoxon signed-rank。不把不同平台的平均表現放在一起比較。

Genes surveyed were LGALS1, LGALS2, LGALS3, LGALS4, LGALS7, LGALS7B, LGALS8, LGALS9, LGALS9B, LGALS9C, LGALS12, LGALS13, LGALS14, and LGALS16. The figures focus on the eight genes with stable detection. A gene was detected when the counts slot was greater than zero. Mean expression was taken from the data layer. Values were summarized per sample, tissue, and cell type. A cell-type comparison required at least 20 cells of that type in the sample and at least three samples. Unpaired comparisons used a Wilcoxon rank-sum test. Within a primary tumor, malignant or epithelial cells were compared with other types by a Wilcoxon signed-rank test. Mean expression was not compared across platforms.

基因程式與擬時序Gene programs and pseudotime

程式分數是所列基因在 data layer 的平均。基底程式為 KRT15、KRT5、KRT14、TP63、COL17A1、ITGA6。分化程式為 KRT13、KRT4、IVL、SPRR1B、SPINK5、KRT1。黏附為 CDH1、EPCAM、DSG3、PKP1、JUP。增殖為 MKI67、TOP2A、PCNA、TYMS、BIRC5、CDK1。EMT 為 VIM、SNAI1、SNAI2、ZEB1、ZEB2、CDH2、FN1、TWIST1。侵襲為 MMP1、MMP9、MMP10、LAMC2、PTHLH、CXCL8。致癌訊息為 EGFR、MYC、CCND1、CDKN2A、STAT3、JUN、FOS。Spearman 相關以 cor.test 計算，exact = FALSE，並在每個資料集與子集（全部腫瘤上皮，以及 LGALS7 檢出細胞若至少 100 個）內對檢定做 BH 校正。細胞少於 30，或任一變數的唯一值少於 3，則不計算。

A program score is the mean of the listed genes on the data layer. Basal genes were KRT15, KRT5, KRT14, TP63, COL17A1, and ITGA6. Differentiation genes were KRT13, KRT4, IVL, SPRR1B, SPINK5, and KRT1. Adhesion genes were CDH1, EPCAM, DSG3, PKP1, and JUP. Proliferation genes were MKI67, TOP2A, PCNA, TYMS, BIRC5, and CDK1. EMT genes were VIM, SNAI1, SNAI2, ZEB1, ZEB2, CDH2, FN1, and TWIST1. Invasion genes were MMP1, MMP9, MMP10, LAMC2, PTHLH, and CXCL8. Oncogenic-signaling genes were EGFR, MYC, CCND1, CDKN2A, STAT3, JUN, and FOS. Spearman correlations used cor.test with exact = FALSE. Benjamini–Hochberg adjustment was applied inside each dataset and subset: all tumor epithelial cells, and LGALS7-detected cells when at least 100 were available. A test was skipped when fewer than 30 cells remained or when either variable had fewer than three unique values.

擬時序子集為上皮狀態屬於原發惡性、淋巴結惡性或標記定義腫瘤上皮的細胞。GSE172577 在軌跡分析前隨機降至最多 6,000 細胞（seed 7）；相關分析仍用全部腫瘤上皮。已是 log 尺度的資料用 dispersion 找高變基因，計數資料用 vst。PCA 取 20 個主成分，鄰居使用前 10 個，分群解析度 0.4。每個群集計算基底基因平均減去分化基因平均，至少 25 個細胞的群集才列入；分數最高者為起點，最低者為終點。Slingshot 在 PCA 上擬合。若指定終點失敗，則只保留起點再擬合。多條曲線時保留有限擬時序細胞數最多的一條。方向檢查是這條擬時序對基底分數與分化分數的 Spearman 相關。

The trajectory subset was cells whose epithelial state was primary malignant, lymph-node malignant, or marker-defined tumor epithelium. GSE172577 was randomly reduced to at most 6,000 cells before the trajectory (seed 7). Correlations still used every tumor epithelial cell. Highly variable genes used dispersion for data that were already on a log scale and vst for counts. PCA retained 20 components, neighbors used the first 10, and clustering resolution was 0.4. Each cluster was scored as the mean of basal genes minus the mean of differentiation genes. Clusters with at least 25 cells were eligible. The highest score was the start and the lowest score was the end. Slingshot was fit on PCA. If the end cluster failed, the fit was repeated from the start cluster only. When several curves were returned, the curve with the most finite pseudotime values was kept. Direction was checked by Spearman correlation of that pseudotime with the basal and differentiated scores.

軟體Software

分析使用 R 4.6.0、Seurat 5.5.0、SeuratObject 5.4.0、slingshot 與其依賴 SingleCellExperiment。圖表以 Python 3 與 Matplotlib 3.10 由已輸出的統計表重繪。腳本為 analysis/galectin7_oscc_scrna_seurat.R、analysis/galectin7_trajectory_cancer_genes.R 與 analysis/make_report_figures.py。隨機程序的種子在擬時序腳本設為 7。

Analyses used R 4.6.0, Seurat 5.5.0, SeuratObject 5.4.0, and slingshot with SingleCellExperiment. Figures were redrawn in Python 3 with Matplotlib 3.10 from the exported statistics. Scripts are analysis/galectin7_oscc_scrna_seurat.R, analysis/galectin7_trajectory_cancer_genes.R, and analysis/make_report_figures.py. The pseudotime script set the random seed to 7.

結論Conclusion

三個獨立 OSCC 單細胞隊列給出同一個細胞分佈結論：galectin-7 與 LGALS7B 屬於惡性或上皮細胞，而 LGALS1、LGALS2、LGALS3 與 LGALS9 屬於更廣或不同的基質與免疫區室。Galectin-7 在腫瘤上皮內對應角質細胞身分，而不是增殖或典型 EMT。這個身分在 Choi 隊列偏基底與細胞黏附，在 Puram 隊列同時包含基底、分化與 LAMC2 相關的侵襲角蛋白，在 Peng 隊列偏上層分化。只有 Peng 隊列的擬時序走出預期的基底到分化方向，並伴隨 LGALS7 上升。正常黏膜的對照樣本太少，白斑上皮的檢出率已與癌症相近。後續若要主張 galectin-7 參與癌細胞分化，需要在同一技術、具拷貝數確認的惡性細胞，以及具層級位置資訊的資料中重複這條軌跡。

Three independent OSCC single-cell cohorts give the same compartment result. Galectin-7 and LGALS7B belong to malignant or epithelial cells, whereas LGALS1, LGALS2, LGALS3, and LGALS9 belong to broader or different stromal and immune compartments. Within tumor epithelium, galectin-7 corresponds to keratinocyte identity rather than to proliferation or canonical EMT. That identity is basal and adhesive in the Choi cohort, basal plus differentiated and LAMC2-linked invasion keratins in the Puram cohort, and suprabasal in the Peng cohort. Only the Peng pseudotime runs in the expected basal-to-differentiated direction, and LGALS7 rises along it. The normal-mucosa comparison is small, and leukoplakia epithelium is already similar to carcinoma. A claim that galectin-7 participates in cancer-cell differentiation needs the trajectory to be repeated in one technology, in copy-number-confirmed malignant cells, and in data that retain epithelial-layer position.

參考文獻References

1. Näsiaho J, Nissi L, Ventelä S, Irjala H, Salmi M. Spatial single-cell analysis reveals tumor microenvironment signatures predictive of oral cavity cancer outcome. Cell Rep Med. 2026;7(2):102615. doi:10.1016/j.xcrm.2026.102615.
2. Choi JH, Lee BS, Jang JY, Lee YS, et al. Single-cell transcriptome profiling of the stepwise progression of head and neck cancer. Nat Commun. 2023;14:1055. doi:10.1038/s41467-023-36691-x.
3. Puram SV, Tirosh I, Parikh AS, Patel AP, Yizhak K, Gillespie S, et al. Single-cell transcriptomic analysis of primary and metastatic tumor ecosystems in head and neck cancer. Cell. 2017;171(7):1611-1624.e24. doi:10.1016/j.cell.2017.10.044.
4. Peng Y, Xiao L, Rong H, Ou Z, Cai T, Liu N, et al. Single-cell profiling of tumor-infiltrating TCF1/TCF7+ T cells reveals a T lymphocyte subset associated with tertiary lymphoid structures/organs and a superior prognosis in oral cancer. Oral Oncol. 2021;119:105348. doi:10.1016/j.oraloncology.2021.105348.
5. Magnaldo T, Bernerd F, Darmon M. Galectin-7, a human 14-kDa S-lectin, specifically expressed in keratinocytes and sensitive to retinoic acid. Dev Biol. 1995;168(2):259-271. doi:10.1006/dbio.1995.1078.
6. Hao Y, Stuart T, Kowalski MH, Choudhary S, Hoffman P, Hartman A, et al. Dictionary learning for integrative, multimodal and scalable single-cell analysis. Nat Biotechnol. 2024;42(2):293-304. doi:10.1038/s41587-023-01767-y.
7. Street K, Risso D, Fletcher RB, Das D, Ngai J, Yosef N, et al. Slingshot: cell lineage and pseudotime inference for single-cell transcriptomics. BMC Genomics. 2018;19:477. doi:10.1186/s12864-018-4772-0.
8. Benjamini Y, Hochberg Y. Controlling the false discovery rate: a practical and powerful approach to multiple testing. J R Stat Soc Series B Stat Methodol. 1995;57(1):289-300.

資料與程式。計數與後設資料來自 GEO 登錄號 GSE181919、GSE103322 與 GSE172577。樣本層級檢定、galectin 彙總與擬時序相關表位於 results/scrna_galectin7/。本頁圖檔位於 results/scrna_galectin7/report/figures/。

Data and code. Counts and metadata are the GEO accessions GSE181919, GSE103322, and GSE172577. Sample-level tests, galectin summaries, and pseudotime correlation tables are in results/scrna_galectin7/. Figures for this page are in results/scrna_galectin7/report/figures/.