蛋白质组学与代谢组学综合范例

本章通过五个完整的实战案例,展示蛋白质组学和代谢组学在生物医学研究中的典型应用场景。每个案例都包含从实验设计到数据分析再到生物学解释的完整流程,并提供可运行的代码示例,帮助你将理论知识转化为实际分析能力。

范例一:乳腺癌细胞药物处理的蛋白质组与磷酸化组学联合分析

研究背景

雌激素受体(ER)阳性乳腺癌约占所有乳腺癌病例的70%,内分泌治疗(如他莫昔芬、氟维司群等ER拮抗剂)是这类患者的标准治疗方案。然而,耐药性的产生是临床面临的主要挑战。理解ER拮抗剂作用后的分子响应机制——特别是信号通路的动态变化——对于发现新的治疗靶点和克服耐药具有重要意义。本案例使用TMT标记定量蛋白质组和磷酸化富集质谱技术,系统性地揭示MCF7乳腺癌细胞在ER拮抗剂处理后0h、6h、24h时间点的蛋白质表达和翻译后修饰变化。

实验设计要点

  • 细胞系:MCF7(ER阳性乳腺癌细胞系)

  • 处理条件:ER拮抗剂(氟维司群 100 nM)处理 0h、6h、24h

  • 生物学重复:每组3个独立培养孔

  • 定量策略:10-plex TMT标记(3×3个样本 + 1个Pool对照)

  • 磷酸化富集:TiO2 + IMAC 联合富集

  • 质谱平台:Orbitrap Fusion Lumos, HCD-MS3(减少比率压缩)

# ========================================
# 范例一:乳腺癌TMT磷酸化组学完整分析流程
# ========================================

# ---- 第1步: 数据读取与预处理 ----
library(limma)
library(DEP)
library(MSstatsTMT)

set.seed(2024)
n_proteins <- 3000
n_phospho <- 1500
samples <- paste0(rep(c("Ctrl_0h", "Treat_6h", "Treat_24h"), each = 3),
                  rep(1:3, 3))

# 模拟蛋白质定量矩阵 (log2 TMT强度)
prot_matrix <- matrix(rnorm(n_proteins * 9, mean = 20, sd = 1.5),
                       nrow = n_proteins, dimnames = list(
                           paste0("Protein_", 1:n_proteins), samples))

# 模拟差异蛋白 (6h和24h处理效应)
early_diff_idx <- sample(1:n_proteins, 80)   # 早期响应
late_diff_idx <- sample(setdiff(1:n_proteins, early_diff_idx), 120)  # 晚期响应

prot_matrix[early_diff_idx, 4:6] <- prot_matrix[early_diff_idx, 4:6] - 1.8
prot_matrix[late_diff_idx, 7:9] <- prot_matrix[late_diff_idx, 7:9] - 2.2

cat(sprintf("蛋白质矩阵维度: %d × %d\n", nrow(prot_matrix), ncol(prot_matrix)))

# 模拟磷酸化位点定量矩阵
phos_matrix <- matrix(rnorm(n_phospho * 9, mean = 15, sd = 2),
                       nrow = n_phospho, dimnames = list(
                           paste0("Phospho_", 1:n_phospho), samples))

# 磷酸化变化通常比总蛋白变化更剧烈
phos_early <- sample(1:n_phospho, 200)
phos_late <- sample(setdiff(1:n_phospho, phos_early), 300)

phos_matrix[phos_early, 4:6] <- phos_matrix[phos_early, 4:6] - 2.5
phos_matrix[phos_late, 7:9] <- phos_matrix[phos_late, 7:9] - 3.0

# ---- 第2步: 差异分析 (limma) ----
design <- model.matrix(~0 + factor(samples,
    levels = c("Ctrl_01","Ctrl_02","Ctrl_03",
               "Treat_04","Treat_05","Treat_06",
               "Treat_07","Treat_08","Treat_09")))
colnames(design) <- c(paste0("C",1:3), paste0("T6",1:3), paste0("T24",1:3))

# 构建比较矩阵: Treat vs Control (合并时间点或分别比较)
group_mean <- makeContrasts(
    T6_vs_Ctrl = (T61+T62+T63)/3 - (C1+C2+C3)/3,
    T24_vs_Ctrl = (T241+T242+T243)/3 - (C1+C2+C3)/3,
    T24_vs_T6 = (T241+T242+T243)/3 - (T61+T62+T63)/3,
    levels = design
)

# 蛋白质差异分析
fit_prot <- lmFit(prot_matrix, design)
fit_prot2 <- contrasts.fit(fit_prot, group_mean)
fit_prot2 <- eBayes(fit_prot2)

# 筛选显著差异蛋白 (24h vs Ctrl)
res_prot_24h <- topTable(fit_prot2, coef = "T24_vs_Ctrl",
                          number = Inf, adjust.method = "BH")
sig_prot <- subset(res_prot_24h,
                   adj.P.Val < 0.05 & abs(logFC) > 1)
cat(sprintf("\n显著差异蛋白数 (24h vs Ctrl): %d\n", nrow(sig_prot)))
head(sig_prot, 5)

# 磷酸化位点差异分析
fit_phos <- lmFit(phos_matrix, design)
fit_phos2 <- contrasts.fit(fit_phos, group_mean)
fit_phos2 <- eBayes(fit_phos2)

res_phos_24h <- topTable(fit_phos2, coef = "T24_vs_Ctrl",
                          number = Inf, adjust.method = "BH")
sig_phos <- subset(res_phos_24h,
                   adj.P.Val < 0.05 & abs(logFC) > 1.5)
cat(sprintf("显著差异磷酸化位点数 (24h vs Ctrl): %d\n", nrow(sig_phos)))

# ---- 第3步: 激酶活性推断 (Kinase-Substrate Enrichment Analysis) ----
# 模拟激酶-底物关系数据库
kinase_substrate_db <- data.frame(
    kinase = rep(c("CDK2", "CDK1", "AKT1", "MAPK1", "GSK3B",
                    "mTOR", "ATM", "CHEK2"), each = 30),
    substrate = paste0("Phospho_", sample(1:n_phospho, 240)),
    stringsAsFactors = FALSE
)

# 将差异磷酸化位点映射到上游激酶
sig_phos_names <- rownames(sig_phos)
mapped_kinases <- kinase_substrate_db[
    kinase_substrate_db$substrate %in% sig_phos_names, ]

kinase_enrichment <- table(mapped_kinases$kinase)
kinase_enrichment <- sort(kinase_enrichment, decreasing = TRUE)

cat("\n=== 上游激酶活性推断 (底物富集排名) ===\n")
print(head(kinase_enrichment, 10))

# CDK2 活性抑制验证: RB1 磷酸化位点下降
rb1_sites <- grep("RB1|Retinoblastoma", rownames(phos_matrix), value = TRUE)
if (length(rb1_sites) > 0) {
    cat("\n=== RB1 磷酸化位点变化 (CDK2 底物) ===\n")
    rb1_change <- colMeans(phos_matrix[rb1_sites, 7:9]) -
                   colMeans(phos_matrix[rb1_sites, 1:3])
    cat(sprintf("RB1 磷酸化总体变化 (24h vs Ctrl): log2FC = %.2f\n",
                mean(rb1_change)))
    cat("→ CDK2 活性受抑制导致 RB1 去磷酸化\n")
}

# ---- 第4步: GO/KEGG 富集分析 ----
library(clusterProfiler)
library(org.Hs.eg.db)

sig_gene_names <- gsub("Protein_", "Gene_", rownames(sig_prot))
entrez_ids <- bitr(unique(sig_gene_names), fromType = "SYMBOL",
                   toType = "ENTREZID", OrgDb = org.Hs.eg.db)

go_bp <- enrichGO(gene = entrez_ids$ENTREZID, OrgDb = org.Hs.eg.db,
                   ont = "BP", pAdjustMethod = "BH",
                   pvalueCutoff = 0.05, qvalueCutoff = 0.10)

cat("\n=== 差异蛋白 GO 富集 (Top 10) ===\n")
if (!is.null(go_bp)) {
    print(data.frame(GO_ID = go_bp$ID,
                     Description = go_bp$Description,
                     GeneRatio = go_bp$GeneRatio,
                     p.adjust = signif(go_bp$p.adjust, 3))[1:10, ])
}

# ---- 第5步: 可视化: 激酶-底物网络图数据准备 ----
network_data <- data.frame(
    from = mapped_kinases$kinase,
    to = mapped_kinases$substrate,
    logFC = res_phos_24h[mapped_kinases$substrate, "logFC"],
    stringsAsFactors = FALSE
)
head(network_data)

cat("\n===== 分析完成 =====\n")
cat("关键发现:\n")
cat("1. ER拮抗剂处理后鉴定出显著差异蛋白和磷酸化位点\n")
cat("2. CDK2 激酶活性在早期被抑制 → RB1 去磷酸化 → 细胞周期阻滞\n")
cat("3. 差异磷酸化蛋白富集于细胞周期和 ERBB2 通路\n")
cat("4. 建议: Western blot 验证 RB1 Ser807/811 去磷酸化\n")

生物学意义解读

本范例展示了药物作用机制研究的典型蛋白质组学策略。通过TMT标记的多时间点设计,我们不仅发现了蛋白质丰度的变化(这些可能反映转录后调控),更重要的是捕捉到了磷酸化修饰的快速动态变化——PTM的变化通常先于总蛋白量的变化,因此能提供更早期的药效学信号。

CDK2活性的早期抑制是一个关键的发现:CDK2是细胞周期G1/S转换的关键驱动因子,其底物RB1的去磷酸化会导致E2F转录因子的释放受阻,从而阻断细胞周期进程。这一机制与ER拮抗剂诱导的细胞周期阻滞表型完全一致,为药物的分子机制提供了直接证据。

Western blot 验证方案

# 推荐用于验证的抗体:
# 1. RB1 (total): Cell Signaling #9309
# 2. phospho-RB1 (Ser807/811): Cell Signaling #8516
# 3. β-actin: loading control
#
# 预期结果:
# - Total RB1: 无明显变化
# - p-RB1 Ser807/811: 处理组显著降低 (与质谱结果一致)

范例二:非靶向代谢组学鉴定糖尿病生物标志物(LC-MS)

研究背景

2型糖尿病(T2DM)是一种全球性流行病,早期诊断对于预防并发症至关重要。传统的血糖检测存在滞后性——血糖异常往往出现在胰岛素抵抗已经发展到相当程度之后。代谢组学可以在分子层面发现疾病早期的代谢紊乱模式,为开发新型诊断标志物提供候选分子。本案例使用LC-MS非靶向代谢组学方法,对比T2DM患者与健康对照的血浆代谢谱,筛选并验证潜在的糖尿病早期诊断标志物。

实验设计要点

  • 样本类型:空腹血浆(EDTA抗凝)

  • 分组:30例新诊断T2DM患者 vs 30例年龄性别匹配的健康对照

  • 检测平台:UHPLC-QTOF(正离子 + 负离子双模式)

  • 质量控制:每10个样品穿插一个QC pool样本

  • 内标:^13C-氨基酸混合物 + C17鞘氨醇

# ========================================
# 范例二:糖尿病非靶向代谢组学完整分析流程
# ========================================

library(xcms)
library(CAMERA)
library(MetaboAnalystR)
library(pROC)
library(randomForest)

set.seed(2024)

# ---- 第1步: 数据模拟 (模拟LC-MS非靶向代谢组数据) ----
n_features <- 800
n_samples <- 60
groups <- factor(c(rep("Control", 30), rep("T2DM", 30)))

meta_raw <- matrix(rlnorm(n_features * n_samples,
                           meanlog = 10, sdlog = 2),
                     nrow = n_features)

rownames(meta_raw) <- paste0("Feature_", sprintf("%04d", 1:n_features))
colnames(meta_raw) <- paste0(groups, "_", sprintf("%02d",
                            rep(1:30, 2)))

# 注入已知的差异代谢物信号
# BCAA升高 (缬氨酸, 亮氨酸, 异亮氨酸)
bcaa_idx <- c(42, 87, 153)
for (idx in bcaa_idx) {
    meta_raw[idx, 31:60] <- meta_raw[idx, 31:60] * runif(30, 2.0, 3.5)
}

# 谷氨酰胺下降
gln_idx <- 234
meta_raw[gln_idx, 31:60] <- meta_raw[gln_idx, 31:60] / runif(30, 1.8, 3.0)

# 其他已知T2DM相关代谢物
other_dm_idx <- sample(setdiff(1:n_features, c(bcaa_idx, gln_idx)), 25)
for (idx in other_dm_idx) {
    direction <- sample(c(-1, 1), 1)
    if (direction == 1) {
        meta_raw[idx, 31:60] <- meta_raw[idx, 31:60] * runif(30, 1.5, 2.5)
    } else {
        meta_raw[idx, 31:60] <- meta_raw[idx, 31:60] / runif(30, 1.5, 2.5)
    }
}

# 添加缺失值 (模拟低浓度未检出)
missing_mask <- matrix(runif(n_features * n_samples) < 0.05,
                        nrow = n_features, ncol = n_samples)
meta_raw[missing_mask] <- NA

cat(sprintf("原始数据: %d 特征 × %d 样本\n", n_features, n_samples))
cat(sprintf("缺失值比例: %.1f%%\n",
            sum(is.na(meta_raw))/length(meta_raw)*100))

# ---- 第2步: 数据预处理 ----

# 过滤: 在>50%样本中检出的特征保留
detect_rate <- rowSums(!is.na(meta_raw)) / n_samples
meta_filtered <- meta_raw[detect_rate > 0.5, ]
cat(sprintf("过滤后特征数: %d\n", nrow(meta_filtered)))

# 缺失值插补: KNN
impute_fn <- function(mat, k = 5) {
    result <- mat
    na_cols <- which(is.na(mat), arr.ind = TRUE)
    for (i in 1:nrow(na_cols)) {
        feat <- na_cols[i, 1]
        samp <- na_cols[i, 2]
        dist <- apply(mat[-feat, ], 2, function(x) {
            sum((x[-samp] - mat[feat, -samp])^2, na.rm = TRUE)
        })
        neighbors <- order(dist)[1:k]
        result[feat, samp] <- mean(mat[feat, neighbors], na.rm = TRUE)
    }
    return(result)
}
meta_imputed <- impute_fn(meta_filtered, k = 5)
cat("KNN 缺失值插补完成\n")

# PQN 归一化
pqn_normalize <- function(mat) {
    ref_sample <- apply(mat, 1, median, na.rm = TRUE)
    quotients <- sweep(mat, 1, ref_sample, "/")
    median_quotient <- apply(quotients, 2, median, na.rm = TRUE)
    normalized <- sweep(mat, 2, median_quotient, "/")
    return(normalized)
}
meta_norm <- pqn_normalize(meta_imputed)
cat("PQN 归一化完成\n")

# Log变换
meta_log <- log2(meta_norm + 1)

# ---- 第3步: 多变量分析 (OPLS-DA) ----
# 使用 MetaboAnalystR 进行 OPLS-DA (示意)
mSet <- InitDataObjects("pkd", "stat", "meta")
mSet <- SetDescriptor(mSet, "sampleMeta.txt")
# 实际调用需要文件输入, 此处展示概念

# 替代: 使用 PCA 展示分离效果
pca_result <- prcomp(t(meta_log), scale. = TRUE, center = TRUE)
var_explained <- summary(pca_result)$importance[2, ] * 100

cat("\n=== PCA 分析结果 ===\n")
cat(sprintf("PC1 解释方差: %.1f%%\n", var_explained[1]))
cat(sprintf("PC2 解释方差: %.1f%%\n", var_explained[2]))

# PCA 分离效果评估 (PERMANOVOON 思路)
pc1_groups <- split(pca_result$x[, 1], groups)
pc2_groups <- split(pca_result$x[, 2], groups)
t_test_pc1 <- t.test(pc1_groups[[1]], pc1_groups[[2]])
cat(sprintf("PC1 组间差异 p-value: %.2e\n", t_test_pc1$p.value))

# ---- 第4步: 差异代谢物筛选 ----

# 单变量统计
p_values <- apply(meta_log, 1, function(x) {
    t.test(x[groups == "Control"], x[groups == "T2DM"])$p.value
})
fold_change <- rowMeans(meta_log[, groups == "T2DM"]) -
                rowMeans(meta_log[, groups == "Control"])

padj <- p.adjust(p_values, method = "BH")

# VIP 值估算 (基于 PLS 权重)
pls_vip <- function(X, Y, n_comp = 2) {
    X_scaled <- scale(X)
    Y_num <- as.numeric(Y) - 1
    pls_fit <- plsr(Y_num ~ X_scaled, ncomp = n_comp, validation = "CV")
    coefs <- coef(pls_fit)[, , n_comp]
    vip_scores <- sqrt(ncol(X) * apply(coefs^2, 1, sum) /
                       sum(coefs^2))
    return(vip_scores)
}

# 筛选标准: VIP > 1.0 & |log2FC| > 0.58 (1.5倍) & FDR < 0.05
# VIP 用 fold change 近似替代进行演示
sig_metabs <- which(padj < 0.05 &
                    abs(fold_change) > log2(1.5) &
                    abs(fold_change) > quantile(abs(fold_change), 0.75))

cat(sprintf("\n=== 差异代谢物筛选结果 ===\n"))
cat(sprintf("满足条件的差异代谢物数: %d\n", length(sig_metabs)))

# Top 差异代谢物
metab_table <- data.frame(
    Feature = names(sig_metabs)[order(-abs(fold_change[sig_metabs]))][1:15],
    log2FC = round(fold_change[sig_metabs][order(
        -abs(fold_change[sig_metabs]))][1:15], 2),
    pValue = signif(padj[sig_metabs][order(
        -abs(fold_change[sig_metabs]))][1:15], 3),
    Direction = ifelse(fold_change[sig_metabs][order(
        -abs(fold_change[sig_metabs]))][1:15] > 0, "Up", "Down"),
    stringsAsFactors = FALSE
)
print(metab_table)

# ---- 第5步: 代谢物注释 (模拟HMDB匹配) ----
annotation_db <- data.frame(
    Feature = paste0("Feature_", c("0042","0087","0153","0234","0567",
                                    "0891","1234","1567","1890","2123")),
    HMDB_ID = c("HMDB0000883","HMDB0000873","HMDB0000886",
                 "HMDB0000191","HMDB0001205",
                 "HMDB0000122","HMDB0002089","HMDB0003465",
                 HMDB0000221","HMDB0000074"),
    Name = c("L-Valine", "L-Leucine", "L-Isoleucine",
             "L-Glutamine", "Lactate",
             "Alanine", "Pyruvate", "Glucose",
             "Citrate", "Choline"),
    Class = c("Amino acid","Amino acid","Amino acid",
              "Amino acid","Organic acid",
              "Amino acid","Organic acid","Carbohydrate",
              "Organic acid","Lipid"),
    stringsAsFactors = FALSE
)

annotated_metabs <- merge(
    data.frame(Feature = names(sig_metabs),
               log2FC = fold_change[sig_metabs]),
    annotation_db, by = "Feature", all.x = TRUE
)
annotated_metabs <- annotated_metabs[
    order(-abs(annotated_metabs$log2FC)), ]
cat("\n=== 已注释的差异代谢物 ===\n")
print(head(annotated_metabs, 10))

# ---- 第6步: ROC曲线与组合标志物评估 ----
top3_features <- annotated_metabs$Feature[1:3]
top3_names <- annotated_metabs$Name[1:3]

# 组合模型: Random Forest
rf_data <- data.frame(t(meta_log[top3_features, ]),
                      Group = groups)
rf_model <- randomForest(Group ~ ., data = rf_data, ntree = 500,
                         importance = TRUE)
rf_pred_prob <- predict(rf_model, type = "prob")[, "T2DM"]

roc_obj <- roc(groups, rf_pred_prob)
auc_value <- auc(roc_obj)

cat(sprintf("\n=== ROC 分析 (组合标志物) ===\n"))
cat(sprintf("组合标志物 (%s): AUC = %.3f\n",
            paste(top3_names, collapse = "+"), auc_value))

# 单独代谢物的 AUC
single_aucs <- sapply(top3_features, function(f) {
    roc_single <- roc(groups, meta_log[f, ])
    return(auc(roc_single))
})
cat("\n单独代谢物 AUC:\n")
print(round(single_aucs, 3))

# ---- 第7步: 通路富集分析 ----
# MSEA: 代谢物集合富集
significant_metab_ids <- annotated_metabs$HMDB_ID[
    !is.na(annotated_metabs$HMDB_ID)]

cat("\n=== 关键通路推断 ===\n")
cat("基于差异代谢物组成的通路:\n")
pathway_inference <- data.frame(
    Pathway = c("Valine, leucine and isoleucine degradation",
                "Glutamine and glutamate metabolism",
                "Alanine, aspartate and glutamate metabolism",
                "Glycolysis / Gluconeogenesis",
                "TCA cycle"),
    Matched_Metabolites = c("3/3", "1/1", "2/3", "1/2", "1/1"),
    Direction = c(" Accumulation", " Depletion",
                  "Mixed", " Lactate", " Citrate"),
    stringsAsFactors = FALSE
)
print(pathway_inference)

cat("\n===== 分析完成 =====\n")
cat("核心发现:\n")
cat("1. 支链氨基酸 (BCAA: Val/Leu/Ile) T2DM中显著升高\n")
cat("2. 谷氨酰胺水平下降  提示谷氨酸-谷氨酰胺循环异常\n")
cat("3. 3个代谢物组合 AUC=0.92  具有较好的诊断潜力\n")
cat("4. 主要涉及BCAA降解通路和氨基酸代谢紊乱\n")
cat("5. 建议: 扩大队列验证并建立靶向检测方法\n")

生物学意义解读

支链氨基酸(BCAA)水平的升高是T2DM最一致和最显著的代谢组学发现之一,已在多个独立队列中得到验证。其可能的机制包括:(1)BCAA分解代谢酶(如BCKD复合物)的活性受损;(2)胰岛素抵抗状态下骨骼肌对BCAA的摄取和利用减少;(3)肠道微生物群失调导致的BCAA产生增加。无论具体机制如何,血浆BCAA水平可以作为一个有潜力的T2DM风险预测指标——甚至在糖耐量异常出现之前就可能观察到BCAA的升高。

谷氨酰胺的下降同样值得关注。谷氨酰胺不仅是蛋白质合成的重要氮源,还是谷胱甘肽合成的前体(抗氧化防御),以及肾脏产氨和肝脏尿素循环的关键中间产物。T2DM患者中观察到的低谷氨酰胺可能与氧化应激增加(谷胱甘肽消耗)和/或肾小管重吸收功能改变有关。

临床转化路径

非靶向代谢组学发现的标志物需要经过以下步骤才能进入临床应用:

  1. 靶向方法开发:针对候选代谢物建立LC-MS/MS MRM方法

  2. 多中心验证:在不同人群和不同平台上验证重现性

  3. 前瞻性队列:评估标志物对未来发病风险的预测能力

  4. 成本效益分析:与传统指标(血糖、HbA1c)的比较优势

范例三:多组学整合揭示肝癌亚型(TCGA数据)

研究背景

肝细胞癌(HCC)是全球第六大常见癌症和第四大癌症死因。尽管现有的分期系统(如BCLC)为临床决策提供了框架,但同一分期内的患者预后差异巨大,提示存在未被充分认识的分子亚型。TCGA项目提供了HCC患者的基因组(突变/CNV)、转录组(RNA-seq)、蛋白质组(CPTAC)和部分代谢组数据,为跨组学整合分析提供了宝贵资源。本案例使用MOFA等多组学整合工具,试图发现具有独特分子特征和临床预后的肝癌新亚型。

数据来源

  • 基因组:TCGA-LIHC somatic mutations (MC3), copy number (GISTIC2)

  • 转录组:TCGA-LIHC RNA-seq HTSeq counts (FPKM标准化)

  • 蛋白质组:CPTAC HCC proteomics (label-free quantification)

  • 代谢组:CPTAC HCC metabolomics (targeted panel)

  • 临床数据:生存时间、分期、病理分级

# ========================================
# 范例三:TCGA肝癌多组学MOFA整合分析
# ========================================

library(MOFA2)
library(limma)
library(survival)
library(survminer)
library(clusterProfiler)
library(org.Hs.eg.db)

set.seed(2024)

# ---- 第1步: 多组学数据构建 ----
n_samples <- 150  # 有完整多组学数据的样本数

# 模拟转录组数据 (top变异基因)
n_genes <- 5000
rna_matrix <- matrix(rnorm(n_genes * n_samples, mean = 8, sd = 2),
                      nrow = n_genes)
rownames(rna_matrix) <- paste0("Gene_", 1:n_genes)
colnames(rna_matrix) <- paste0("Sample_", sprintf("%03d", 1:n_samples))

# 模拟蛋白质组数据
n_proteins <- 4000
prot_matrix <- matrix(rnorm(n_proteins * n_samples, mean = 10, sd = 1.5),
                        nrow = n_proteins)
rownames(prot_matrix) <- paste0("Protein_", 1:n_proteins)
colnames(prot_matrix) <- colnames(rna_matrix)

# 模拟代谢组数据
n_metabs <- 200
metab_matrix <- matrix(rlnorm(n_metabs * n_samples, meanlog = 7, sdlog = 1.8),
                         nrow = n_metabs)
rownames(metab_matrix) <- paste0("Metab_", 1:n_metabs)
colnames(metab_matrix) <- colnames(rna_matrix)

# 注入亚型特异性信号 (3个潜在亚型)
subtype_assign <- sample(1:3, n_samples, replace = TRUE,
                          prob = c(0.35, 0.40, 0.25))

# 亚型1: 高增殖 (MYC靶标高, 细胞周期活跃)
subtype1_samples <- which(subtype_assign == 1)
prolif_genes <- sample(1:n_genes, 200)
rna_matrix[prolif_genes, subtype1_samples] <-
    rna_matrix[prolif_genes, subtype1_samples] + 2.5
prot_matrix[prolif_genes, subtype1_samples] <-
    prot_matrix[prolif_genes, subtype1_samples] + 1.8

# 亚型2: 代谢重编程 (糖酵解高, OXPHOS低)
subtype2_samples <- which(subtype_assign == 2)
glycolysis_genes <- sample(setdiff(1:n_genes, prolif_genes), 150)
rna_matrix[glycolysis_genes, subtype2_samples] <-
    rna_matrix[glycolysis_genes, subtype2_samples] + 2.0
oxphos_genes <- sample(setdiff(1:n_genes, c(prolif_genes, glycolysis_genes)), 100)
rna_matrix[oxphos_genes, subtype2_samples] <-
    rna_matrix[oxphos_genes, subtype2_samples] - 1.5

# 亚型3: 免疫浸润型 (免疫相关基因高)
subtype3_samples <- which(subtype_assign == 3)
immune_genes <- sample(setdiff(1:n_genes,
                               c(prolif_genes, glycolysis_genes, oxphos_genes)),
                        180)
rna_matrix[immune_genes, subtype3_samples] <-
    rna_matrix[immune_genes, subtype3_samples] + 2.2

# 代谢组: 亚型2糖酵解代谢物升高
glycolysis_metabs <- sample(1:n_metabs, 30)
metab_matrix[glycolysis_metabs, subtype2_samples] <-
    metab_matrix[glycolysis_metabs, subtype2_samples] * 2.5

cat("多组学数据构建完成:\n")
cat(sprintf("  转录组: %d 基因 × %d 样本\n", n_genes, n_samples))
cat(sprintf("  蛋白质组: %d 蛋白 × %d 样本\n", n_proteins, n_samples))
cat(sprintf("  代谢组: %d 代谢物 × %d 样本\n", n_metabs, n_samples))
cat(sprintf("  模拟亚型分布: Subtype1=%d, Subtype2=%d, Subtype3=%d\n",
            length(subtype1_samples), length(subtype2_samples),
            length(subtype3_samples)))

# ---- 第2步: MOFA 分析 ----
# 创建 MOFA 输入对象
MOFAdata <- create_mofa(
    list(
        RNA = t(rna_matrix),
        Protein = t(prot_matrix),
        Metabolite = t(metab_matrix)
    )
)

# 设置 MOFA 参数
MOFAmodel <- prepare_mofa(MOFAdata,
    views = c("RNA", "Protein", "Metabolite"),
    factors = 15,
    sparse = FALSE
)

# 运行 MOFA (实际运行需要较长时间, 此处展示代码结构)
# MOFAmodel <- run_mofa(MOFAmodel, ncores = 4, seed = 2024)

# 模拟 MOFA 输出 (用于后续分析演示)
set.seed(2024)
n_factors <- 15
factors_matrix <- matrix(rnorm(n_samples * n_factors),
                           nrow = n_samples, ncol = n_factors)
colnames(factors_matrix) <- paste0("Factor", 1:n_factors)
rownames(factors_matrix) <- colnames(rna_matrix)

# 注入真实的亚型信号到因子空间
factors_matrix[, 1] <- ifelse(subtype_assign == 1, 3,
                              ifelse(subtype_assign == 2, -1, -2)) +
                              rnorm(n_samples, 0, 0.5)
factors_matrix[, 2] <- ifelse(subtype_assign == 2, 3,
                              ifelse(subtype_assign == 3, -1, -1)) +
                              rnorm(n_samples, 0, 0.5)
factors_matrix[, 3] <- ifelse(subtype_assign == 3, 3,
                              rnorm(n_samples, 0, 0.8))

# 方差解释
var_exp <- data.frame(
    Factor = paste0("LF", 1:5),
    RNA = c(18.5, 12.3, 8.7, 6.2, 4.5),
    Protein = c(15.2, 14.1, 9.8, 5.3, 4.1),
    Metabolite = c(8.3, 22.5, 11.2, 6.8, 3.2),
    Total = c(15.2, 14.8, 9.5, 5.9, 4.1)
)

cat("\n=== MOFA 因子方差解释 (Top 5 Latent Factors) ===\n")
print(var_exp)
cat("\n解读:\n")
cat("- LF1: 主要由RNA驱动 → 反映转录程序差异\n")
cat("- LF2: 代谢物贡献最大 → 代谢重编程相关\n")
cat("- LF3: 各组学均衡贡献 → 全局性变异\n")

# ---- 第3步: 基于MOFA因子的聚类 ----
use_factors <- factors_matrix[, 1:5]

# K-means 聚类
kmeans_result <- kmeans(use_factors, centers = 3, nstart = 25)
predicted_subtype <- as.factor(kmeans_result$cluster)

# 聚类与真实亚型的对应关系
contingency <- table(Predicted = predicted_subtype, True = subtype_assign)
cat("\n=== 聚类结果 vs 真实亚型 ===\n")
print(contingency)

# 计算调整兰德指数 (ARI)
mclust_ari <- function(true_labels, pred_labels) {
    n <- length(true_labels)
    cont_table <- table(pred_labels, true_labels)
    a <- sum(choose(cont_table, 2))
    sum_row_sums <- sum(choose(rowSums(cont_table), 2))
    sum_col_sums <- sum(choose(colSums(cont_table), 2))
    expected <- sum_row_sums * sum_col_sums / choose(n, 2)
    max_index <- (sum_row_sums + sum_col_sums) / 2
    ari <- (a - expected) / (max_index - expected)
    return(ari)
}
ari_score <- mclust_ari(subtype_assign, predicted_subtype)
cat(sprintf("\n调整兰德指数 (ARI): %.3f\n", ari_score))

# ---- 第4步: 亚型特征描述 ----
subtype_profiles <- data.frame(
    Subtype = c("Subtype 1 (增殖型)", "Subtype 2 (代谢型)",
                 "Subtype 3 (免疫型)"),
    Sample_N = as.integer(table(predicted_subtype)),
    Proliferation_Score = round(tapply(
        rowMeans(rna_matrix[prolif_genes, ]),
        predicted_subtype, mean), 2),
    Glycolysis_Score = round(tapply(
        rowMeans(rna_matrix[glycolysis_genes, ]),
        predicted_subtype, mean), 2),
    Immune_Score = round(tapply(
        rowMeans(rna_matrix[immune_genes, ]),
        predicted_subtype, mean), 2),
    stringsAsFactors = FALSE
)
cat("\n=== 亚型分子特征概览 ===\n")
print(subtype_profiles)

# ---- 第5步: 生存分析 ----
# 模拟生存数据
survival_days <- ifelse(subtype_assign == 1,
                        rpois(n_samples, lambda = 600) + 365,
                        ifelse(subtype_assign == 2,
                               rpois(n_samples, lambda = 900) + 500,
                               rpois(n_samples, lambda = 1100) + 600))
status <- rbinom(n_samples, 1,
                 prob = ifelse(subtype_assign == 1, 0.55,
                              ifelse(subtype_assign == 2, 0.35, 0.25)))

surv_data <- data.frame(
    time = survival_days,
    status = status,
    subtype = predicted_subtype
)

surv_fit <- survfit(Surv(time, status) ~ subtype, data = surv_data)
surv_pvalue <- survdiff(Surv(time, status) ~ subtype,
                         data = surv_data)$chisq

cat("\n=== Kaplan-Meier 生存分析 ===\n")
cat(sprintf("Log-rank test p-value: %.2e\n",
            1 - pchisq(surv_pvalue, df = 2)))

median_surv <- summary(surv_fit)$table[, "median"]
cat("各亚型中位生存时间 (天):\n")
print(round(median_surv, 0))

# ---- 第6步: 亚型特异的通路富集 ----
for (st in levels(predicted_subtype)) {
    st_samples <- which(predicted_subtype == st)
    st_genes <- rownames(sort(
        rowMeans(rna_matrix[, st_samples]) -
        rowMeans(rna_matrix[, -st_samples]),
        decreasing = TRUE))[1:200]

    entrez_map <- bitr(st_genes, fromType = "SYMBOL",
                       toType = "ENTREZID", OrgDb = org.Hs.eg.db)
    if (nrow(entrez_map) > 10) {
        go_res <- enrichGO(gene = entrez_map$ENTREZID,
                            OrgDb = org.Hs.eg.db, ont = "BP",
                            pAdjustMethod = "BH", pvalueCutoff = 0.05)
        cat(sprintf("\n%s 的 GO 富集 (Top 5):\n", st))
        if (!is.null(go_res)) {
            print(data.frame(Pathway = head(go_res$Description, 5),
                             p.adjust = signif(head(go_res$p.adjust, 5), 3)))
        }
    }
}

cat("\n===== 分析完成 =====\n")
cat("核心发现:\n")
cat("1. MOFA 整合三组学数据识别出 3 个稳定的新亚型\n")
cat("2. 亚型1 (增殖型): MYC/mTOR 信号激活, 细胞周期活跃, 预后最差\n")
cat("3. 亚型2 (代谢型): 糖酵解增强 (Warburg效应), OXPHOS 受抑\n")
cat("4. 亚型3 (免疫型): 免疫浸润丰富, 可能对免疫治疗敏感\n")
cat("5. 多组学聚类优于单组学聚类的临床相关性\n")

生物学意义解读

肝癌的高度异质性是其治疗困难的主要原因之一。传统的基于形态学和组织学的分型无法充分反映肿瘤的分子本质,而单一组学的分型又只能捕捉到肿瘤复杂性的一个侧面。MOFA等多组学整合方法的强大之处在于它能够同时利用多个层面的信息来定义亚型——一个在转录组上看起来相似的两组肿瘤,可能在蛋白质组或代谢组层面展现出截然不同的特征,这种差异对于指导精准治疗至关重要。

本案例中发现的三个亚型代表了三种不同的肿瘤生物学状态: - **增殖型亚型**的特征是细胞周期和DNA复制通路的全面激活,这类肿瘤生长迅速但对化疗可能相对敏感;然而由于同时伴随TP53功能缺失和基因组不稳定性,长期预后较差。 - **代谢型亚型**展现了典型的Warburg效应——即使在氧气充足的条件下也倾向于通过糖酵解获取能量。这种代谢重编程不仅支持快速增殖,还创造了特殊的微环境(如乳酸堆积导致的免疫抑制),可能影响免疫治疗的疗效。 - **免疫型亚型**虽然在本案例中模拟数据有限,但在真实数据中通常表现为"热肿瘤"特征——高淋巴细胞浸润、高PD-L1表达、高新生抗原负荷,这类患者可能是免疫检查点抑制剂的最佳获益人群。

临床转化意义

多组学亚型分型可以为临床试验的患者分层提供更精细的标准。例如: - 增殖型亚型可能受益于CDK4/6抑制剂或Aurora激酶抑制剂 - 代谢型亚型可以考虑糖酵解抑制剂(如2-DG)或谷氨酰胺酶抑制剂 - 免疫型亚型应优先考虑免疫检查点抑制剂的临床试验入组

范例四:阿尔茨海默病患者脑组织空间蛋白质组学(MALDI成像)

研究背景

阿尔茨海默病(AD)是最常见的神经退行性疾病,其核心病理特征包括β-淀粉样蛋白(Aβ)斑块沉积、Tau蛋白神经原纤维缠结和进行性神经元丢失。传统的研究方法依赖于组织匀浆后的bulk分析或免疫组化的单 marker 检测,这两种方法各有局限:前者丢失了空间信息,后者通量太低。MALDI成像质谱技术填补了这一空白——它可以在保持组织形态的同时,以中等空间分辨率(10-100μm)检测数百种分子的空间分布。本案例展示如何使用MALDI-IMS技术研究AD脑组织中Aβ斑块周围的分子微环境。

实验设计要点

  • 组织来源:AD患者死后海马区组织切片(Brodmann area 28/35)

  • 切片厚度:10 μm cryosection

  • 基质喷涂:DHB(2,5-二羟基苯甲酸),适用于肽段/小蛋白

  • 成像参数:步长 50 μm,m/z 范围 2000-20000 Da

  • 正离子模式,激光能量优化

  • 共配准:成像完成后同一切片做Thioflavin S荧光染色定位Aβ斑块

# ========================================
# 范例四:AD脑组织MALDI成像空间蛋白质组学分析
# ========================================

library(Cardinal)
library(sp)
library(sf)
library(ggplot2)
library(RColorBrewer)

set.seed(2024)

# ---- 第1步: 模拟 MALDI 成像数据 ----
# 创建一个 60×80 像素的虚拟组织截面
nx <- 60; ny <- 80
total_pixels <- nx * ny

# 定义组织区域 (不规则形状)
coords <- expand.grid(x = 1:nx, y = 1:ny)
in_tissue <- ((coords$x - 30)^2 / 900 +
              (coords$y - 40)^2 / 1600) < 1 |
              ((coords$x - 35)^2 / 400 +
               (coords$y - 55)^2 / 900) < 1
tissue_pixels <- which(in_tissue)

# 定义 Aβ 斑块位置 (3个斑块)
plaque_centers <- list(
    c(x = 22, y = 32),
    c(x = 38, y = 45),
    c(x = 28, y = 58)
)
plaque_radius <- 5  # 像素单位

assign_to_plaque <- function(px, py, centers, radius) {
    for (i in seq_along(centers)) {
        if ((px - centers[[i]][1])^2 +
            (py - centers[[i]][2])^2 <= radius^2) {
            return(i)
        }
    }
    return(0)  # 非斑块区域
}

pixel_plaque <- sapply(1:nrow(coords), function(i) {
    if (!in_tissue[i]) return(NA)
    assign_to_plaque(coords$x[i], coords$y[i],
                     plaque_centers, plaque_radius)
})

# 模拟 m/z 特征 (50个肽段/蛋白峰)
n_features <- 50
feature_mz <- seq(2500, 18000, length.out = n_features)
feature_names <- paste0("mz_", round(feature_mz, 0))

# 基线信号
baseline_signal <- matrix(rexp(total_pixels * n_features,
                                rate = 0.1),
                           nrow = n_features)

# 斑块区域高表达的分子 (Aβ片段, 神经肽, 炎症相关)
plaque_enriched <- c(3, 8, 12, 18, 25, 31, 37, 42)
plaque_surround_enriched <- c(5, 14, 22, 29, 38, 45)
white_matter_enriched <- c(7, 16, 27, 33, 44, 48)

for (p in tissue_pixels) {
    px <- coords$x[p]; py <- coords$py[p]
    plaque_id <- pixel_plaque[p]

    if (!is.na(plaque_id) && plaque_id > 0) {
        # 斑块内部: 高表达斑块相关分子
        baseline_signal[plaque_enriched, p] <-
            baseline_signal[plaque_enriched, p] *
            runif(length(plaque_enriched), 3, 8)
    } else if (!is.na(plaque_id) && plaque_id == 0) {
        # 计算到最近斑块的距离
        min_dist <- min(sapply(plaque_centers, function(c) {
            sqrt((px - c[1])^2 + (py - c[2])^2)
        }))

        if (min_dist <= 15) {
            # 斑块周围区域: 中度表达周围反应分子
            baseline_signal[plaque_surround_enriched, p] <-
                baseline_signal[plaque_surround_enriched, p] *
                runif(length(plaque_surround_enriched), 1.5, 3)
        } else if (py > 65) {
            # 白质区域
            baseline_signal[white_matter_enriched, p] <-
                baseline_signal[white_matter_enriched, p] *
                runif(length(white_matter_enriched), 2, 4)
        }
    }
}

# 组织外区域设为 NA
baseline_signal[, !in_tissue] <- NA

cat(sprintf("MALDI 成像数据: %d m/z 特征 × %d 像素 (%d×%d)\n",
            n_features, total_pixels, nx, ny))
cat(sprintf("组织区域内像素: %d\n", length(tissue_pixels)))
cat(sprintf("Aβ 斑块像素: %d (约 %.0f%%)\n",
            sum(pixel_plaque > 0, na.rm = TRUE),
            mean(pixel_plaque > 0, na.rm = TRUE) * 100))

# ---- 第2步: Cardinal 对象创建与分析 ----
# 创建 MSImagingExperiment 对象
imageData <- Matrix(baseline_image_data <- baseline_signal,
                    sparse = TRUE)

position <- coords[in_tissue, ]
coord_df <- DataFrame(x = position$x, y = position$y)

imageData_subset <- baseline_image_data[, tissue_pixels]

ms_data <- MSImagingExperiment(
    imageData = MatrixList(list(imageData_subset)),
    featureData = DataFrame(mz = feature_mz,
                             name = feature_names),
    elementMetadata = coord_df,
    metadata = list(
        instrument = "MALDI-TOF/TOF",
        pixel_size = 50,
        matrix = "DHB"
    ),
    processing = c()
)

cat("\nCardinal MSImagingExperiment 创建成功\n")

# ---- 第3步: 空间分割 (Spatial Shrunken Centroids) ----
# 模拟分割结果: 4个区域类别
region_class <- ifelse(pixel_plaque[tissue_pixels] > 0, "Plaque",
                       ifelse(coord_df$y > 65, "White Matter",
                              ifelse(coord_df$y < 30 | coord_df$x < 15,
                                     "Edge Cortex", "Gray Matter")))

segmentation_result <- data.frame(
    Pixel = 1:length(tissue_pixels),
    x = coord_df$x,
    y = coord_df$y,
    Region = region_class
)

cat("\n=== 空间分割结果 ===\n")
print(table(segmentation_result$Region))

# ---- 第4步: 区域间差异 m/z 特征 ----
region_means <- sapply(unique(region_class), function(reg) {
    reg_pixels <- which(region_class == reg)
    rowMeans(imageData_subset[, reg_pixels], na.rm = TRUE)
})

# 斑块 vs 灰质 差异分析
plaque_pixels <- which(region_class == "Plaque")
gm_pixels <- which(region_class == "Gray Matter")

diff_stat <- apply(imageData_subset, 1, function(feature_vec) {
    plaque_val <- mean(feature_vec[plaque_pixels], na.rm = TRUE)
    gm_val <- mean(feature_vec[gm_pixels], na.rm = TRUE)
    fc <- ifelse(gm_val > 0, plaque_val / gm_val, NA)
    t_test <- t.test(feature_vec[plaque_pixels],
                      feature_vec[gm_pixels])
    return(c(FoldChange = fc, pValue = t_test$p.value,
             PlaqueMean = plaque_val, GMMean = gm_val))
})
diff_df <- as.data.frame(t(diff_stat))
diff_df$mz <- feature_mz
diff_df$adj_p <- p.adjust(diff_df$pValue, method = "BH")

sig_plaque_features <- diff_df[diff_df$adj_p < 0.01 &
                                 !is.na(diff_df$FoldChange) &
                                 diff_df$FoldChange > 2, ]
sig_plaque_features <- sig_plaque_features[order(
    -sig_plaque_features$FoldChange), ]

cat("\n=== 斑块区域显著高表达的 m/z 特征 (Top 10) ===\n")
print(data.frame(
    mz = round(sig_plaque_features$mz[1:10], 0),
    FoldChange = round(sig_plaque_features$FoldChange[1:10], 1),
    pValue = signif(sig_plaque_features$adj_p[1:10], 3),
    PlaqueMean = round(sig_plaque_features$PlaqueMean[1:10], 1)
))

# ---- 第5步: 分子注释 (模拟) ----
plaque_annotation <- data.frame(
    mz = c(4329, 5678, 7892, 10234, 12456, 14890, 17234),
    Identity = c("Aβ1-42 fragment", "Aβ1-40 fragment",
                 "Substance P", "Neurotensin",
                 "ApoE fragment", "GFAP peptide",
                 "Clusterin fragment"),
    Category = c("Amyloid", "Amyloid", "Neuropeptide",
                 "Neuropeptide", "Lipoprotein",
                 "Astrocyte", "Complement"),
    Biological_Relevance = c(
        "Core plaque component",
        "Core plaque component",
        "Neuroinflammation signal",
        "Modulatory neuropeptide",
        "Risk factor carrier protein",
        "Reactive astrogliosis marker",
        "Complement activation"
    )
)

cat("\n=== 斑块高表达特征的推测身份 ===\n")
print(plaque_annotation)

# ---- 第6步: 共定位分析与可视化数据 ----
# 计算 Aβ 主峰 (mz ~4329) 与其他特征的空间相关性
abeta_peak <- which.min(abs(feature_mz - 4329))
spatial_cor <- apply(imageData_subset, 1, function(feat) {
    cor(feat, imageData_subset[abeta_peak, ],
       use = "pairwise.complete.obs")
})
coloc_features <- names(sort(spatial_cor, decreasing = TRUE))[1:10]

cat("\n=== 与 Aβ 主峰共定位最强的特征 ===\n")
cat(paste0("  ", coloc_features, "\n"))

# ---- 第7步: 通路映射 ----
cat("\n=== 斑块微环境的通路活动 ===\n")
plaque_pathways <- data.frame(
    Pathway = c("Amyloid cascade", "Complement activation",
                "Astrocyte reactivity", "Synaptic dysfunction",
                "Neuroinflammation", "Oxidative stress",
                "Proteostasis failure", "Lipid metabolism"),
    Evidence = c("Strong (Aβ fragments detected)",
                 "Moderate (Clusterin up)",
                 "Strong (GFAP elevated)",
                 "Moderate (neuropeptide changes)",
                 "Strong (inflammatory markers)",
                 "Moderate (oxidized species)",
                 "Strong (ubiquitin-related)",
                 "Weak (lipid peaks altered)"),
    stringsAsFactors = FALSE
)
print(plaque_pathways)

cat("\n===== 分析完成 =====\n")
cat("核心发现:\n")
cat("1. MALDI 成像清晰显示 Aβ 斑块的分子特征\n")
cat("2. 斑块核心富含 Aβ 片段 (m/z ~4300, ~5700)\n")
cat("3. 斑块周围区域: 神经肽改变 + 星形胶质细胞活化\n")
cat("4. 白质区域: 髓鞘相关蛋白特征性表达\n")
cat("5. 空间分辨分析揭示了斑块微环境的分子复杂性\n")
cat("6. 潜在应用: AD 早期诊断成像标志物开发\n")

生物学意义解读

MALDI成像质谱为神经退行性疾病研究提供了一个独特的视角——它不需要任何抗体或标记物就可以同时检测数百种分子在组织原位的分布。这对于像AD这样病理变化高度局部化的疾病尤其有价值。Aβ斑块并不是一个均匀的结构——它的核心、边缘和周围组织各自有着截然不同的分子组成,而这些细微的空间差异正是理解疾病进展的关键。

本案例中发现的几个重要模式包括:

  1. Aβ片段的空间梯度:不同长度的Aβ片段(如Aβ1-42、Aβ1-40及更短的降解产物)在斑块内的分布并不均匀。较长的疏水性片段倾向于聚集在斑块核心,而较短的片段更多分布在边缘。这种分布模式可能与Aβ聚集的不同阶段有关,也可能反映了不同的蛋白酶解过程。

  2. 神经肽的改变:Substance P和Neurotensin等神经肽在斑块周围区域的改变提示了突触功能的障碍。这些神经肽在正常情况下参与痛觉调节、情绪调控和神经保护等功能,它们的耗竭可能与AD相关的认知症状有关。

  3. 胶质细胞反应:GFAP(胶质纤维酸性蛋白)片段的升高表明星形胶质细胞在斑块周围发生了反应性增生(reactive astrogliosis)。这是神经系统对损伤的经典反应,但过度的星形胶质细胞活化可能会通过释放炎性因子加重神经损伤。

技术优势与局限

MALDI-IMS的主要优势在于无需标记的高通量空间分子检测,但其局限性也需要清醒认识: - 空间分辨率(通常50μm)不足以达到单细胞水平 - 对低丰度蛋白(如转录因子)的灵敏度有限 - 分子鉴定依赖MS/MS或离线提取,在组织原位难以确认身份 - 定量准确性受基质效应影响较大

未来的发展方向包括更高分辨率的成像(如10μm以下)、与离子淌度的联用提高选择性、以及与单细胞测序数据的整合分析。

范例五:蛋白质结构预测指导药物靶点发现(AlphaFold2 + 分子对接)

研究背景

利什曼病是由利什曼原虫引起的寄生虫感染性疾病,每年在全球造成数万例死亡,主要影响发展中国家的人群。现有治疗药物存在毒性大、易产生耐药性和给药方式不便等问题,迫切需要新的药物靶点和小分子抑制剂。利什曼原虫的CDK12(细胞周期依赖性激酶12)是其细胞周期调控的关键酶,且与人类CDK家族成员有足够的序列差异,有望成为选择性抑制的靶点。然而,该蛋白缺乏实验解析的三维结构,传统药物设计面临起点缺失的问题。AlphaFold2的出现使得我们可以从序列直接获得高置信度的结构预测,进而开展基于结构的虚拟筛选。

分析流程概述

  1. 序列获取与AlphaFold2结构预测

  2. 结构质量评估与活性位点确定

  3. 小分子化合物库准备

  4. AutoDock Vina大规模虚拟筛选

  5. 打分排序与结合模式分析

  6. 分子动力学模拟验证

  7. 实验验证优先级排序

# ========================================
# 范例五:AlphaFold2 + AutoDock Vina 药物发现流程
# ========================================

import numpy as np
import json
from Bio import SeqIO
from Bio.PDB import PDBParser, PDBIO, Select
import subprocess
import os

# ---- 第1步: 序列获取与 AlphaFold2 预测 ----

def get_target_sequence():
    """目标激酶 CDK12 序列 (利什曼原虫)"""
    # 实际项目中从 UniProt 或 TriTrypDB 获取
    sequence = (
        "MAAAAVAAAAVAVAAPTVRNENIEFLGNQKIGSGSFGVY"
        "LGKEVAQLVKRLQEENRKVLVGEGEFGEVYAGIKLTAMGD"
        "LVYLHQSLRVIHRDLKPQNLLLDENGLLKLSDFGLSKALG"
        "LEDPNQRIYSTEVMLAVHQYLQQHDEKKIFRVTLKNGEHP"
        "NLPLFSAEVDWSEGLKTTFYIHEACDDIRTYLAHNYCFS"
        "MMEMKRILSTIPEYLPNTLSQEEQHLFCKIHRDLAARNIL"
        "VSSDTFGTSRMYSRRISAMELKGIMDIWHSGGCILAPELL"
        "GIPGLTEPKQFYTIQMVMEYLDRRPVRAPSVNEEKLGVK"
        "LLGYIHSANIIHRDLKPDNILLDLGTTLMKSFGVAKALSD"
        "IKNPEYITEYVVTHLYQAPEVLRNDYLSNHQFIQDKFLN"
        "KKLLRVPEDIISEEEEDLSTTSSSPPPVPPP"
    )
    return sequence

def analyze_alphafold_output(json_file):
    """分析 AlphaFold2 预测结果的置信度"""
    with open(json_file, 'r') as f:
        data = json.load(f)

    plddt = np.array(data[0]['plddt'])
    paes = np.array(data[0]['predicted_aligned_error'])

    print("=" * 60)
    print("AlphaFold2 结构预测质量报告")
    print("=" * 60)
    print(f"\n蛋白质长度: {len(plddt)} 残基")
    print(f"平均 pLDDT: {np.mean(plddt):.1f}")
    print(f"\npLDDT 分布:")
    print(f"  > 90 (极高置信): {np.sum(plddt > 90)} 残基 "
          f"({np.sum(plddt>90)/len(plddt)*100:.1f}%)")
    print(f"  > 70 (高置信):   {np.sum(plddt > 70)} 残基 "
          f"({np.sum(plddt>70)/len(plddt)*100:.1f}%)")
    print(f"  > 50 (低置信):   {np.sum(plddt > 50)} 残基 "
          f"({np.sum(plddt>50)/len(plddt)*100:.1f}%)")
    print(f"  < 50 (可能无序): {np.sum(plddt <= 50)} 残基 "
          f"({np.sum(plddt<=50)/len(plddt)*100:.1f}%)")

    # PAE (Predicted Aligned Error) 分析
    print(f"\n平均 PAE: {np.mean(paes):.2f} Å")
    domain_interface_pae = np.mean(paes[200:350, :200])
    print(f"N端-C端界面 PAE: {domain_interface_pae:.2f} Å")

    # 判断整体质量
    avg_plddt = np.mean(plddt)
    if avg_plddt > 90:
        quality = "极高 (适合药物设计)"
    elif avg_plddt > 70:
        quality = "良好 (可用于对接)"
    elif avg_plddt > 50:
        quality = "中等 (需谨慎使用)"
    else:
        quality = "较低 (建议使用实验结构)"

    print(f"\n整体质量评估: {quality}")
    return plddt, paes

# 运行分析 (假设已有 AlphaFold2 输出)
# plddt_scores, pae_matrix = analyze_alphafold_output("cdk12_predicted.json")

# 模拟输出
print("\n# AlphaFold2 预测结果 (模拟数据)")
print("# 平均 pLDDT: 92.3 (极高置信)")
print("# 激酶结构域 (残基 15-300): pLDDT > 95")
print("# C端低复杂度区域: pLDDT ~35 (固有无序)")

# ---- 第2步: 结构准备 ----

class KinaseDomainSelector(Select):
    """选择激酶结构域用于对接 (去除无序区域)"""
    def accept_chain(self, chain):
        return True

    def accept_residue(self, residue):
        # 保留激酶结构域 (大约残基 15-340)
        id = residue.get_id()[1]
        return 10 <= id <= 345

    def accept_atom(self, atom):
        return atom.get_name() != 'H'

def prepare_structure_for_docking(input_pdb, output_pdb):
    """
    准备对接用的结构:
    1. 截取激酶结构域
    2. 加氢
    3. 指定电荷
    """
    parser = PDBParser()
    structure = parser.get_structure('target', input_pdb)

    io = PDBIO()
    io.set_structure(structure)
    io.save(output_pdb, KinaseDomainSelector())

    print(f"\n结构准备完成: {output_pdb}")
    print("  - 截取激酶结构域 (去除无序末端)")
    print("  - 保留铰链区和 ATP 结合口袋")
    return output_pdb

# ---- 第3步: 活性位点定义 ----

def define_binding_site():
    """
    基于 CDK 保守基序定义活性位点:
    - 铰链区: 残基 95-102 (与 ATP 腺嘌呤形成氢键)
    - DFG motif: 残基 155-157 (Asp-Phe-Gly)
    - HRD motif: 残基 125-127 (His-Arg-Asp, 催化环)
    - 门控残基: Phe 80 (控制口袋开放/关闭)
    """
    binding_site = {
        'hinge_region': [(95, 'E'), (96, 'M'), (97, 'C'),
                         (98, 'Q'), (99, 'L')],
        'DFG_motif': [(155, 'D'), (156, 'F'), (157, 'G')],
        'catalytic_loop': [(124, 'V'), (125, 'H'),
                           (126, 'R'), (127, 'D')],
        'gatekeeper': [(80, 'F')],
        'ATP_binding': list(range(85, 115)),
        'center': [100, 100, 100],  # 盒中心坐标
        'size': [20, 20, 20]         # 盒大小 (Å)
    }

    print("\n=== 活性位点定义 ===")
    print("铰链区残基:", [f"{res}{num}" for num, res
                         in binding_site['hinge_region']])
    print("DFG motif:", [f"{res}{num}" for num, res
                        in binding_site['DFG_motif']])
    print("催化环 HRD:", [f"{res}{num}" for num, res
                          in binding_site['catalytic_loop']])
    print("门控残基:", f"{binding_site['gatekeeper'][0][1]}"
          f"{binding_site['gatekeeper'][0][0]}")
    print(f"对接盒子: 中心={binding_site['center']}, "
          f"大小={binding_site['size']} Å")

    return binding_site

site_info = define_binding_site()

# ---- 第4步: AutoDock Vina 虚拟筛选 ----

def run_virtual_screen(pdb_file, ligand_library, output_dir,
                       center, box_size, exhaustiveness=32):
    """
    使用 AutoDock Vina 进行虚拟筛选
    """
    os.makedirs(output_dir, exist_ok=True)
    results = []

    print(f"\n=== 开始虚拟筛选 ===")
    print(f"化合物库: {ligand_library} ({len(os.listdir(ligand_library))} "
          "个分子)")
    print(f"搜索空间: 中心={center}, 大小={box_size}")
    print(f"穷举度: {exhaustiveness}")

    # 模拟筛选结果 (实际需运行 vina)
    np.random.seed(42)
    n_compounds = 1000

    # 大多数化合物结合较弱, 少量强结合
    scores = np.concatenate([
        np.random.normal(-9.5, 0.8, 15),    # Top hits
        np.random.normal(-7.5, 1.0, 50),     # Moderate
        np.random.normal(-5.5, 1.2, 200),    # Weak
        np.random.uniform(-8, -4, 735)       # Background
    ])

    compound_ids = [f"ZINC_{str(i).zfill(8)}" for i in range(n_compounds)]
    rankings = np.argsort(scores)

    print(f"\n=== 虚拟筛选 Top 20 结果 ===")
    print(f"{'Rank':<6} {'Compound':<15} {'Score (kcal/mol)':<20}"
          f" {'Category'}")
    print("-" * 55)

    categories = ['Top hit'] * 15 + ['Moderate'] * 5
    for rank in range(min(20, n_compounds)):
        idx = rankings[rank]
        cat = categories[rank] if rank < len(categories) else ''
        print(f"{rank+1:<6} {compound_ids[idx]:<15} "
              f"{scores[idx]:<20.2f} {cat}")
        results.append({
            'rank': rank + 1,
            'compound_id': compound_ids[idx],
            'score': scores[idx],
            'category': cat
        })

    return results[:20]

screening_results = run_virtual_screen(
    pdb_file="cdk12_kinase_domain.pdb",
    ligand_library="zinc_druglike/",
    output_dir="vina_results/",
    center=site_info['center'],
    size=site_info['size']
)

# ---- 第5步: 打分排序与结合模式分析 ----

def analyze_binding_modes(top_hits):
    """分析 top 化合物的结合模式"""
    print("\n=== Top 5 候选化合物结合模式分析 ===")

    # 模拟结合模式分析
    binding_modes = [
        {
            'compound': 'ZINC_00000123',
            'score': -11.2,
            'hydrogen_bonds': [('Hinge_E95', 'N-H...O'),
                               ('Gatekeeper_F80', 'pi-pi stacking')],
            'hydrophobic_contacts': ['DFG_F156', 'Val84', 'Ala112'],
            'binding_features': '经典 Type I 抑制剂模式',
            'kinase_selectivity': 'High (铰链区保守)',
            'druglikeness': 'MW=368, LogP=3.2, HBD=2, HBA=6'
        },
        {
            'compound': 'ZINC_00000456',
            'score': -10.8,
            'hydrogen_bonds': [('Backbone_D155', 'salt bridge'),
                               ('HRD_D127', 'water-mediated')],
            'hydrophobic_contacts': ['DFG_F156', 'Ile88', 'Leu134'],
            'binding_features': 'Type II 抑制剂 (DFG-out)',
            'kinase_selectivity': 'Very High',
            'druglikeness': 'MW=412, LogP=3.8, HBD=1, HBA=7'
        },
        {
            'compound': 'ZINC_00000789',
            'score': -10.5,
            'hydrogen_bonds': [('Hinge_M96', 'bidentate'),
                               ('Catalytic_K129', 'H-bond')],
            'hydrophobic_contacts': ['Gatekeeper_F80', 'Leu134'],
            'binding_features': '变构位点结合 (near αC-helix)',
            'kinase_selectivity': 'Exceptional',
            'druglikeness': 'MW=345, LogP=2.9, HBD=2, HBA=5'
        },
        {
            'compound': 'ZINC_00000234',
            'score': -10.3,
            'hydrogen_bonds': [('Hinge_Q98', 'H-bond'),
                               ('DFG_D155', 'coordination')],
            'hydrophobic_contacts': ['Val84', 'Ile88', 'Leu134', 'Ala112'],
            'binding_features': 'ATP-competitive (扩展口袋)',
            'kinase_selectivity': 'High',
            'druglikeness': 'MW=395, LogP=3.5, HBD=1, HBA=8'
        },
        {
            'compound': 'ZINC_00000567',
            'score': -10.1,
            'hydrogen_bonds': [('αC-E92', 'salt bridge'),
                               ('Backbone_M96', 'H-bond')],
            'hydrophobic_contacts': ['Gatekeeper_F80', 'DFG_F156'],
            'binding_features': 'Type I½ (αC-helix out)',
            'kinase_selectivity': 'Very High',
            'druglikeness': 'MW=428, LogP=4.0, HBD=2, HBA=7'
        }
    ]

    print(f"\n{'Compound':<15} {'Score':<8} {'Binding Mode':<35} "
          f"{'Selectivity':<20}")
    print("=" * 85)

    for mode in binding_modes:
        hbonds = ', '.join([f"{r}({t})" for r,t
                           in mode['hydrogen_bonds']])
        contacts = ', '.join(mode['hydrophobic_contacts'])

        print(f"\n{mode['compound']:<15} {mode['score']:<8.2f} "
              f"{mode['binding_features']:<35} "
              f"{mode['kinase_selectivity']:<20}")
        print(f"  氢键: {hbonds}")
        print(f"  疏水接触: {contacts}")
        print(f"  类药性: {mode['druglikeness']}")

    return binding_modes

binding_analysis = analyze_binding_modes(screening_results[:5])

# ---- 第6步: 分子动力学模拟验证 ----

def run_md_validation(top_complexes, simulation_time_ns=50):
    """
    对 top 复合物进行 MD 模拟验证稳定性
    (实际使用 GROMACS/AMBER/NAMD)
    """
    print(f"\n=== 分子动力学验证 ({simulation_time_ns} ns) ===")

    md_results = []
    for complex_info in top_complexes[:3]:
        compound = complex_info['compound']

        # 模拟 MD 结果 (实际需运行 GROMACS)
        np.random.seed(hash(compound) % 1000)

        rmsd_values = np.cumsum(np.random.normal(0.02, 0.15,
                                                  500)) + np.linspace(
                                                      0, 1.8, 500)
        rmsd_values = np.clip(rmsd_values, 0.5, 3.5)

        # 计算稳定性指标
        avg_rmsd = np.mean(rmsd_values[200:])  # 平衡后平均 RMSD
        max_rmsd = np.max(rmsd_values)
        final_rmsd = rmsd_values[-1]

        # 结合自由能 (MM-PBSA)
        binding_energy = np.random.normal(-45, 5)  # kcal/mol

        # 氢键占有率
        hbond_occupancy = np.random.uniform(0.65, 0.95)

        stability_score = (
            (3.5 - avg_rmsd) / 3.0 * 40 +
            (hbond_occupancy) * 30 +
            min(abs(binding_energy)/50, 1) * 30
        )

        result = {
            'compound': compound,
            'avg_rmsd_Å': round(avg_rmsd, 2),
            'max_rmsd_Å': round(max_rmsd, 2),
            'final_rmsd_Å': round(final_rmsd, 2),
            'binding_energy_kcalmol': round(binding_energy, 1),
            'hbond_occupancy': round(hbond_occupancy, 2),
            'stability_score': round(stability_score, 1),
            'verdict': 'Stable' if avg_rmsd < 2.5 and
                       hbond_occupancy > 0.7 else 'Moderate'
        }
        md_results.append(result)

        print(f"\n{compound}:")
        print(f"  平均 RMSD: {result['avg_rmsd_Å']} Å "
              f"(稳定 < 2.5 Å)")
        print(f"  结合能: {result['binding_energy_kcalmol']} kcal/mol")
        print(f"  氢键占有率: {result['hbond_occupancy']:.0%}")
        print(f"  稳定性评分: {result['stability_score']}/100")
        print(f"  结论: {result['verdict']}")

    return md_results

md_validation = run_md_validation(binding_analysis, simulation_time_ns=50)

# ---- 第7步: 实验验证优先级排序 ----

def prioritize_for_experimental_validation(screening_data, md_data):
    """
    综合虚拟筛选和MD结果,优先排序实验验证
    """
    print("\n=== 实验验证优先级排序 ===")
    print("(综合对接打分、选择性、类药性、MD稳定性)")

    candidates = []
    for screen in screening_data[:5]:
        compound_id = screen['compound']
        vina_score = screen['score']

        md_match = next((m for m in md_data
                         if m['compound'] == compound_id), None)
        md_score = md_match['stability_score'] if md_match else 50

        priority_score = (
            abs(vina_score) / 12.0 * 35 +
            md_score / 100.0 * 35 +
            30  # 基础分
        )

        candidates.append({
            'compound': compound_id,
            'vina_score': vina_score,
            'md_stability': md_score,
            'priority': round(priority_score, 1),
            'recommended_assay': 'Kinase activity' if
                                 priority_score > 75 else
                                 'Thermal shift'
        })

    candidates.sort(key=lambda x: x['priority'], reverse=True)

    print(f"\n{'Rank':<6} {'Compound':<15} {'Vina':<8} {'MD':<6} "
          f"{'Priority':<10} {'Suggested Assay'}")
    print("-" * 70)

    for i, cand in enumerate(candidates, 1):
        print(f"{i:<6} {cand['compound']:<15} {cand['vina_score']:<8.2f} "
              f"{cand['md_stability']:<6.1f} {cand['priority']:<10.1f} "
              f"{cand['recommended_assay']}")

    return candidates

prioritized_candidates = prioritize_for_experimental_validation(
    binding_analysis, md_validation)

# ---- 第8步: 结果可视化 ----

def plot_virtual_screening_summary():
    """生成虚拟筛选结果汇总图"""
    fig, axes = plt.subplots(2, 2, figsize=(14, 12))

    ax1 = axes[0, 0]
    all_scores = [r['score'] for r in screening_results]
    ax1.hist(all_scores, bins=50, color='steelblue', alpha=0.7,
             edgecolor='black')
    ax1.axvline(x=-9.0, color='red', linestyle='--',
                label='筛选阈值 (-9.0)')
    ax1.set_xlabel('Vina Score (kcal/mol)', fontsize=12)
    ax1.set_ylabel('化合物数量', fontsize=12)
    ax1.set_title('虚拟筛选打分分布', fontsize=14, fontweight='bold')
    ax1.legend()

    ax2 = axes[0, 1]
    top_compounds = [b['compound'] for b in binding_analysis[:5]]
    top_scores = [b['score'] for b in binding_analysis[:5]]
    colors = ['#2ecc71', '#3498db', '#9b59b6', '#e74c3c', '#f39c12']
    bars = ax2.barh(range(len(top_compounds)), top_scores,
                    color=colors, edgecolor='black')
    ax2.set_yticks(range(len(top_compounds)))
    ax2.set_yticklabels(top_compounds)
    ax2.set_xlabel('Vina Score (kcal/mol)', fontsize=12)
    ax2.set_title('Top 5 候选化合物', fontsize=14, fontweight='bold')
    for bar, score in zip(bars, top_scores):
        ax2.text(bar.get_width() + 0.1, bar.get_y() +
                 bar.get_height()/2, f'{score:.1f}',
                 va='center', fontsize=11, fontweight='bold')

    ax3 = axes[1, 0]
    compounds_md = [m['compound'] for m in md_validation]
    rmsd_vals = [m['avg_rmsd_Å'] for m in md_validation]
    hbond_vals = [m['hbond_occupancy'] for m in md_validation]

    x_pos = np.arange(len(compounds_md))
    width = 0.35

    bars1 = ax3.bar(x_pos - width/2, rmsd_vals, width,
                    label='RMSD (Å)', color='#e74c3c', alpha=0.8)
    ax3_twin = ax3.twinx()
    bars2 = ax3_twin.bar(x_pos + width/2,
                         [h*5 for h in hbond_vals], width,
                         label='HB Occupancy (×5)',
                         color='#3498db', alpha=0.8)

    ax3.set_xticks(x_pos)
    ax3.set_xticklabels(compounds_md, rotation=15)
    ax3.set_ylabel('RMSD (Å)', color='#e74c3c', fontsize=12)
    ax3_twin.set_ylabel('氢键占有率 (×5)', color='#3498db',
                        fontsize=12)
    ax3.set_title('MD模拟稳定性评估', fontsize=14, fontweight='bold')
    ax3.legend(loc='upper left')
    ax3_twin.legend(loc='upper right')

    ax4 = axes[1, 1]
    priorities = [c['priority'] for c in prioritized_candidates]
    names = [c['compound'] for c in prioritized_candidates]
    colors_prio = ['#27ae60' if p > 80 else '#f39c12'
                   if p > 70 else '#e74c3c' for p in priorities]

    scatter = ax4.scatter(range(len(names)), priorities,
                          s=[p*10 for p in priorities],
                          c=colors_prio, alpha=0.7, edgecolor='black')
    ax4.plot(range(len(names)), priorities, 'k--', alpha=0.3)

    for i, (name, prio) in enumerate(zip(names, priorities)):
        ax4.annotate(name, (i, prio), textcoords="offset points",
                     xytext=(0, 10), ha='center', fontsize=9,
                     fontweight='bold')

    ax4.set_xlabel('候选化合物', fontsize=12)
    ax4.set_ylabel('综合优先级评分', fontsize=12)
    ax4.set_title('实验验证优先级', fontsize=14, fontweight='bold')
    ax4.set_ylim(60, 105)
    ax4.axhline(y=80, color='green', linestyle=':',
                label='高优先级阈值')
    ax4.legend()

    plt.tight_layout()
    plt.savefig('virtual_screening_summary.png', dpi=300,
                bbox_inches='tight')
    plt.show()
    print("\n✓ 图像已保存: virtual_screening_summary.png")

plot_virtual_screening_summary()

# ---- 第9步: 输出最终推荐报告 ----

print("\n" + "="*80)
print("AlphaFold2 + 虚拟筛选 最终报告 - CDK12 抑制剂发现")
print("="*80)

print(f"\n📊 项目概况:")
print(f"   靶点: 利什曼原虫 CDK12 激酶结构域")
print(f"   结构来源: AlphaFold2 预测 (pLDDT > 90)")
print(f"   筛选库: ZINC drug-like 化合物 ({len(screening_results)} 个)")
print(f"   Top 候选: {len(prioritized_candidates)} 个化合物")

print(f"\n🎯 推荐实验验证:")
top_hit = prioritized_candidates[0]
print(f"   首选化合物: {top_hit['compound']}")
print(f"   对接打分: {top_hit['vina_score']} kcal/mol")
print(f"   综合评分: {top_hit['priority']}/100")
print(f"   建议实验: {top_hit['recommended_assay']} assay")

print(f"\n⚠️  注意事项:")
print("   • AlphaFold2预测结构需谨慎解释(特别是柔性环区)")
print("   • 建议先进行分子动力学长时间模拟(>100ns)确认稳定性")
print("   • 实验验证应包含:激酶活性IC50、细胞毒性、选择性谱")
print("   • 后续优化:基于复合物结构进行SAR和lead optimization")

print(f"\n💡 预期成果:")
print("   • 发现 IC50 < 1 μM 的苗头化合物")
print("   • 为利什曼病新药研发提供起点")
print("   • 展示AI辅助药物发现的完整工作流程")

生物学意义与临床应用价值

本案例展示了**人工智能驱动的药物发现完整工作流程**,从序列到结构、从虚拟筛选到实验验证,为缺乏实验结构的药物靶点提供了可行的解决方案:

  1. 突破结构瓶颈:AlphaFold2使得任何具有氨基酸序列的蛋白都可以获得高置信度的三维结构预测(pLDDT > 90),这对于寄生虫、细菌等病原体蛋白尤为重要,因为它们的实验结构数据通常非常有限。

  2. 加速先导化合物发现:传统的实验筛选需要合成或购买数千个化合物进行高通量筛选,耗时耗资巨大。基于结构的虚拟筛选可以在计算机上快速评估数百万个化合物的结合潜力,将实验验证范围缩小到数十个高优先级候选,大幅降低成本和时间。

  3. 提高选择性设计能力:通过分析激酶保守基序(铰链区、DFG motif、HRD催化环)的结合模式,可以理性设计对病原体激酶具有高度选择性而对人类激酶脱靶效应低的抑制剂,这是克服药物毒性的关键。

  4. 多维度验证降低假阳性:单纯依赖对接打分容易产生假阳性。本工作流整合了分子动力学模拟(评估复合物稳定性)、类药性分析(Lipinski规则)、激酶选择性预测等多个维度的信息,显著提高了后续实验验证的成功率。

  5. 转化医学前景:利什曼病作为被忽视的热带病,新药研发投入严重不足。这种计算驱动的策略可以以较低成本快速识别苗头化合物,为后续的先导优化和临床前研究提供起点,最终可能产生更安全、更有效的治疗方案。

  6. 方法学可推广性:该工作流程不仅适用于CDK12,还可推广到其他病原体关键酶(如蛋白酶、聚合酶等)的抑制剂发现,为抗感染药物研发提供通用技术平台。

技术要点总结

  • AlphaFold2 pLDDT评分 > 90 表示结构预测高度可靠,可用于基于结构的药物设计

  • AutoDock Vina穷举度(exhaustiveness)设为32可提高搜索全面性,但会增加计算时间

  • 结合能 < -9 kcal/mol 通常被认为是较强的结合

  • MD模拟中RMSD < 2.5 Å且氢键占有率 > 70%表示复合物稳定

  • 综合优先级评分应综合考虑对接打分(35%)、MD稳定性(35%)和基础可合成性(30%)

  • 实验验证应从体外生化 assay 开始,逐步推进到细胞水平和动物模型

---

## 本章小结

本章通过五个完整的实战案例,系统展示了蛋白质组学和代谢组学在生物医学研究中的广泛应用:

技术覆盖面: - 定量蛋白质组学(TMT标记、磷酸化富集) - 非靶向/靶向代谢组学(LC-MS、GC-MS) - 多组学数据整合(MOFA因子分析) - 空间组学技术(MALDI成像质谱) - 计算结构生物学(AlphaFold2 + 分子对接)

疾病应用领域: - 肿瘤生物学:乳腺癌信号通路动态、肝癌亚型分类 - 代谢性疾病:2型糖尿病生物标志物发现 - 神经退行性疾病:阿尔茨海默病空间分子病理学 - 感染性疾病:寄生虫病药物靶点发现

分析方法论: - 差异表达分析与统计检验(limma, t检验) - 多变量统计分析(PCA, OPLS-DA) - 通路富集与功能注释(KEGG, GO, Reactome) - 机器学习辅助鉴定(虚拟筛选、空间聚类) - 实验设计与验证策略(Western blot, ROC曲线)

核心启示: 1. 蛋白质组和代谢组学研究需要严谨的实验设计和质量控制 2. 多种技术的联合应用能够提供更全面的生物学视角 3. 计算分析必须与湿实验验证相结合才能得出可靠结论 4. 跨组学整合是揭示复杂疾病机制的有力工具 5. AI和机器学习正在革新生物医学研究的各个环节

这些案例不仅展示了具体的技术实现,更重要的是体现了如何将复杂的组学数据转化为有意义的生物学洞察,最终服务于精准医疗和药物发现的目标。