Skip to content

README.md

Latest commit

 

History

History
144 lines (108 loc) · 6.33 KB

README.md

File metadata and controls

144 lines (108 loc) · 6.33 KB
Raw

Modern SubMap (R Implementation)

Core Algorithm: Hoshida Y, et al. "Subclass Mapping: Identifying Common Subtypes in Independent Disease Data Sets". PLoS ONE (2007).

Original Script: submap.R by Broad Institute (Oct. 14, 2008).

The original submap.R relies on the specific old file formats of GenePattern (.gct, .cls), which leads to low running efficiency (with many loops) and cannot directly interface with modern R analysis workflows (like Seurat, Biobase).

This project has undergone Modern Refactoring of the original code, fundamentally improving performance and usability while maintaining complete consistency with the mathematical core.

1. Why Refactor

Pain Points

  • Dependency on File I/O: Data must be written out as .gct and .cls text files to run, which is very cumbersome.
  • Slow Execution: The original script was designed for data scales from 2008, using many inefficient R loops (for loops), making it run extremely slowly on today's single-cell or large cohort data.
  • Memory Overflow: The old version stored large 3D arrays during permutation tests, which easily leads to memory crashes.
  • Outdated Visualization: Utilizes basic plotting systems, making it difficult to generate publication-quality images.

Modern Features

  • Memory Pass-Through: Directly accepts R Matrix (expression profiles) and Vector (grouping) inputs, without the need for any temporary files.
  • Blazing Performance: * Introduced matrixStats (C++) for matrix operations, speeding up 10-50 times. * Built-in parallel multi-core computing support.
  • Result Objectification: Results are returned as a clean List object, containing all P-value matrices and intermediate statistics.
  • Visualization Upgrade: Automatically generates publication-quality heatmaps with significance annotations based on ComplexHeatmap.

2. Deep Comparison

Feature Original Version (2008) Modern Version (2025) Notes
Algorithm Core KS Statistic + Fisher's Combination KS Statistic + Fisher's Combination Core algorithm is identical
Input Method File Path (.gct, .cls) R Objects (Matrix, Vector) No disk I/O required
S2N Calculation Double Loop + apply matrixStats::rowMeans2/rowSds 20x+ faster
Permutation Test Single-threaded mclapply Multi-core Parallel Default 8-core parallel
Memory Usage Extremely high (stores all Permutations for file writing) Extremely low (only returns Pool statistics) Avoids large memory crashes
Plotting Engine Base Graphics ComplexHeatmap Automatic legends and annotations
Code Volume ~1000 lines ~350 lines Removed a lot of redundant code

3. Usage

3.0 Data Preparation Requirements

  • Expression matrices must be numeric with genes in rows and samples in columns; rownames should be gene IDs and colnames sample IDs.
  • Gene identifiers must match between the two datasets; the shared genes after intersect() should be at least n_markers * 2.
  • Avoid duplicated gene IDs and drop any rows with all-zero or missing values to keep S2N stable.
  • Group vectors (group_a, group_b) must align with column order, have the same length as the number of samples, and contain >=2 classes; keep class sizes reasonable (avoid groups of size 1).
  • Provide log-normalized or otherwise comparable expression values across datasets; do not mix units (e.g., TPM vs raw counts) without normalization.

3.1 Core Function

Simply load submap_modern.R to use.

source("submap_modern.R")

# Prepare data (Standard R objects)
# exp_ref:   Gene x Sample matrix (Reference)
# group_ref: Sample grouping vector
# exp_query: Gene x Sample matrix (Query)
# group_query: Sample grouping vector (Query)

# Run analysis
res <- run_submap(
  exp_a = exp_ref,    group_a = group_ref,     # Reference Dataset
  exp_b = exp_query,  group_b = group_query,   # Query Dataset
  n_markers = 100,    # Recommended 100-300
  n_perm = 1000,      # Recommended 1000 permutations
  n_cores = 8,        # Number of parallel cores
  weighted_score_type = 1 # 1=Weighted (Recommended), 0=Unweighted
)

write.csv(res$sa_p_matrix,    file = file.path(out_dir, "01.SA_P_value_matrix.csv"))
write.csv(res$sa_fdr_matrix,  file = file.path(out_dir, "02.SA_FDR_matrix.csv"))
write.csv(res$sa_bonf_matrix, file = file.path(out_dir, "03.SA_Bonferroni_matrix.csv"))
write.csv(res$fisher_stats,   file = file.path(out_dir, "04.Fisher_Stats_matrix.csv"))

# Result visualization
plot_submap_heatmap(res, 
                    row_title = "Reference Group", 
                    col_title = "Query Group",
                    correction_method = "Bonferroni") # Optional FDR/Bonferroni

3.2 Minimal Runnable Example (Toy Data)

library(Matrix)

set.seed(123)
genes <- paste0("G", sprintf("%04d", 1:500))

# Reference dataset: 500 genes x 20 samples, two classes
exp_ref <- matrix(rnorm(length(genes) * 20, mean = 0, sd = 1),
                  nrow = length(genes),
                  dimnames = list(genes, paste0("R", 1:20)))
group_ref <- rep(c("Ref_A", "Ref_B"), each = 10)

# Query dataset: 500 genes x 18 samples, two classes
exp_query <- matrix(rnorm(length(genes) * 18, mean = 0, sd = 1),
                    nrow = length(genes),
                    dimnames = list(genes, paste0("Q", 1:18)))
group_query <- rep(c("Qry_X", "Qry_Y"), each = 9)

# Optional: convert to Matrix::Matrix for memory efficiency
exp_ref <- Matrix(exp_ref, sparse = TRUE)
exp_query <- Matrix(exp_query, sparse = TRUE)

source("submap_modern.R")

res <- run_submap(
  exp_a = exp_ref,    group_a = group_ref,
  exp_b = exp_query,  group_b = group_query,
  n_markers = 50,     # small toy example
  n_perm = 200,       # increase (e.g., 1000) for real analysis
  n_cores = 4,
  weighted_score_type = 1
)

# Inspect and plot
str(res)
plot_submap_heatmap(res, row_title = "Ref", col_title = "Query")

4. Remarks

The above README is generated by Gemini 3 pro.

As this project does not have significant relevance to my own work, I currently do not have more energy to devote to it. Suggestions and questions are welcome.

Original Author Credit:

  • Hoshida Y, Brunet JP, Tamayo P, Golub TR, Mesirov JP
  • Broad Institute of MIT and Harvard University

Modern Refactor Author:

  • Yuanwei
  • Guizhou University