Add Logistic PCA (LPCA) analysis

.github/workflows/build-containers.yml

@@ -73,3 +73,8 @@ jobs:
       # of detected changes.
       always_build: true
       context: ./
+  build-and-remove-lpca:
+    uses: "./.github/workflows/build-and-remove-template.yml"
+    with:
+      path: docker-wrappers/LPCA
+      container: reedcompbio/lpca

Snakefile

@@ -91,6 +91,18 @@ def make_final_input(wildcards):
         final_input.extend(expand('{out_dir}{sep}{dataset}-ml{sep}ensemble-pathway.txt',out_dir=out_dir,sep=SEP,dataset=dataset_labels,algorithm_params=algorithms_with_params))
         final_input.extend(expand('{out_dir}{sep}{dataset}-ml{sep}jaccard-matrix.txt',out_dir=out_dir,sep=SEP,dataset=dataset_labels,algorithm_params=algorithms_with_params))
         final_input.extend(expand('{out_dir}{sep}{dataset}-ml{sep}jaccard-heatmap.png',out_dir=out_dir,sep=SEP,dataset=dataset_labels,algorithm_params=algorithms_with_params))
+
+    if _config.config.analysis_include_lpca:
+        final_input.extend(expand('{out_dir}{sep}{dataset}-ml{sep}lpca.png',out_dir=out_dir, sep=SEP, dataset=dataset_labels))
+        final_input.extend(expand('{out_dir}{sep}{dataset}-ml{sep}lpca-scores.csv',out_dir=out_dir, sep=SEP, dataset=dataset_labels))
+        final_input.extend(expand('{out_dir}{sep}{dataset}-ml{sep}lpca-coordinates.txt',out_dir=out_dir, sep=SEP, dataset=dataset_labels))
+        final_input.extend(expand('{out_dir}{sep}{dataset}-ml{sep}lpca-binary-matrix.csv',out_dir=out_dir, sep=SEP, dataset=dataset_labels))
+
+    if _config.config.analysis_include_lpca_aggregate_algo:
+        final_input.extend(expand('{out_dir}{sep}{dataset}-ml{sep}{algorithm}-lpca-scores.csv',out_dir=out_dir, sep=SEP, dataset=dataset_labels, algorithm=algorithms_mult_param_combos))
+        final_input.extend(expand('{out_dir}{sep}{dataset}-ml{sep}{algorithm}-lpca.png',out_dir=out_dir, sep=SEP,dataset=dataset_labels,algorithm=algorithms_mult_param_combos))
+        final_input.extend(expand('{out_dir}{sep}{dataset}-ml{sep}{algorithm}-lpca-coordinates.txt',out_dir=out_dir, sep=SEP, dataset=dataset_labels,algorithm=algorithms_mult_param_combos))
+        final_input.extend(expand('{out_dir}{sep}{dataset}-ml{sep}{algorithm}-lpca-binary-matrix.csv',out_dir=out_dir, sep=SEP, dataset=dataset_labels,algorithm=algorithms_mult_param_combos))
 
     if _config.config.analysis_include_ml_aggregate_algo:
         final_input.extend(expand('{out_dir}{sep}{dataset}-ml{sep}{algorithm}-pca.png',out_dir=out_dir,sep=SEP,dataset=dataset_labels,algorithm=algorithms_mult_param_combos))
@@ -356,6 +368,7 @@ rule ml_analysis:
         ml.hac_horizontal(summary_df, output.hac_image_horizontal, output.hac_clusters_horizontal, **hac_params)
         ml.pca(summary_df, output.pca_image, output.pca_variance, output.pca_coordinates, **pca_params)
 
+
 # Calculated Jaccard similarity between output pathways for each dataset
 rule jaccard_similarity:
     input:
@@ -403,6 +416,58 @@ rule ml_analysis_aggregate_algo:
         ml.hac_horizontal(summary_df, output.hac_image_horizontal, output.hac_clusters_horizontal, **hac_params)
         ml.pca(summary_df, output.pca_image, output.pca_variance, output.pca_coordinates, **pca_params)
 
+rule lpca_analysis_all:
+    input:
+        pathways = expand('{out_dir}{sep}{{dataset}}-{algorithm_params}{sep}pathway.txt', out_dir=out_dir, sep=SEP, algorithm_params=algorithms_with_params)
+    output:
+        lpca_scores = SEP.join([out_dir, '{dataset}-ml', 'lpca-scores.csv']),
+        lpca_png = SEP.join([out_dir, '{dataset}-ml', 'lpca.png']),
+        lpca_coord = SEP.join([out_dir, '{dataset}-ml', 'lpca-coordinates.txt']),
+        lpca_matrix = SEP.join([out_dir, '{dataset}-ml', 'lpca-binary-matrix.csv'])
+    run:
+        from spras.analysis import lpca
+        summary_df = ml.summarize_networks(input.pathways)
+        lpca.run_lpca(
+            summary_df,
+            output.lpca_scores,
+            output.lpca_matrix,
+            k=_config.config.lpca_params.k,
+            m=_config.config.lpca_params.m,
+            cv=_config.config.lpca_params.cv,
+            container_settings=container_settings
+        )
+        lpca.plot_lpca(
+            output.lpca_scores,
+            output.lpca_png,
+            output.lpca_coord
+        )
+
+rule lpca_analysis_aggregate_algo:
+    input:
+        pathways = collect_pathways_per_algo
+    output:
+        lpca_scores = SEP.join([out_dir, '{dataset}-ml', '{algorithm}-lpca-scores.csv']),
+        lpca_png = SEP.join([out_dir, '{dataset}-ml', '{algorithm}-lpca.png']),
+        lpca_coord = SEP.join([out_dir, '{dataset}-ml', '{algorithm}-lpca-coordinates.txt']),
+        lpca_matrix = SEP.join([out_dir, '{dataset}-ml', '{algorithm}-lpca-binary-matrix.csv'])
+    run:
+        from spras.analysis import lpca
+        summary_df = ml.summarize_networks(input.pathways)
+        lpca.run_lpca(
+            summary_df,
+            output.lpca_scores,
+            output.lpca_matrix,
+            k=_config.config.lpca_params.k,
+            m=_config.config.lpca_params.m,
+            cv=_config.config.lpca_params.cv,
+            container_settings=container_settings
+        )
+        lpca.plot_lpca(
+            output.lpca_scores,
+            output.lpca_png,
+            output.lpca_coord
+        )
+
 # Ensemble the output pathways for each dataset per algorithm
 rule ensemble_per_algo:
     input:

config/config.yaml

@@ -272,3 +272,17 @@ analysis:
     # adds evaluation per algorithm per dataset-goldstandard pair
     # evaluation per algorithm will not run unless ml include and ml aggregate_per_algorithm are set to true
     aggregate_per_algorithm: true
+  lpca:
+    # if true, runs LPCA in addition to the existing PCA (both analyses will run in parallel).
+    # Running both helps compare the two methods on the same data.
+    include: false
+    # adds separate LPCA analysis for each algorithm that has multiple parameter combinations
+    aggregate_per_algorithm: false
+    # number of principal components to compute
+    k: 2
+    # fixed value of the logisticPCA tuning parameter m, used when cv is false.
+    # default of 6 was selected based on cross-validation experiments across multiple
+    # SPRAS algorithms (see https://github.com/Jeebjean/lpca-spras for details)
+    m: 6
+    # if true, choose m by cross-validation; if false, use the fixed m above
+    cv: false

docker-wrappers/LPCA/Dockerfile

@@ -0,0 +1,31 @@
+# Logistic PCA (logisticPCA) wrapper for SPRAS.
+# Invoked by spras/analysis/lpca.py, which supplies the full command
+# (Rscript /app/run_lpca.R ... or /app/run_cv.R ...), so no ENTRYPOINT is set.
+
+# Pinned R version for reproducibility. Bump deliberately, not to :latest.
+FROM rocker/r-base:4.4.2
+
+LABEL org.opencontainers.image.source="https://github.com/Reed-CompBio/spras"
+LABEL org.opencontainers.image.description="Logistic PCA (logisticPCA) wrapper for SPRAS"
+
+# System libraries needed to compile ggplot2 (a hard Import of logisticPCA)
+# and its dependency stack from source on Debian.
+RUN apt-get update && apt-get install -y --no-install-recommends \
+        libcurl4-openssl-dev \
+        libssl-dev \
+        libxml2-dev \
+        libfontconfig1-dev \
+        libfreetype6-dev \
+        libpng-dev \
+        libtiff5-dev \
+        libjpeg-dev \
+    && rm -rf /var/lib/apt/lists/*
+
+# logisticPCA is still on CRAN (last published 2016) and pulls in ggplot2.
+RUN Rscript -e "install.packages(c('logisticPCA', 'rARPACK'), repos='https://cran.r-project.org')" \
+    && Rscript -e "library(logisticPCA); library(rARPACK)"
+
+COPY run_lpca.R /app/run_lpca.R
+COPY run_cv.R /app/run_cv.R
+
+WORKDIR /app

docker-wrappers/LPCA/README.md

@@ -0,0 +1,69 @@
+# LPCA (Logistic PCA) wrapper
+
+Docker image: https://hub.docker.com/r/reedcompbio/lpca
+
+This wrapper runs [logisticPCA](https://github.com/andland/logisticPCA)
+
+([Landgraf & Lee, 2020](https://doi.org/10.1016/j.jmva.2020.104668)) as a SPRAS analysis step. It reduces the binary
+edge-by-run matrix built from a set of pathway reconstruction outputs to a small
+number of components and reports the proportion of deviance explained.
+
+The analysis is driven by the `analysis.lpca` config block and the
+`lpca_analysis` Snakemake rule, and is implemented in `spras/analysis/lpca.py`.
+
+## Configuration
+
+    analysis:
+      lpca:
+        include: false   # run the LPCA analysis per algorithm
+        k: 2             # number of principal components
+        m: 6             # fixed logisticPCA tuning parameter, used when cv is false
+        cv: false        # true: choose m by cross-validation; false: use the fixed m
+
+LPCA only runs for algorithms with multiple parameter combinations, so that the
+binary matrix has more than one column. It also needs a reasonable number of
+observations to be meaningful; very small inputs (such as the bundled example
+datasets) produce degenerate results, which is why it is disabled by default.
+
+  ### `partial_decomp`
+
+  The LPCA wrapper always runs `logisticSVD` with `partial_decomp = TRUE`, which
+  uses a truncated (rARPACK-based) decomposition instead of a full one. This is
+  hardcoded rather than exposed as a parameter:
+
+  - On small datasets it has no practical effect on the result.
+  - On large datasets it is required to avoid out-of-memory (OOMKilled) errors
+    that occur with the full decomposition.
+
+  Because it is beneficial on large inputs and harmless on small ones, it is
+  enabled unconditionally and is not a user-facing configuration option.
+
+## Scripts
+
+The image contains two R scripts under `/app`:
+
+- `run_lpca.R <input> <output> <k> <m>`: runs logisticPCA with a fixed `m` and
+  writes the scores CSV plus a sibling `<output basename>_deviance.txt`.
+- `run_cv.R <input> <output> <k>`: cross-validates `m` over 1..20 and writes a
+  CSV with a `best_m` column (plus a `_curve.csv` with the full CV curve). Only
+  used when `cv: true`.
+
+Both read a CSV whose first column holds row labels and whose remaining columns
+are binary (0/1) features, and coerce missing values to 0.
+
+## Dependency note
+
+`logisticPCA` declares `ggplot2` as a hard `Imports` dependency, so building the
+image compiles the ggplot2 stack. The Dockerfile installs the required Debian
+system libraries for that. To build from a source CRAN mirror the `repos`
+argument already points at `https://cran.r-project.org`.
+
+## Building and publishing the image
+
+For the SPRAS default registry to resolve the image, it must be published as
+`docker.io/reedcompbio/lpca:v1`, which requires access to the `reedcompbio`
+Docker Hub organization:
+
+    docker build -t reedcompbio/lpca:v1 docker-wrappers/lpca/
+    docker push reedcompbio/lpca:v1
+

docker-wrappers/LPCA/run_cv.R

@@ -0,0 +1,36 @@
+# run_cv.R
+# Finds the optimal m for a given k using cross-validation
+
+args = commandArgs(trailingOnly = TRUE)
+input_file = args[1]
+output_file = args[2]
+k = as.integer(args[3])
+
+set.seed(42)
+
+# Load data
+library(logisticPCA)
+data = read.csv(input_file, row.names = NULL)
+data = data[, -1]
+data_matrix = as.matrix(data)
+data_matrix[is.na(data_matrix)] = 0
+
+# Cross-validation over m, fixed k
+cv_result = cv.lpca(data_matrix, ks = k, ms = 1:20)
+best_m = which.min(cv_result)
+
+cat("Cross-validation done for k =", k, "\n")
+cat("Best m:", best_m, "\n")
+
+# Save best m
+write.csv(data.frame(k = k, best_m = best_m), output_file, row.names = FALSE)
+
+# Save full CV curve (all m values and their reconstruction error)
+cv_curve_file = sub("\\.csv$", "_curve.csv", output_file)
+cv_df = data.frame(
+  m = 1:20,
+  reconstruction_error = as.numeric(cv_result),
+  is_best = (1:20) == best_m
+)
+write.csv(cv_df, cv_curve_file, row.names = FALSE)
+cat("CV curve saved to", cv_curve_file, "\n")

docker-wrappers/LPCA/run_lpca.R

@@ -0,0 +1,30 @@
+# Read command line arguments
+args = commandArgs(trailingOnly = TRUE)
+input_file = args[1]
+output_file = args[2]
+k = as.integer(args[3])
+m = as.numeric(args[4])
+
+# Load data
+library(logisticPCA)
+data = read.csv(input_file, row.names = NULL)
+row_labels = data[, 1]
+data = data[, -1]
+data_matrix = as.matrix(data)
+data_matrix[is.na(data_matrix)] = 0
+
+# Run LPCA
+model = logisticPCA(data_matrix, k = k, m = m, partial_decomp = TRUE)
+
+# Save scores
+scores = model$PCs
+rownames(scores) = row_labels
+write.csv(scores, output_file, row.names = TRUE)
+
+# Save deviance explained
+deviance_file = sub("\\.csv$", "_deviance.txt", output_file)
+writeLines(as.character(model$prop_deviance_expl), deviance_file)
+
+cat("LPCA done! Scores saved to", output_file, "\n")
+cat("Score dimensions:", nrow(scores), "x", ncol(scores), "\n")
+cat("Proportion of deviance explained:", model$prop_deviance_expl, "\n")