pycmplot 0.3.0__tar.gz → 0.4.0__tar.gz

{pycmplot-0.3.0 → pycmplot-0.4.0}/PKG-INFO

@@ -1,6 +1,6 @@
 Metadata-Version: 2.4
 Name: pycmplot
-Version: 0.3.0
+Version: 0.4.0
 Summary: Multi-track circular and linear Manhattan plot generation for GWAS summary statistics
 Author: Kevin Esoh
 Author-email: Kevin Esoh <kesohku1@jh.edu>
@@ -39,16 +39,60 @@ Dynamic: license-file
 
 # pycmplot
 
-Multi-track **circular** and **linear** Manhattan plot generation for GWAS summary statistics.
+<div>
+  <h1 align="center">Multi-track circular and linear Manhattan plotting in Python.</h1>
+</div>
+
+<p align="center">
+  <img width="600" height="400" src="https://github.com/esohkevin/pycmplot/blob/main/docs/pycmplot-logo-circular.png">
+</p>
+
 
 ```
-#~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~#
-|  PACKAGE FOR CIRCULAR AND LINEAR MANHATTAN PLOTTING  |
-|                    Kevin Esoh, 2026                  |
-|                    kesohku1@jh.edu                   |
-#~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~#
+                    #~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~#
+                    |  MULTI-TRACK CIRCULAR AND LINEAR MANHATTAN PLOTTING  |
+                    |                      in Python                       |
+                    |                    Kevin Esoh, 2026                  |
+                    |                    kesohku1@jh.edu                   |
+                    #~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~#
 ```
 
+----------------------------------------------
+
+## Content
+1. [Overview](#overview)
+2. [Key features](#key-features)
+    - [Column auto-detection](#column-auto-detection)
+    - [Density-aware sub-sampling](#density-aware-sub-sampling)
+    - [Trim insignificant variants for faster plotting](#trim-insignificant-variants-for-faster-plotting)
+    - [Genome build conversion (liftover)](#genome-build-conversion-liftover)
+    - [Nearest-gene annotation for GWAS lead SNPs](#nearest-gene-annotation-for-gwas-lead-snps)
+    - [Caching & warm resume](#caching--warm-resume)
+    - [User-editable hits overlay](#user-editable-hits-overlay)
+    - [Per-locus highlight colours & custom legend](#per-locus-highlight-colours--custom-legend)
+    - [Multi-panel canvas](#multi-panel-canvas)
+3. [Application](#application)
+4. [Tip](#tip)
+5. [Installation](#installation)
+    - [From PyPI](#from-pypi)
+    - [From GitHub](#from-github)
+    - [Python virtual environment](#python-virtual-environment)
+6. [Test the installation](#test-the-installation)
+7. [Dependencies](#dependencies)
+8. [Command-line usage](#command-line-usage)
+    - [Linear Manhattan (default)](#linear-manhattan-default)
+    - [Circular Manhattan](#circular-manhattan)
+    - [Key options](#key-options)
+9. [Python API](#python-api)
+10. [Contributing](#contributing)
+11. [Contributors](#contributors)
+
+
+---------------------------------------------
+
+
+## Overview 
+
 This package will take any number of per SNP/variant summary statistics, be it GWAS, 
 selection scans (e.g. iHS, EHH, FST), etc and generate Manhattan plots. If given a single
 file, a single one-track Manhattan plot will be generated. Multiple files will result in 
@@ -68,8 +112,8 @@ option of the package should be used to indicate the column and then the package
 postions in hg19 to hg38 ensuring that hits table generation and plotting are done with one unified 
 corrdinate system.
 
-# Key features
-## Column auto-detection
+## Key features
+### Column auto-detection
 A key functionality of the package is its ability to auto-detect certain columns if ommited on the 
 command-line or python API:
 - Chromosome column: `-chr, --chrom_column` or ommited
@@ -92,7 +136,7 @@ bld_candidates = [build, 'BUILD', 'Genome', 'Genome_Build', 'Genome-build']
 
 > NB: Upper and lower cases of the candidates are also considered, making each candidate expanded 3 times.
 
-## Density-aware sub-sampling
+### Density-aware sub-sampling
 Another key feature is density-aware sub-sampling for Manhattan-style scatter plots.
 This was inspired by ``gwaslab``'s default behaviour (https://cloufield.github.io/gwaslab/). 
 
@@ -103,14 +147,14 @@ scan with the defaults below, this typically cuts the plotted point count from 1
 to ~200 K + a few hundred peaks — visually indistinguishable above the suggestive 
 band, but two orders of magnitude faster to render.
 
-## Trim insignificant variants for faster plotting
+### Trim insignificant variants for faster plotting
 An optional parameter `-tp, --trim_pval` is provided to increase speed even further. 
 Set with a value to exclude variants with p-value above a certain threshold, 
 e.g. `0.01 (1e-2)` or `0.001 (1e-3)`. Performed on top of the default auto-thin 
 feature above, it siginificant increases speed and reduces peak memory usage. 
 See benchmark figure (manuscript in preparation).
 
-## Genome build conversion (liftover)
+### Genome build conversion (liftover)
 Conversion of a both hg18 and hg19 positions to their hg38 equivalent is included through
 `pyliftover.LiftOver`.
 
@@ -120,7 +164,7 @@ column to specify the genome build of each position ('hg18', 'hg19', or 'hg38')
 using one coordinate system. If only 'hg18' or 'hg19' positions are present, no liftover 
 be necessary. Hence, liftover is only performed in cases of mixed genome builds.
 
-## Nearest-gene annotation for GWAS lead SNPs
+### Nearest-gene annotation for GWAS lead SNPs
 The package bundles GFF3 files in hg19 and hg38 coordinates processed to reduce size 
 for gene annotation. Also included are UCSC chain files for coordinate conversion (liftover).
   - ``chain_hg19_hg38`` -- UCSC LiftOver chain file for hg19 to hg38
@@ -139,12 +183,52 @@ for gene annotation. Also included are UCSC chain files for coordinate conversio
     ``Homo_sapiens.GRCh37.geneinfo.tsv.gz``.
 
 
-# Application
+### Caching & warm resume
+Loading is the expensive step (I/O + trim + liftover + lead extraction), so
+pycmplot ships a per-track cache keyed on `SHA-256(raw_file_sha256 + version + Stage-1 params)`.
+Warm re-runs of the same `(files, parameters)` combination complete in
+milliseconds; changing any parameter transparently invalidates only the
+affected tracks and regenerates them. Enable with `--cache` (CLI) or
+`cache=True` (Python API); use `--clear_cache` to wipe the tree.
+
+```bash
+pycmplot --sum_stats hb.tsv,mcv.tsv --labels Hb,MCV \
+         --logp --highlight --cache --cache_dir ./.pycmplot
+```
+
+### User-editable hits overlay
+When caching is on, the auto-generated hits table is written to a
+group-scoped TSV at `<cache_dir>/annotations/hits.<group_key>.tsv`
+that you're *expected to hand-edit*. Rows added or changed there feed
+straight back into the next plot — no re-plumbing required. Each row
+carries `source` (`auto` vs `user`), `highlight_color`, and `category`
+columns; user edits are inherited across cache regenerations by
+`(CHR, POS)` lookup, so re-running with a new parameter never loses
+your annotations.
+
+### Per-locus highlight colours & custom legend
+Set the `highlight_color` column on any row of the hits overlay to a
+matplotlib-parseable colour (name, `#rrggbb`, or an RGB tuple) to give
+that locus its own highlight colour; leave `auto` to fall back to the
+plot-wide `--highlight_color`. Set the `category` column
+(e.g. `novel`, `replicated`, `MHC`) to group loci in a
+**"Highlighted Categories"** legend that both the linear and circular
+plotters render automatically. If nothing is edited, no legend is
+added — the pre-feature layout is preserved.
+
+### Multi-panel canvas
+Place multiple *groups* of sumstats on the same figure by passing an
+explicit matplotlib `Axes` or `SubFigure` via `ax=` to the plotter.
+Cache files and hits overlays are group-scoped, so two panels with
+different sumstats never clobber each other's artefacts.
+
+
+## Application
 A potential useful application is **comparative visualization** of results from multiple imputation panels, 
 multiple populations, or multiple traits to observe shared genetic architecture.
 
 
-# Tip
+## Tip
 
 For signed non-pvalue statistics such as iHS, use `--signif_threshold` and 
 `--suggest_threshold` to specify lines to draw for iHS cutoff above and below zero, 
@@ -153,7 +237,7 @@ that is in the positive and negative axes.
 
 Read more in the package documentation page: https://pycmplot.readthedocs.io/en/latest/
 
----
+----------------------------
 
 ## Installation
 
@@ -180,24 +264,25 @@ pip install -e . --break-system-packages
 ```
 
 
-### Use python virtual environment if local installation is not possible
-```bash
-python -m venv ~/bin/pycmplot
+### Python virtual environment
 
-source ~/bin/pycmplot/bin/activate
+Use python virtual environment if local installation is not possible
 
-pip install --upgrade pip setuptools wheel
+```bash
+python -m venv ~/envs/pycmplot
+source ~/envs/pycmplot/bin/activate
 
-# then follow any of the installation steps above
+pip install --upgrade pip setuptools wheel
+pip install pycmplot
 ```
 
 
-# Test the installation
+## Test the installation
 ```bash
 pycmplot -h
 ```
 
-### Dependencies
+## Dependencies
 
 | Package | Purpose |
 |---------|---------|
@@ -264,6 +349,16 @@ pycmplot \
 | `-st, --sort_track` | Sort tracks by `label` or `chrom_len` | input order |
 | `-od, --output_dir` | Output directory | `.` |
 | `-of, --output_format` | Output format (`png`, `pdf`, `svg`, `jpg`) | `png` |
+| `--cache` | Enable per-track cache + user-editable hits overlay | off |
+| `--cache_dir` | Where to store cache artefacts | `./.pycmplot` |
+| `--no_resume` | Regenerate but still write fresh cache entries | resume on |
+| `--clear_cache` | Delete the cache tree and exit | off |
+| `-V, --version` | Print version and exit | — |
+
+> **QQ note (0.4.0+):** the loader no longer materialises the full sorted
+> p-value array by default. `-qq/--qq_plot` sets `compute_pvals=True`
+> automatically on the CLI; Python-API callers must pass it explicitly
+> to feed `bundle['pvals']` into a QQ plotter.
 
 Run `pycmplot -h` for the full option list.
 
@@ -273,6 +368,12 @@ Run `pycmplot -h` for the full option list.
 
 A demonstration of how to use the python API is provided in this notebook: https://github.com/esohkevin/pycmplot/blob/main/pycmplot_python_api.ipynb
 
+For an end-to-end walkthrough of every feature (caching, hits overlay,
+per-locus colours & categories, multi-panel canvas, mixed-build
+liftover, QQ opt-in), see the
+[Tutorial](https://pycmplot.readthedocs.io/en/latest/tutorial.html)
+page in the docs.
+
 
 ---
 

{pycmplot-0.3.0 → pycmplot-0.4.0}/README.md

@@ -1,15 +1,59 @@
 # pycmplot
 
-Multi-track **circular** and **linear** Manhattan plot generation for GWAS summary statistics.
+<div>
+  <h1 align="center">Multi-track circular and linear Manhattan plotting in Python.</h1>
+</div>
+
+<p align="center">
+  <img width="600" height="400" src="https://github.com/esohkevin/pycmplot/blob/main/docs/pycmplot-logo-circular.png">
+</p>
+
 
 ```
-#~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~#
-|  PACKAGE FOR CIRCULAR AND LINEAR MANHATTAN PLOTTING  |
-|                    Kevin Esoh, 2026                  |
-|                    kesohku1@jh.edu                   |
-#~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~#
+                    #~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~#
+                    |  MULTI-TRACK CIRCULAR AND LINEAR MANHATTAN PLOTTING  |
+                    |                      in Python                       |
+                    |                    Kevin Esoh, 2026                  |
+                    |                    kesohku1@jh.edu                   |
+                    #~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~#
 ```
 
+----------------------------------------------
+
+## Content
+1. [Overview](#overview)
+2. [Key features](#key-features)
+    - [Column auto-detection](#column-auto-detection)
+    - [Density-aware sub-sampling](#density-aware-sub-sampling)
+    - [Trim insignificant variants for faster plotting](#trim-insignificant-variants-for-faster-plotting)
+    - [Genome build conversion (liftover)](#genome-build-conversion-liftover)
+    - [Nearest-gene annotation for GWAS lead SNPs](#nearest-gene-annotation-for-gwas-lead-snps)
+    - [Caching & warm resume](#caching--warm-resume)
+    - [User-editable hits overlay](#user-editable-hits-overlay)
+    - [Per-locus highlight colours & custom legend](#per-locus-highlight-colours--custom-legend)
+    - [Multi-panel canvas](#multi-panel-canvas)
+3. [Application](#application)
+4. [Tip](#tip)
+5. [Installation](#installation)
+    - [From PyPI](#from-pypi)
+    - [From GitHub](#from-github)
+    - [Python virtual environment](#python-virtual-environment)
+6. [Test the installation](#test-the-installation)
+7. [Dependencies](#dependencies)
+8. [Command-line usage](#command-line-usage)
+    - [Linear Manhattan (default)](#linear-manhattan-default)
+    - [Circular Manhattan](#circular-manhattan)
+    - [Key options](#key-options)
+9. [Python API](#python-api)
+10. [Contributing](#contributing)
+11. [Contributors](#contributors)
+
+
+---------------------------------------------
+
+
+## Overview 
+
 This package will take any number of per SNP/variant summary statistics, be it GWAS, 
 selection scans (e.g. iHS, EHH, FST), etc and generate Manhattan plots. If given a single
 file, a single one-track Manhattan plot will be generated. Multiple files will result in 
@@ -29,8 +73,8 @@ option of the package should be used to indicate the column and then the package
 postions in hg19 to hg38 ensuring that hits table generation and plotting are done with one unified 
 corrdinate system.
 
-# Key features
-## Column auto-detection
+## Key features
+### Column auto-detection
 A key functionality of the package is its ability to auto-detect certain columns if ommited on the 
 command-line or python API:
 - Chromosome column: `-chr, --chrom_column` or ommited
@@ -53,7 +97,7 @@ bld_candidates = [build, 'BUILD', 'Genome', 'Genome_Build', 'Genome-build']
 
 > NB: Upper and lower cases of the candidates are also considered, making each candidate expanded 3 times.
 
-## Density-aware sub-sampling
+### Density-aware sub-sampling
 Another key feature is density-aware sub-sampling for Manhattan-style scatter plots.
 This was inspired by ``gwaslab``'s default behaviour (https://cloufield.github.io/gwaslab/). 
 
@@ -64,14 +108,14 @@ scan with the defaults below, this typically cuts the plotted point count from 1
 to ~200 K + a few hundred peaks — visually indistinguishable above the suggestive 
 band, but two orders of magnitude faster to render.
 
-## Trim insignificant variants for faster plotting
+### Trim insignificant variants for faster plotting
 An optional parameter `-tp, --trim_pval` is provided to increase speed even further. 
 Set with a value to exclude variants with p-value above a certain threshold, 
 e.g. `0.01 (1e-2)` or `0.001 (1e-3)`. Performed on top of the default auto-thin 
 feature above, it siginificant increases speed and reduces peak memory usage. 
 See benchmark figure (manuscript in preparation).
 
-## Genome build conversion (liftover)
+### Genome build conversion (liftover)
 Conversion of a both hg18 and hg19 positions to their hg38 equivalent is included through
 `pyliftover.LiftOver`.
 
@@ -81,7 +125,7 @@ column to specify the genome build of each position ('hg18', 'hg19', or 'hg38')
 using one coordinate system. If only 'hg18' or 'hg19' positions are present, no liftover 
 be necessary. Hence, liftover is only performed in cases of mixed genome builds.
 
-## Nearest-gene annotation for GWAS lead SNPs
+### Nearest-gene annotation for GWAS lead SNPs
 The package bundles GFF3 files in hg19 and hg38 coordinates processed to reduce size 
 for gene annotation. Also included are UCSC chain files for coordinate conversion (liftover).
   - ``chain_hg19_hg38`` -- UCSC LiftOver chain file for hg19 to hg38
@@ -100,12 +144,52 @@ for gene annotation. Also included are UCSC chain files for coordinate conversio
     ``Homo_sapiens.GRCh37.geneinfo.tsv.gz``.
 
 
-# Application
+### Caching & warm resume
+Loading is the expensive step (I/O + trim + liftover + lead extraction), so
+pycmplot ships a per-track cache keyed on `SHA-256(raw_file_sha256 + version + Stage-1 params)`.
+Warm re-runs of the same `(files, parameters)` combination complete in
+milliseconds; changing any parameter transparently invalidates only the
+affected tracks and regenerates them. Enable with `--cache` (CLI) or
+`cache=True` (Python API); use `--clear_cache` to wipe the tree.
+
+```bash
+pycmplot --sum_stats hb.tsv,mcv.tsv --labels Hb,MCV \
+         --logp --highlight --cache --cache_dir ./.pycmplot
+```
+
+### User-editable hits overlay
+When caching is on, the auto-generated hits table is written to a
+group-scoped TSV at `<cache_dir>/annotations/hits.<group_key>.tsv`
+that you're *expected to hand-edit*. Rows added or changed there feed
+straight back into the next plot — no re-plumbing required. Each row
+carries `source` (`auto` vs `user`), `highlight_color`, and `category`
+columns; user edits are inherited across cache regenerations by
+`(CHR, POS)` lookup, so re-running with a new parameter never loses
+your annotations.
+
+### Per-locus highlight colours & custom legend
+Set the `highlight_color` column on any row of the hits overlay to a
+matplotlib-parseable colour (name, `#rrggbb`, or an RGB tuple) to give
+that locus its own highlight colour; leave `auto` to fall back to the
+plot-wide `--highlight_color`. Set the `category` column
+(e.g. `novel`, `replicated`, `MHC`) to group loci in a
+**"Highlighted Categories"** legend that both the linear and circular
+plotters render automatically. If nothing is edited, no legend is
+added — the pre-feature layout is preserved.
+
+### Multi-panel canvas
+Place multiple *groups* of sumstats on the same figure by passing an
+explicit matplotlib `Axes` or `SubFigure` via `ax=` to the plotter.
+Cache files and hits overlays are group-scoped, so two panels with
+different sumstats never clobber each other's artefacts.
+
+
+## Application
 A potential useful application is **comparative visualization** of results from multiple imputation panels, 
 multiple populations, or multiple traits to observe shared genetic architecture.
 
 
-# Tip
+## Tip
 
 For signed non-pvalue statistics such as iHS, use `--signif_threshold` and 
 `--suggest_threshold` to specify lines to draw for iHS cutoff above and below zero, 
@@ -114,7 +198,7 @@ that is in the positive and negative axes.
 
 Read more in the package documentation page: https://pycmplot.readthedocs.io/en/latest/
 
----
+----------------------------
 
 ## Installation
 
@@ -141,24 +225,25 @@ pip install -e . --break-system-packages
 ```
 
 
-### Use python virtual environment if local installation is not possible
-```bash
-python -m venv ~/bin/pycmplot
+### Python virtual environment
 
-source ~/bin/pycmplot/bin/activate
+Use python virtual environment if local installation is not possible
 
-pip install --upgrade pip setuptools wheel
+```bash
+python -m venv ~/envs/pycmplot
+source ~/envs/pycmplot/bin/activate
 
-# then follow any of the installation steps above
+pip install --upgrade pip setuptools wheel
+pip install pycmplot
 ```
 
 
-# Test the installation
+## Test the installation
 ```bash
 pycmplot -h
 ```
 
-### Dependencies
+## Dependencies
 
 | Package | Purpose |
 |---------|---------|
@@ -225,6 +310,16 @@ pycmplot \
 | `-st, --sort_track` | Sort tracks by `label` or `chrom_len` | input order |
 | `-od, --output_dir` | Output directory | `.` |
 | `-of, --output_format` | Output format (`png`, `pdf`, `svg`, `jpg`) | `png` |
+| `--cache` | Enable per-track cache + user-editable hits overlay | off |
+| `--cache_dir` | Where to store cache artefacts | `./.pycmplot` |
+| `--no_resume` | Regenerate but still write fresh cache entries | resume on |
+| `--clear_cache` | Delete the cache tree and exit | off |
+| `-V, --version` | Print version and exit | — |
+
+> **QQ note (0.4.0+):** the loader no longer materialises the full sorted
+> p-value array by default. `-qq/--qq_plot` sets `compute_pvals=True`
+> automatically on the CLI; Python-API callers must pass it explicitly
+> to feed `bundle['pvals']` into a QQ plotter.
 
 Run `pycmplot -h` for the full option list.
 
@@ -234,6 +329,12 @@ Run `pycmplot -h` for the full option list.
 
 A demonstration of how to use the python API is provided in this notebook: https://github.com/esohkevin/pycmplot/blob/main/pycmplot_python_api.ipynb
 
+For an end-to-end walkthrough of every feature (caching, hits overlay,
+per-locus colours & categories, multi-panel canvas, mixed-build
+liftover, QQ opt-in), see the
+[Tutorial](https://pycmplot.readthedocs.io/en/latest/tutorial.html)
+page in the docs.
+
 
 ---
 

pycmplot-0.4.0/benchmark/build_composite_figure.py

@@ -0,0 +1,199 @@
+#!/usr/bin/env python3
+"""Build the composite scaling figure used in the manuscript.
+
+Output: ``benchmark_composite_scaling.png`` (and an accompanying PDF), written
+to the directory passed via ``--outdir`` (default: same directory as this
+script).
+
+Layout (2 rows x 3 columns):
+
+  Top row    -- single-track wall-time scaling at 500K-10M variants:
+    (A) Manhattan
+    (B) Circular Manhattan
+    (C) QQ
+
+  Bottom row -- multi-track wall-time scaling at 1M and 2M variants
+                (multi-track benchmarks were only collected at 1M / 2M):
+    (D) Multi-track Manhattan          (pycmplot vs CMplot)
+    (E) Multi-track circular Manhattan (pycmplot vs CMplot)
+    (F) Shared legend
+
+The figure reuses :data:`collect_results.SERIES_STYLE` so styling stays
+consistent with the per-plot-type PDFs produced by
+``collect_results.py --plot``.
+
+Usage
+-----
+::
+
+    cd benchmark
+    python collect_results.py --resultsdir results --out results/summary.csv
+    python build_composite_figure.py                      # writes alongside this script
+    python build_composite_figure.py --outdir ../figures  # custom output dir
+    python build_composite_figure.py --summary path/to/summary.csv
+"""
+from __future__ import annotations
+
+import argparse
+import os
+import sys
+from pathlib import Path
+
+import matplotlib
+matplotlib.use("Agg")
+import matplotlib.pyplot as plt
+import pandas as pd
+
+# Resolve ``collect_results`` relative to this script so the figure builder
+# works regardless of the current working directory.
+SCRIPT_DIR = Path(__file__).resolve().parent
+sys.path.insert(0, str(SCRIPT_DIR))
+
+from collect_results import SERIES_STYLE, SIZE_N  # noqa: E402
+
+
+PANELS = [
+    # (row, col, letter, plot_type, title)
+    (0, 0, "A", "manhattan",            "Manhattan plot"),
+    (0, 1, "B", "circular",             "Circular Manhattan plot"),
+    (0, 2, "C", "qq",                   "QQ plot"),
+    (1, 0, "D", "multitrack_manhattan", "Multi-track Manhattan plot"),
+    (1, 1, "E", "multitrack_circular",  "Multi-track circular Manhattan plot"),
+]
+
+
+def build(summary_path: Path, outdir: Path, basename: str = "benchmark_composite_scaling") -> Path:
+    """Render the composite figure and return the path to the PNG output."""
+    summary = pd.read_csv(summary_path)
+
+    #fig, axes = plt.subplots(2, 3, figsize=(13, 7.5))
+    fig, axes = plt.subplots(2, 3, figsize=(6, 3.5))
+    plt.subplots_adjust(
+        top=0.94, bottom=0.07, left=0.06, right=0.98, wspace=0.30, hspace=0.45,
+    )
+
+    all_handles: list = []
+    all_labels: list = []
+    seen: set = set()
+
+    for r, c, letter, pt, title in PANELS:
+        ax = axes[r, c]
+        sub = summary[summary["plot_type"] == pt].copy()
+        sub["n_approx"] = sub["size_label"].map(SIZE_N)
+        sub = sub.dropna(subset=["n_approx"])
+
+        tools_in_data = list(sub["tool"].unique())
+        ordered_tools = (
+            [t for t in SERIES_STYLE if t in tools_in_data]
+            + [t for t in tools_in_data if t not in SERIES_STYLE]
+        )
+
+        for tool in ordered_tools:
+            tdf = sub[sub["tool"] == tool].sort_values("n_approx")
+            if tdf.empty:
+                continue
+            s = SERIES_STYLE.get(
+                tool, dict(label=tool, color="#888780", ls="-", marker="o"),
+            )
+            line, = ax.plot(
+                tdf["n_approx"], tdf["wall_time_mean"],
+                marker=s["marker"], linestyle=s["ls"], color=s["color"],
+                label=s["label"] or tool,
+                linewidth=1, markersize=2,
+            )
+            ax.fill_between(
+                tdf["n_approx"],
+                tdf["wall_time_mean"] - tdf["wall_time_sd"].fillna(0),
+                tdf["wall_time_mean"] + tdf["wall_time_sd"].fillna(0),
+                alpha=0.12, color=s["color"], linewidth=0,
+            )
+            label = s["label"] or tool
+            if label not in seen:
+                all_handles.append(line)
+                all_labels.append(label)
+                seen.add(label)
+
+        ax.set_xscale("log")
+        ax.set_yscale("log")
+        ax.set_xlabel("Number of variants", fontsize=7) #fontsize=10
+        if c == 0:
+            ax.set_ylabel("Wall-clock time (s)", fontsize=7) #fontsize=10
+        ax.set_title(f"({letter}) {title}", fontsize=7, loc="left") #fontsize=10
+        ax.grid(True, which="both", linestyle="--", linewidth=0.3, alpha=0.5)
+        ax.tick_params(labelsize=5)
+
+    # Bottom-right cell (1, 2) holds the shared legend so each axes stays
+    # uncluttered.
+    legend_ax = axes[1, 2]
+    legend_ax.axis("off")
+    legend_ax.legend(
+        all_handles, all_labels,
+        loc="center", ncol=1, fontsize=5, frameon=False,
+        title="Tool", title_fontsize=5,
+    )
+
+    outdir.mkdir(parents=True, exist_ok=True)
+    png_path = outdir / f"{basename}.png"
+    pdf_path = outdir / f"{basename}.pdf"
+    tif_path = outdir / f"{basename}.TIF"
+    #eps_path = outdir / f"{basename}.eps"
+    fig.savefig(png_path, dpi=300, bbox_inches="tight")
+    fig.savefig(tif_path, dpi=300, bbox_inches="tight")
+    #fig.savefig(eps_path, dpi=300, bbox_inches="tight")
+    fig.savefig(pdf_path, bbox_inches="tight")
+    plt.close(fig)
+
+    print(f"Wrote {png_path}")
+    print(f"Wrote {tif_path}")
+    #print(f"Wrote {eps_path}")
+    print(f"Wrote {pdf_path}")
+    return png_path
+
+
+def main() -> None:
+    parser = argparse.ArgumentParser(
+        description="Build the composite scaling figure for the manuscript.",
+    )
+    parser.add_argument(
+        "--summary",
+        type=Path,
+        default=None,
+        help=(
+            "Path to summary.csv produced by collect_results.py.  "
+            "Default: <script_dir>/summary.csv, then <script_dir>/results/summary.csv."
+        ),
+    )
+    parser.add_argument(
+        "--outdir",
+        type=Path,
+        default=SCRIPT_DIR,
+        help="Output directory for the figure files (default: this script's directory).",
+    )
+    parser.add_argument(
+        "--basename",
+        type=str,
+        default="benchmark_composite_scaling",
+        help="Output file basename (without extension).  Default: benchmark_composite_scaling.",
+    )
+    args = parser.parse_args()
+
+    if args.summary is None:
+        candidates = [SCRIPT_DIR / "summary.csv", SCRIPT_DIR / "results" / "summary.csv"]
+        for c in candidates:
+            if c.exists():
+                args.summary = c
+                break
+        else:
+            sys.exit(
+                "Could not locate summary.csv.  Pass --summary or run "
+                "`python collect_results.py --resultsdir results --out results/summary.csv` first."
+            )
+
+    if not args.summary.exists():
+        sys.exit(f"summary file not found: {args.summary}")
+
+    build(args.summary, args.outdir, basename=args.basename)
+
+
+if __name__ == "__main__":
+    main()