multipanel
Overview
A multi-panel figure is one figure, built one of two ways depending on what
you have:
- Option 1 — redraw every panel (you have the data or plotting code): draw
each data panel with a python script into its own
subfigure so it packs
to its own labels — no empty bands, and axes need NOT align across the grid.
Follow the discipline below so legends stay inside their panels, panel letters
sit at each panel's own top-left, and text never overlaps.
- Option 2 — composite finished images (you only have rendered PNG/PDF panels):
paste them onto a
plt.subplot_mosaic canvas — fine here, since images carry no
tick labels to misalign — add panel letters, and export.
A mix is allowed: if one or two panels are image-only (no data/code), imshow
them onto their own subfigure axes and redraw the rest into the same figure. Both
modes export a vector PDF and a high-DPI PNG.
Always export the individual panels AND the composite. Every run outputs both:
one standalone figure per panel (figure1A.png, figure1B.png, …) and the combined
figure (combined_figure1.pdf + .png) — not just the composite. Because a
matplotlib subfigure cannot be saved on its own, factor every data panel's plotting
body into a draw_<letter>(ax) function (option 1); the same function then draws onto
the composite's subfigure axis AND onto a fresh standalone figure, so the panels stay
identical across both outputs with no duplicated drawing code. See "Exporting
individual panels" below.
This skill covers composition. For how to draw each individual plot type
(volcano, GSEA bar, heatmap, box/violin, PCA, Kaplan–Meier, …), use the sibling
omics-plotting skill — copy each recipe's body onto a subfigure's axis rather than
calling it as a standalone figure. Everything you need here (shared style,
composite recipe, panel-label helper) is in this document.
When to use
- The user asks for a multi-panel / composite / journal figure (panels A, B,
C…) combining two or more plots into one page of image.
- The user hands you or points out already-rendered panels (PNG/PDF) and wants them combined
into one figure (image assembly — see "Assembling user-provided panels").
- You are assembling a figure for a report, a paper submission, or a presentation
and want all panels to read as one consistent system.
Do NOT use for
- A single plot from a data table — use the sibling
omics-plotting skill.
- Interactive dashboards or web charts (this is static matplotlib output).
- 3D molecular structure rendering (that is the structure viewer, not a plot).
Key Concepts
Redraw vs composite — two composition modes
There are two fundamentally different ways to build a composite, and the user
chooses. Redraw (option 1) rebuilds every panel from data or
code in one script, giving uniform style, fonts, colors, and panel letters — best
when you hold the underlying data/DataFrame or the plotting code. Composite
(option 2) pastes already-rendered PNG/PDF panels onto a canvas and only adds
panel letters — image assembly, not plotting — best when you have only the
finished images. A mix is allowed: image-only panels are imshow-pasted while
data panels are redrawn, all into one figure.
Independent subfigures vs shared mosaic
The central layout decision. Giving each panel its own subfigure lets it run
its own constrained_layout and pack tightly to its OWN labels — panels sit flush
with no empty bands, and axes deliberately do NOT align across the grid. A single
shared subplot_mosaic gridspec instead equalizes every column's margin to its
widest y-label, leaving wide empty bands beside short-label panels. Independent
subfigures are the default here because composites usually mix heterogeneous plot
types; a shared mosaic is correct only when panels genuinely share a scale and are
meant to be read against each other.
Panel letters in the subfigure frame
Panel letters (bold A, B, C…) must sit at each panel's OWN outer top-left, left
of that panel's y-axis labels — never merged into the title and never snapped to a
shared column x-position. Placing each letter at (0, 1) in its subfigure's
coordinate frame (transform=sf.transSubfigure) guarantees it hugs its panel
regardless of neighbors' label widths.
Decision Framework
Start from what you have, then how panels relate:
What sources do you have?
├─ Data / code for every panel .................. Option 1: redraw all
├─ Only finished PNG/PDF images ................. Option 2: composite images
└─ Mix (some data, some image-only) ............. Option 1 + imshow the image-only panels
│
▼
How do the panels relate?
├─ Heterogeneous plot types (default) ........... Independent subfigures (tight pack, axes need NOT align)
└─ Same scale, read against each other .......... Shared subplot_mosaic (aligned axes)
│
▼
Layout: sketch the grid [[...]], nest subfigures for spanning panels, fill every cell
| Situation |
Approach |
Layout primitive |
Panel letters |
| Have data/code for all panels |
Redraw (option 1) |
fig.subfigures(...) per panel |
subfigure frame (0,1) |
| Only rendered images |
Composite (option 2) |
plt.subplot_mosaic + imshow |
mosaic axes top-left |
| Some data, some image-only |
Redraw + paste |
subfigures + imshow leaf |
subfigure frame (0,1) |
| Panels share a common scale |
Shared mosaic |
subplot_mosaic aligned |
axes top-left |
| Spanning panel (e.g. bottom row) |
Nested subfigures |
top[0].subfigures(1, 2) |
leaf subfigure frame |
Workflow
Ask which approach first — ask the user, then wait. Both approaches
below are usually viable and the choice is the user's, so before drawing or writing any
script, ask the user to choose between these two concrete options:
- Option 1 — Redraw every panel into one unified figure (from data/code): consistent
style, fonts, colors, and panel letters across all panels. Best when you have the
underlying data (CSV/TSV/DataFrame) or the plotting code.
- Option 2 — Composite already-rendered images: paste the finished PNG/PDF panels
onto a canvas and add panel letters — image assembly, not plotting. Best when you only
have the finished images (no data/code) or the user wants to keep the originals as-is.
Skip the question only when one option is impossible (e.g. only images and no data/code →
option 2 is forced; or a data table with no rendered images → option 1) and say why. If a
mix (some panels have data, one or two are images-only), tell the user
that the image-only panels will be pasted regardless (discipline in the intro).
Decide the layout (the grid [[...]] sketch is just to plan the tiling; you build
it with nested subfigures, not subplot_mosaic — see discipline #1). Fill every cell.
— e.g. two on top, one spanning the bottom → [["A", "B"], ["C", "C"]] →
top = fig.subfigures(2, 1); tc = top[0].subfigures(1, 2) (A,B in tc; C in top[1]).
— e.g. three on top, two on the bottom → [["A", "B", "C"], ["D", "E", "E"]].
- e.g. one big panel on the left, two stacked on the right →
[["A", "B"], ["A", "C"]] →
lr = fig.subfigures(1, 2); A = lr[0]; rr = lr[1].subfigures(2, 1).
- Gather each panel's source — a workspace-relative CSV/TSV (or DataFrame)
for data panels, or a user-supplied PNG/PDF for image panels.
- Write one python script: paste the style block, factor each data panel's
plotting body into a
draw_<letter>(ax) function (so it can render onto both a
subfigure axis and a standalone figure), build the subfigures (nest for spanning
panels), call each draw_<letter> onto its axis (data) or imshow the image,
collect the subfigures into a panels dict, and add panel letters with the helper.
Then always save both outputs to workspace-relative paths under figures/:
- the composite as
figures/combined_figure1.pdf + figures/combined_figure1.png, and
- each individual panel as
figures/figure1A.png, figures/figure1B.png, … (plus
matching .pdf) by rendering every draw_<letter> onto a fresh standalone figure.
See "Exporting individual panels" for the exact loop.
- Report the saved paths back to the user — the combined figure and every
individual panel file.
Shared style — paste at the top of the script
import matplotlib.pyplot as plt
# Publication style (colorblind-friendly, editable vector text, no top/right spines)
PUB_STYLE = {
"figure.dpi": 110, "savefig.dpi": 300, "savefig.bbox": "tight",
"font.family": "sans-serif",
"font.sans-serif": ["Arial", "Liberation Sans", "Nimbus Sans", "Helvetica", "DejaVu Sans"],
"font.size": 11, "axes.titlesize": 13, "axes.titleweight": "bold",
"figure.titlesize": 13, "figure.titleweight": "bold",
"axes.labelsize": 12, "axes.linewidth": 1.0,
"axes.spines.top": False, "axes.spines.right": False,
"xtick.labelsize": 10, "ytick.labelsize": 10,
"xtick.direction": "out", "ytick.direction": "out",
"legend.frameon": False, "legend.fontsize": 9,
"svg.fonttype": "none", "pdf.fonttype": 42, "ps.fonttype": 42,
}
plt.rcParams.update(PUB_STYLE)
# Palette — reuse the SAME colors across every panel
UP, DOWN, NS = "#d73721", "#204897", "#d9d9d9" # up / down / not-significant
PALETTE = ["#2a78d6", "#eb6834", "#1baf7a", "#eda100", "#e87ba4",
"#008300", "#4a3aa7", "#e34948", "#12a4c0", "#a66a2e"] # categorical (CVD-safe)
DIVERGING_CMAP = "RdBu_r" # z-score / log2FC — set center=0, vmin=-vmax
SEQUENTIAL_CMAP = "viridis" # magnitude / -log10 p / density
For a dense composite, lower the font: plt.rcParams.update({"font.size": 7, "axes.titlesize": 8, "axes.labelsize": 7, "legend.fontsize": 6}).
Multi-panel discipline
This is what keeps a composite clean — every rule prevents a specific failure.
- One figure, independent subfigures, constrained layout. Give each panel its
own subfigure so it packs to its OWN labels:
fig = plt.figure(layout="constrained", figsize=(width_mm/25.4, height_mm/25.4)), then sfs = fig.subfigures(nrows, ncols, width_ratios=..., height_ratios=...) and ax = sfs[r, c].subplots() per panel. Each
subfigure runs its own constrained_layout, so a panel with long y-tick labels no
longer shoves its column-neighbors' plots sideways — axes deliberately do NOT align
across the grid; panels sit flush with no empty bands (a single shared
subplot_mosaic gridspec, by contrast, equalizes each column's margin to its widest
y-label and leaves a wide gap beside the short-label panels). Reserve a hair of margin
so panel letters never clip: fig.get_layout_engine().set(rect=(0.012, 0, 0.988, 0.985)). Never add tight_layout() or manual subplots_adjust. Size in mm (single
column = 88 mm, double = 180 mm).
- Spanning panels: nest subfigures — e.g. two panels on top, one spanning the
bottom →
top = fig.subfigures(2, 1); tc = top[0].subfigures(1, 2) (A, B in tc[0],
tc[1]; C in top[1]). One .subplots() per leaf subfigure.
- Match each panel to its plot's shape via the subfigures'
width_ratios/
height_ratios (pin with ax.set_box_aspect(...) if it still deforms): scatter
panels (volcano/PCA) near-square; for bar / box / histogram, protect the value
axis in both orientations — horizontal (barh, horizontal box) kept wide,
vertical (bar, box, hist) kept tall. Never let a neighbor squeeze that axis flat.
- When panels genuinely share a scale (same y-range, meant to be read against each
other), a shared
subplot_mosaic with aligned axes is the right choice instead — but
this skill usually combines heterogeneous plot types, so independent subfigures are
the default.
- Fill every cell. No empty grid slots. If a panel would be blank, span a
neighbor across it:
[["A", "B"], ["C", "C"]].
- Legends & colorbars belong to their own panel — a legend in that panel's
free corner (
ax.legend(loc="lower right", frameon=False)) or a colorbar on
that one axis (fig.colorbar(im, ax=ax, fraction=0.025, pad=0.02)). Never float
a figure-level legend in empty space or stack two in a margin; for a dot plot,
keep only the colorbar and drop the size legend (count range → panel title).
- Too wide? Thin a colorbar with
aspect=40 + shrink=0.6 (and small
fraction); tighten a legend with handlelength=1.0, handletextpad=0.2,
borderpad=0.2, or fold long legends into ncol.
- Panel letters at each panel's OWN outer top-left. Bold capitals
A, B, C…
(lowercase fine; never numeric), placed to the LEFT of that panel's y-axis tick and
axis labels — not merged into the title. Place each letter in its own subfigure's
coordinate frame — sf.text(0.0, 1.0, letter, transform=sf.transSubfigure, ...) (the
helper below). The subfigure's top-left corner is always left of that panel's y-labels
and hugs that panel, so the letter never overlaps a wide label and never floats over an
empty band. Do not snap letters to a shared column-x — with independent packing that
would drag a short-label panel's letter far from its plot (the empty-gap failure the
user sees). The reserved margin from discipline #1 keeps edge letters in-canvas.
- Text must stay readable — the #1 way composites go wrong.
- Point labels (volcano/scatter): cap to ≤5 strongest hits in a small
panel, italic ~6 pt, and repel with
adjustText; if it is not installed,
skip labels rather than dumping overlapping text.
- Long category names (pathways/gene sets): put them on the y-axis
(horizontal, one per row), never crammed/rotated on a narrow x-axis. If they
must go on x, rotate (45° to save vertical space, or 90° when very long,
ha="right" so the tick end aligns under its bar), wrap to ≤26 chars, and
give enough width per column — crammed x-labels otherwise collide into
unreadable text.
Shorten over-long names (common with MSigDB/GO/Reactome): strip the DB
prefix (HALLMARK_, GO_, REACTOME_, KEGG_), swap _→space and title-case,
and replace verbose terms with standard abbreviations (e.g.
HALLMARK_INTERFERON_GAMMA_RESPONSE → IFN-γ response); truncate with an
ellipsis only if still too long. Keep the full name in the underlying
data/tooltip, not on the axis tick.
- Size-encoded markers (dot plot): floor the size range (
s in ~[25,150])
so small dots stay visible.
- In-cell heatmap numbers: annotate only when the values are needed (small
font ~4–5 pt, no decimals); otherwise omit them and let the colorbar carry the
values.
- Reuse one palette and axis convention across panels so the composite reads
as a single system.
- Export vector PDF + PNG and report the relative path.
Panel-label helper
Place each letter at the top-left corner of its own subfigure. Because every panel
lives in its own tightly-packed subfigure, that corner is always left of the panel's
y-labels and hugs the panel — so letters never float over an empty band (the shared-column
failure) and never overlap a wide y-label, no matter how the panels' label widths differ:
def add_panel_labels(panels, size=11):
"""Bold letter at each panel's OWN outer top-left, in its subfigure frame.
panels : dict {letter: subfigure} — the subfigure that holds each panel's axes,
collected as you build them (for a spanning panel, its leaf subfigure). Placing
the letter at (0, 1) in the subfigure's coordinates puts it at that cell's top-left
corner: always LEFT of the panel's y-labels and hugging the panel, with no
dependence on any neighbor's label width. Reserve a hair of figure margin first
(`fig.get_layout_engine().set(rect=(0.012, 0, 0.988, 0.985))`, discipline #1) so the
letters of edge panels are not clipped at the canvas edge.
"""
for letter, sf in panels.items():
sf.text(0.0, 1.0, letter, transform=sf.transSubfigure,
fontsize=size, fontweight="bold", va="top", ha="left")
Usage: collect the subfigures as you create them, e.g. panels = {"A": sfs[0, 0], "B": sfs[0, 1], "C": top[1]}, then call add_panel_labels(panels).
Exporting individual panels
Every run produces both the individual panels (figure1A.png, figure1B.png, …)
and the composite (combined_figure1.pdf + .png) — this is the default output,
not an extra. A matplotlib subfigure cannot be saved on its own, so put each panel's
plotting body in a draw_<letter>(ax) function and call it twice: once onto the
composite's subfigure axis, and once onto a fresh standalone figure. One source of
truth per panel — the panels stay identical across both outputs.
import os
os.makedirs("plots", exist_ok=True)
# 1) Factor each DATA panel's body into a function of a single Axes.
# (Copy the omics-plotting recipe body here, drawing onto `ax` instead of a new figure.)
def draw_A(ax):
ax.scatter(df["log2FC"], -np.log10(df["padj"]), s=8, c=NS) # volcano, etc.
ax.set_xlabel("log2 fold change"); ax.set_ylabel("-log10 FDR")
def draw_B(ax):
... # PCA / box / heatmap body onto ax
def draw_C(ax):
...
DATA_PANELS = {"A": draw_A, "B": draw_B, "C": draw_C}
# Per-panel standalone figure size (mm) — match each plot's shape (discipline #1).
PANEL_SIZE_MM = {"A": (88, 75), "B": (88, 75), "C": (180, 70)}
# Image-only panels stay separate: keep the PNG/PDF the user supplied as their
# standalone file, and only imshow them onto the composite axis (see intro).
# 2) Composite — draw each function onto its subfigure axis, add letters, save.
panels = {"A": sfs[0, 0], "B": sfs[0, 1], "C": top[1]}
for letter, sf in panels.items():
DATA_PANELS[letter](sf.subplots())
add_panel_labels(panels)
fig.savefig("figures/combined_figure1.pdf")
fig.savefig("figures/combined_figure1.png", dpi=300)
# 3) Individual panels — same functions onto fresh standalone figures (no letter).
for letter, draw in DATA_PANELS.items():
w_mm, h_mm = PANEL_SIZE_MM[letter]
fp = plt.figure(layout="constrained", figsize=(w_mm / 25.4, h_mm / 25.4))
draw(fp.subplots())
fp.savefig(f"figures/figure1{letter}.pdf")
fp.savefig(f"figures/figure1{letter}.png", dpi=300)
plt.close(fp)
Output files (Figure 1 with panels A, B, C):
figures/combined_figure1.pdf, figures/combined_figure1.png,
figures/figure1A.{pdf,png}, figures/figure1B.{pdf,png}, figures/figure1C.{pdf,png}.
Notes:
- No panel letter on standalones — the
A/B/C label belongs to the composite
frame only; a lone figure1A.png needs no letter baked in.
- Size each standalone to its plot's shape (discipline #1) via
PANEL_SIZE_MM:
scatter/PCA near-square, barh/horizontal-box wide, vertical bar/box/hist tall —
don't reuse one size for all.
- Legends/colorbars still belong to their own axis (discipline #3) — since the
body lives in
draw_<letter>, attach them inside that function so they appear in
both the composite and the standalone.
- Image-only panels are already standalone files (the user's PNG/PDF); don't
re-export them — just reference the originals.
Best Practices
- One figure, one style. Never stitch separate PNGs or call standalone plot
functions for a composite; copy their bodies onto each subfigure's axis.
- Workspace-relative paths only. Save under
figures/ (create it if needed);
never absolute paths like /tmp or /home/....
- Only plot data that exists. Never invent columns, groups, or values.
- Label every axis, keep every legend inside its panel, fill every cell.
- Always export both a vector
.pdf and a .png (dpi≥300) under figures/.
Common Pitfalls
- Building the whole figure as one shared
subplot_mosaic gridspec. It
equalizes each column's margin to its widest y-label, so a long-label panel
shoves neighbors sideways, leaves empty bands, and strands letters snapped to the
shared column edge. How to avoid: give each panel its own subfigure so it
packs to its own labels (discipline #1); reserve a shared mosaic only for panels
that genuinely share a scale.
- Panel letters merged into titles, snapped to a shared column-x, or clipped at
the edge. They then sit right of the y-labels, float far from their plot, or
vanish off-canvas. How to avoid: place each letter at
(0, 1) in its own
subfigure frame (transform=sf.transSubfigure) with the add_panel_labels
helper, and reserve a hair of margin (rect=(0.012, 0, 0.988, 0.985),
discipline #1) so edge letters stay in-canvas.
- Floating or bulky legends and colorbars. Per-plot figure-level legends
collide in the margins, or a colorbar eats half the panel. How to avoid:
attach each legend/colorbar to its own panel's axis, drop a composite dot plot's
size legend, and thin a wide colorbar (
aspect=40, shrink=0.6, small
fraction) (discipline #3).
- Value axis flattened — scatter dots merge or bars/boxes squash. A neighbor
steals the space the plot's value direction needs. How to avoid: widen or
heighten that cell via
width_ratios / height_ratios (or pin with
ax.set_box_aspect) instead of shrinking the plot; only then bump marker size.
- Cramming long category names onto a narrow x-axis. Pathway/gene-set names
collide into unreadable text. How to avoid: put long names on the horizontal
y-axis, strip DB prefixes (
HALLMARK_, GO_) and abbreviate, or rotate 90° and
wrap to ≤26 chars with enough panel width.
- Fixing cramped panels with manual spacing. Adding
tight_layout() or
subplots_adjust fights constrained_layout and makes it worse. How to
avoid: instead increase figsize (in mm), adjust the ratios, or lower the
font, and let constrained layout re-space.
- Inconsistent or unreadable text. Over-labeled points overlap, heatmap cell
numbers are too dense, and font sizes drift between panels. How to avoid: cap
point labels to ≤5 (repel with
adjustText, else skip); annotate heatmap cells
only at ~4–5 pt with no decimals or drop them for the colorbar; keep the same
font sizes across all panels, including any the user supplies (discipline #5).
Further Reading
1---2name: multipanel3description: Assemble multiple plots into ONE publication-ready multi-panel journal figure (e.g. Figure 1 with panels A, B, C). Use whenever the user asks to combine, compose, or lay out several plots as a single composite figure — newly plotted from data or from already-rendered panels the user supplies (PNG/PDF). Ask the user to pick one of two approaches: (1) redraw every panel into one unified figure using independent, tightly packed `subfigures` (each sized to its own labels, so axes need NOT align), consistent style, correctly placed panel letters, and per-panel legends/colorbars; (2) composite already-rendered PNG/PDF panels onto a mosaic canvas and add panel letters (image compositing, not plotting). Both export vector PDF + high-DPI PNG. For a SINGLE plot from a data table, use the sibling `omics-plotting` skill instead.4license: Proprietary (HITS Inc.)5---67# multipanel89## Overview1011A multi-panel figure is **one** figure, built one of two ways depending on what12you have:1314- **Option 1 — redraw every panel** (you have the data or plotting code): draw15 each data panel with a python script into its **own `subfigure`** so it packs16 to its own labels — no empty bands, and axes need NOT align across the grid.17 Follow the discipline below so legends stay inside their panels, panel letters18 sit at each panel's own top-left, and text never overlaps.19- **Option 2 — composite finished images** (you only have rendered PNG/PDF panels):20 paste them onto a `plt.subplot_mosaic` canvas — fine here, since images carry no21 tick labels to misalign — add panel letters, and export.2223A mix is allowed: if one or two panels are image-only (no data/code), `imshow`24them onto their own subfigure axes and redraw the rest into the same figure. Both25modes export a vector PDF and a high-DPI PNG.2627**Always export the individual panels AND the composite.** Every run outputs both:28one standalone figure per panel (`figure1A.png`, `figure1B.png`, …) and the combined29figure (`combined_figure1.pdf` + `.png`) — not just the composite. Because a30matplotlib `subfigure` cannot be saved on its own, factor every data panel's plotting31body into a `draw_<letter>(ax)` function (option 1); the same function then draws onto32the composite's subfigure axis AND onto a fresh standalone figure, so the panels stay33identical across both outputs with no duplicated drawing code. See "Exporting34individual panels" below.3536This skill covers **composition**. For how to draw each individual plot type37(volcano, GSEA bar, heatmap, box/violin, PCA, Kaplan–Meier, …), use the sibling38`omics-plotting` skill — copy each recipe's **body** onto a subfigure's axis rather than39calling it as a standalone figure. Everything you need here (shared style,40composite recipe, panel-label helper) is in this document.4142## When to use4344- The user asks for a **multi-panel / composite / journal figure** (panels A, B,45 C…) combining two or more plots into one page of image.46- The user hands you or points out **already-rendered panels (PNG/PDF)** and wants them combined47 into one figure (image assembly — see "Assembling user-provided panels").48- You are assembling a figure for a report, a paper submission, or a presentation49 and want all panels to read as one consistent system.5051## Do NOT use for5253- A **single** plot from a data table — use the sibling `omics-plotting` skill.54- Interactive dashboards or web charts (this is static matplotlib output).55- 3D molecular structure rendering (that is the structure viewer, not a plot).5657## Key Concepts5859### Redraw vs composite — two composition modes6061There are two fundamentally different ways to build a composite, and the user62chooses. **Redraw (option 1)** rebuilds every panel from data or63code in one script, giving uniform style, fonts, colors, and panel letters — best64when you hold the underlying data/DataFrame or the plotting code. **Composite65(option 2)** pastes already-rendered PNG/PDF panels onto a canvas and only adds66panel letters — image assembly, not plotting — best when you have only the67finished images. A mix is allowed: image-only panels are `imshow`-pasted while68data panels are redrawn, all into one figure.6970### Independent subfigures vs shared mosaic7172The central layout decision. Giving **each panel its own `subfigure`** lets it run73its own `constrained_layout` and pack tightly to its OWN labels — panels sit flush74with no empty bands, and axes deliberately do NOT align across the grid. A single75shared `subplot_mosaic` gridspec instead equalizes every column's margin to its76widest y-label, leaving wide empty bands beside short-label panels. Independent77subfigures are the default here because composites usually mix heterogeneous plot78types; a shared mosaic is correct only when panels genuinely share a scale and are79meant to be read against each other.8081### Panel letters in the subfigure frame8283Panel letters (bold `A, B, C…`) must sit at each panel's OWN outer top-left, left84of that panel's y-axis labels — never merged into the title and never snapped to a85shared column x-position. Placing each letter at `(0, 1)` in its subfigure's86coordinate frame (`transform=sf.transSubfigure`) guarantees it hugs its panel87regardless of neighbors' label widths.8889## Decision Framework9091Start from what you have, then how panels relate:9293```94What sources do you have?95├─ Data / code for every panel .................. Option 1: redraw all96├─ Only finished PNG/PDF images ................. Option 2: composite images97└─ Mix (some data, some image-only) ............. Option 1 + imshow the image-only panels98 │99 ▼100How do the panels relate?101├─ Heterogeneous plot types (default) ........... Independent subfigures (tight pack, axes need NOT align)102└─ Same scale, read against each other .......... Shared subplot_mosaic (aligned axes)103 │104 ▼105Layout: sketch the grid [[...]], nest subfigures for spanning panels, fill every cell106```107108| Situation | Approach | Layout primitive | Panel letters |109|---|---|---|---|110| Have data/code for all panels | Redraw (option 1) | `fig.subfigures(...)` per panel | subfigure frame `(0,1)` |111| Only rendered images | Composite (option 2) | `plt.subplot_mosaic` + `imshow` | mosaic axes top-left |112| Some data, some image-only | Redraw + paste | subfigures + `imshow` leaf | subfigure frame `(0,1)` |113| Panels share a common scale | Shared mosaic | `subplot_mosaic` aligned | axes top-left |114| Spanning panel (e.g. bottom row) | Nested subfigures | `top[0].subfigures(1, 2)` | leaf subfigure frame |115116## Workflow1171181. **Ask which approach first — ask the user, then wait.** Both approaches119 below are usually viable and the choice is the user's, so **before drawing or writing any120 script, ask the user to choose between these two concrete options**:121 - **Option 1 — Redraw every panel into one unified figure** (from data/code): consistent122 style, fonts, colors, and panel letters across all panels. Best when you have the123 underlying data (CSV/TSV/DataFrame) or the plotting code.124 - **Option 2 — Composite already-rendered images**: paste the finished PNG/PDF panels125 onto a canvas and add panel letters — image assembly, not plotting. Best when you only126 have the finished images (no data/code) or the user wants to keep the originals as-is.127128 Skip the question only when one option is impossible (e.g. only images and no data/code →129 option 2 is forced; or a data table with no rendered images → option 1) and say why. If a130 mix (some panels have data, one or two are images-only), tell the user131 that the image-only panels will be pasted regardless (discipline in the intro).1322. **Decide the layout** (the grid `[[...]]` sketch is just to plan the tiling; you build133 it with nested `subfigures`, not `subplot_mosaic` — see discipline #1). Fill every cell.134 — e.g. two on top, one spanning the bottom → `[["A", "B"], ["C", "C"]]` →135 `top = fig.subfigures(2, 1); tc = top[0].subfigures(1, 2)` (A,B in `tc`; C in `top[1]`).136 — e.g. three on top, two on the bottom → `[["A", "B", "C"], ["D", "E", "E"]]`.137 - e.g. one big panel on the left, two stacked on the right → `[["A", "B"], ["A", "C"]]` →138 `lr = fig.subfigures(1, 2); A = lr[0]; rr = lr[1].subfigures(2, 1)`.1393. **Gather each panel's source** — a workspace-relative CSV/TSV (or DataFrame)140 for data panels, or a user-supplied PNG/PDF for image panels.1414. **Write one python script**: paste the style block, **factor each data panel's142 plotting body into a `draw_<letter>(ax)` function** (so it can render onto both a143 subfigure axis and a standalone figure), build the subfigures (nest for spanning144 panels), call each `draw_<letter>` onto its axis (data) or `imshow` the image,145 collect the subfigures into a `panels` dict, and add panel letters with the helper.146 Then **always save both outputs** to **workspace-relative** paths under `figures/`:147 - the **composite** as `figures/combined_figure1.pdf` + `figures/combined_figure1.png`, and148 - **each individual panel** as `figures/figure1A.png`, `figures/figure1B.png`, … (plus149 matching `.pdf`) by rendering every `draw_<letter>` onto a fresh standalone figure.150 See "Exporting individual panels" for the exact loop.1515. **Report the saved paths** back to the user — the combined figure and every152 individual panel file.153154## Shared style — paste at the top of the script155156```python157import matplotlib.pyplot as plt158159# Publication style (colorblind-friendly, editable vector text, no top/right spines)160PUB_STYLE = {161 "figure.dpi": 110, "savefig.dpi": 300, "savefig.bbox": "tight",162 "font.family": "sans-serif",163 "font.sans-serif": ["Arial", "Liberation Sans", "Nimbus Sans", "Helvetica", "DejaVu Sans"],164 "font.size": 11, "axes.titlesize": 13, "axes.titleweight": "bold",165 "figure.titlesize": 13, "figure.titleweight": "bold",166 "axes.labelsize": 12, "axes.linewidth": 1.0,167 "axes.spines.top": False, "axes.spines.right": False,168 "xtick.labelsize": 10, "ytick.labelsize": 10,169 "xtick.direction": "out", "ytick.direction": "out",170 "legend.frameon": False, "legend.fontsize": 9,171 "svg.fonttype": "none", "pdf.fonttype": 42, "ps.fonttype": 42,172}173plt.rcParams.update(PUB_STYLE)174175# Palette — reuse the SAME colors across every panel176UP, DOWN, NS = "#d73721", "#204897", "#d9d9d9" # up / down / not-significant177PALETTE = ["#2a78d6", "#eb6834", "#1baf7a", "#eda100", "#e87ba4",178 "#008300", "#4a3aa7", "#e34948", "#12a4c0", "#a66a2e"] # categorical (CVD-safe)179DIVERGING_CMAP = "RdBu_r" # z-score / log2FC — set center=0, vmin=-vmax180SEQUENTIAL_CMAP = "viridis" # magnitude / -log10 p / density181```182183For a dense composite, lower the font: `plt.rcParams.update({"font.size": 7,184"axes.titlesize": 8, "axes.labelsize": 7, "legend.fontsize": 6})`.185186## Multi-panel discipline187188This is what keeps a composite clean — every rule prevents a specific failure.1891901. **One figure, independent subfigures, constrained layout.** Give **each panel its191 own subfigure** so it packs to its OWN labels: `fig = plt.figure(layout="constrained",192 figsize=(width_mm/25.4, height_mm/25.4))`, then `sfs = fig.subfigures(nrows, ncols,193 width_ratios=..., height_ratios=...)` and `ax = sfs[r, c].subplots()` per panel. Each194 subfigure runs its own `constrained_layout`, so a panel with long y-tick labels no195 longer shoves its column-neighbors' plots sideways — **axes deliberately do NOT align196 across the grid; panels sit flush with no empty bands** (a single shared197 `subplot_mosaic` gridspec, by contrast, equalizes each column's margin to its widest198 y-label and leaves a wide gap beside the short-label panels). Reserve a hair of margin199 so panel letters never clip: `fig.get_layout_engine().set(rect=(0.012, 0, 0.988,200 0.985))`. Never add `tight_layout()` or manual `subplots_adjust`. Size in mm (single201 column = 88 mm, double = 180 mm).202 - **Spanning panels**: nest subfigures — e.g. two panels on top, one spanning the203 bottom → `top = fig.subfigures(2, 1); tc = top[0].subfigures(1, 2)` (A, B in `tc[0]`,204 `tc[1]`; C in `top[1]`). One `.subplots()` per leaf subfigure.205 - **Match each panel to its plot's shape** via the subfigures' `width_ratios`/206 `height_ratios` (pin with `ax.set_box_aspect(...)` if it still deforms): scatter207 panels (volcano/PCA) near-square; for **bar / box / histogram**, protect the value208 axis in **both orientations** — horizontal (`barh`, horizontal box) kept wide,209 vertical (bar, box, hist) kept tall. Never let a neighbor squeeze that axis flat.210 - **When panels genuinely share a scale** (same y-range, meant to be read against each211 other), a shared `subplot_mosaic` with aligned axes is the right choice instead — but212 this skill usually combines heterogeneous plot types, so independent subfigures are213 the default.2142. **Fill every cell.** No empty grid slots. If a panel would be blank, span a215 neighbor across it: `[["A", "B"], ["C", "C"]]`.2163. **Legends & colorbars belong to their own panel** — a legend in that panel's217 free corner (`ax.legend(loc="lower right", frameon=False)`) or a colorbar on218 that one axis (`fig.colorbar(im, ax=ax, fraction=0.025, pad=0.02)`). Never float219 a figure-level legend in empty space or stack two in a margin; for a dot plot,220 keep only the colorbar and drop the size legend (count range → panel title).221 - **Too wide?** Thin a colorbar with `aspect=40` + `shrink=0.6` (and small222 `fraction`); tighten a legend with `handlelength=1.0`, `handletextpad=0.2`,223 `borderpad=0.2`, or fold long legends into `ncol`.2244. **Panel letters at each panel's OWN outer top-left.** Bold capitals `A, B, C…`225 (lowercase fine; never numeric), placed **to the LEFT of that panel's y-axis tick and226 axis labels** — not merged into the title. Place each letter in its **own subfigure's227 coordinate frame** — `sf.text(0.0, 1.0, letter, transform=sf.transSubfigure, ...)` (the228 helper below). The subfigure's top-left corner is always left of that panel's y-labels229 and hugs that panel, so the letter never overlaps a wide label and never floats over an230 empty band. Do **not** snap letters to a shared column-x — with independent packing that231 would drag a short-label panel's letter far from its plot (the empty-gap failure the232 user sees). The reserved margin from discipline #1 keeps edge letters in-canvas.2335. **Text must stay readable — the #1 way composites go wrong.**234 - **Point labels (volcano/scatter):** cap to **≤5** strongest hits in a small235 panel, italic ~6 pt, and repel with `adjustText`; if it is not installed,236 skip labels rather than dumping overlapping text.237 - **Long category names (pathways/gene sets):** put them on the **y-axis**238 (horizontal, one per row), never crammed/rotated on a narrow x-axis. If they239 must go on x, rotate (45° to save vertical space, or 90° when very long,240 `ha="right"` so the tick end aligns under its bar), wrap to ≤26 chars, and241 give enough width per column — crammed x-labels otherwise collide into242 unreadable text.243 **Shorten over-long names** (common with MSigDB/GO/Reactome): strip the DB244 prefix (`HALLMARK_`, `GO_`, `REACTOME_`, `KEGG_`), swap `_`→space and title-case,245 and replace verbose terms with **standard abbreviations** (e.g.246 `HALLMARK_INTERFERON_GAMMA_RESPONSE` → `IFN-γ response`); truncate with an247 ellipsis only if still too long. Keep the full name in the underlying248 data/tooltip, not on the axis tick.249 - **Size-encoded markers (dot plot):** floor the size range (`s` in ~[25,150])250 so small dots stay visible.251 - **In-cell heatmap numbers:** annotate only when the values are needed (small252 font ~4–5 pt, no decimals); otherwise omit them and let the colorbar carry the253 values.2546. **Reuse one palette and axis convention across panels** so the composite reads255 as a single system.2567. **Export vector PDF + PNG** and report the relative path.257258## Panel-label helper259260Place each letter at the top-left **corner of its own subfigure**. Because every panel261lives in its own tightly-packed subfigure, that corner is always left of the panel's262y-labels and hugs the panel — so letters never float over an empty band (the shared-column263failure) and never overlap a wide y-label, no matter how the panels' label widths differ:264265```python266def add_panel_labels(panels, size=11):267 """Bold letter at each panel's OWN outer top-left, in its subfigure frame.268269 panels : dict {letter: subfigure} — the subfigure that holds each panel's axes,270 collected as you build them (for a spanning panel, its leaf subfigure). Placing271 the letter at (0, 1) in the subfigure's coordinates puts it at that cell's top-left272 corner: always LEFT of the panel's y-labels and hugging the panel, with no273 dependence on any neighbor's label width. Reserve a hair of figure margin first274 (`fig.get_layout_engine().set(rect=(0.012, 0, 0.988, 0.985))`, discipline #1) so the275 letters of edge panels are not clipped at the canvas edge.276 """277 for letter, sf in panels.items():278 sf.text(0.0, 1.0, letter, transform=sf.transSubfigure,279 fontsize=size, fontweight="bold", va="top", ha="left")280```281282Usage: collect the subfigures as you create them, e.g. `panels = {"A": sfs[0, 0],283"B": sfs[0, 1], "C": top[1]}`, then call `add_panel_labels(panels)`.284285## Exporting individual panels286287Every run produces **both** the individual panels (`figure1A.png`, `figure1B.png`, …)288**and** the composite (`combined_figure1.pdf` + `.png`) — this is the default output,289not an extra. A matplotlib `subfigure` cannot be saved on its own, so put each panel's290plotting body in a `draw_<letter>(ax)` function and call it twice: once onto the291composite's subfigure axis, and once onto a fresh standalone figure. One source of292truth per panel — the panels stay identical across both outputs.293294```python295import os296297os.makedirs("plots", exist_ok=True)298299# 1) Factor each DATA panel's body into a function of a single Axes.300# (Copy the omics-plotting recipe body here, drawing onto `ax` instead of a new figure.)301def draw_A(ax):302 ax.scatter(df["log2FC"], -np.log10(df["padj"]), s=8, c=NS) # volcano, etc.303 ax.set_xlabel("log2 fold change"); ax.set_ylabel("-log10 FDR")304305def draw_B(ax):306 ... # PCA / box / heatmap body onto ax307308def draw_C(ax):309 ...310311DATA_PANELS = {"A": draw_A, "B": draw_B, "C": draw_C}312# Per-panel standalone figure size (mm) — match each plot's shape (discipline #1).313PANEL_SIZE_MM = {"A": (88, 75), "B": (88, 75), "C": (180, 70)}314# Image-only panels stay separate: keep the PNG/PDF the user supplied as their315# standalone file, and only imshow them onto the composite axis (see intro).316317# 2) Composite — draw each function onto its subfigure axis, add letters, save.318panels = {"A": sfs[0, 0], "B": sfs[0, 1], "C": top[1]}319for letter, sf in panels.items():320 DATA_PANELS[letter](sf.subplots())321add_panel_labels(panels)322fig.savefig("figures/combined_figure1.pdf")323fig.savefig("figures/combined_figure1.png", dpi=300)324325# 3) Individual panels — same functions onto fresh standalone figures (no letter).326for letter, draw in DATA_PANELS.items():327 w_mm, h_mm = PANEL_SIZE_MM[letter]328 fp = plt.figure(layout="constrained", figsize=(w_mm / 25.4, h_mm / 25.4))329 draw(fp.subplots())330 fp.savefig(f"figures/figure1{letter}.pdf")331 fp.savefig(f"figures/figure1{letter}.png", dpi=300)332 plt.close(fp)333```334335Output files (Figure 1 with panels A, B, C):336`figures/combined_figure1.pdf`, `figures/combined_figure1.png`,337`figures/figure1A.{pdf,png}`, `figures/figure1B.{pdf,png}`, `figures/figure1C.{pdf,png}`.338339Notes:340- **No panel letter on standalones** — the `A/B/C` label belongs to the composite341 frame only; a lone `figure1A.png` needs no letter baked in.342- **Size each standalone to its plot's shape** (discipline #1) via `PANEL_SIZE_MM`:343 scatter/PCA near-square, `barh`/horizontal-box wide, vertical bar/box/hist tall —344 don't reuse one size for all.345- **Legends/colorbars still belong to their own axis** (discipline #3) — since the346 body lives in `draw_<letter>`, attach them inside that function so they appear in347 both the composite and the standalone.348- **Image-only panels** are already standalone files (the user's PNG/PDF); don't349 re-export them — just reference the originals.350351## Best Practices352353- **One figure, one style.** Never stitch separate PNGs or call standalone plot354 functions for a composite; copy their bodies onto each subfigure's axis.355- **Workspace-relative paths only.** Save under `figures/` (create it if needed);356 never absolute paths like `/tmp` or `/home/...`.357- **Only plot data that exists.** Never invent columns, groups, or values.358- **Label every axis, keep every legend inside its panel, fill every cell.**359- **Always export both** a vector `.pdf` and a `.png` (dpi≥300) under `figures/`.360361## Common Pitfalls362363- **Building the whole figure as one shared `subplot_mosaic` gridspec.** It364 equalizes each column's margin to its widest y-label, so a long-label panel365 shoves neighbors sideways, leaves empty bands, and strands letters snapped to the366 shared column edge. **How to avoid:** give each panel its own `subfigure` so it367 packs to its own labels (discipline #1); reserve a shared mosaic only for panels368 that genuinely share a scale.369- **Panel letters merged into titles, snapped to a shared column-x, or clipped at370 the edge.** They then sit right of the y-labels, float far from their plot, or371 vanish off-canvas. **How to avoid:** place each letter at `(0, 1)` in its own372 subfigure frame (`transform=sf.transSubfigure`) with the `add_panel_labels`373 helper, and reserve a hair of margin (`rect=(0.012, 0, 0.988, 0.985)`,374 discipline #1) so edge letters stay in-canvas.375- **Floating or bulky legends and colorbars.** Per-plot figure-level legends376 collide in the margins, or a colorbar eats half the panel. **How to avoid:**377 attach each legend/colorbar to its own panel's axis, drop a composite dot plot's378 size legend, and thin a wide colorbar (`aspect=40`, `shrink=0.6`, small379 `fraction`) (discipline #3).380- **Value axis flattened — scatter dots merge or bars/boxes squash.** A neighbor381 steals the space the plot's value direction needs. **How to avoid:** widen or382 heighten that cell via `width_ratios` / `height_ratios` (or pin with383 `ax.set_box_aspect`) instead of shrinking the plot; only then bump marker size.384- **Cramming long category names onto a narrow x-axis.** Pathway/gene-set names385 collide into unreadable text. **How to avoid:** put long names on the horizontal386 y-axis, strip DB prefixes (`HALLMARK_`, `GO_`) and abbreviate, or rotate 90° and387 wrap to ≤26 chars with enough panel width.388- **Fixing cramped panels with manual spacing.** Adding `tight_layout()` or389 `subplots_adjust` fights `constrained_layout` and makes it worse. **How to390 avoid:** instead increase `figsize` (in mm), adjust the ratios, or lower the391 font, and let constrained layout re-space.392- **Inconsistent or unreadable text.** Over-labeled points overlap, heatmap cell393 numbers are too dense, and font sizes drift between panels. **How to avoid:** cap394 point labels to ≤5 (repel with `adjustText`, else skip); annotate heatmap cells395 only at ~4–5 pt with no decimals or drop them for the colorbar; keep the same396 font sizes across all panels, including any the user supplies (discipline #5).397398## Further Reading399400- Matplotlib subfigures / `Figure.subfigures` — https://matplotlib.org/stable/gallery/subplots_axes_and_figures/subfigures.html401- Matplotlib constrained layout guide — https://matplotlib.org/stable/users/explain/axes/constrained_layout_guide.html402- Matplotlib `subplot_mosaic` tutorial — https://matplotlib.org/stable/users/explain/axes/mosaic.html403- adjustText (label de-overlap) — https://github.com/Phlya/adjustText