Scientific Publication

Scientific Publication Figure Refinement

Expert guidance for systematically improving scientific figures through iterative refinement based on user feedback and publication requirements.

Supporting files in this directory:

- publication-standards.md - DPI, file formats, size specs, color accessibility

- multi-study-results.md - Writing integrated results from multi-study analyses and practical recommendations from complex trade-offs

- methodological-transparency.md - Dual approach pattern for figures vs statistics, outlier handling

- overleaf-packages.md - Creating production-ready Overleaf packages with templates and checklists

When to Use This Skill

• Improving figures based on reviewer or collaborator feedback
• Optimizing figure clarity and readability
• Ensuring all figure elements fit within bounds
• Deciding between layout alternatives (horizontal vs vertical panels)
• Preparing figures for high-impact publications

Iterative Figure Refinement Workflow

Standard Refinement Sequence

When improving a publication figure, follow this systematic approach:

1. Identify the Core Issue

Examples:
- "Violin plots look distorted on log scale"
- "P-values are cut off at the top"
- "Too much visual clutter, hard to see the data"
- "Text overlaps with data points"

2. Fix the Visualization Type/Method

# Example: Replace inappropriate plot type
# Before: Violin plot on log scale (distorted)
ax.violinplot(data)
ax.set_yscale('log')

# After: Boxplot on log scale (accurate)
ax.boxplot(data)
ax.set_yscale('log')

3. Improve Visual Clarity Systematically adjust element sizes:

# Point sizes: Reduce for dense data
# Start: s=60 (exploratory)
# End: s=25 (publication)
ax.scatter(..., s=25, alpha=0.5)

# Line widths: Thinner reduces clutter
# Start: linewidth=2.5
# End: linewidth=1.5
ax.plot(..., linewidth=1.5)

# Text sizes: Prevent overlap
# Start: fontsize=10-12
# End: fontsize=8-9
ax.text(..., fontsize=8)

# Error bar caps: Keep readable
ax.errorbar(..., capsize=5)

4. Test Layout Alternatives

# Option A: Side-by-side panels
fig, axes = plt.subplots(1, 2, figsize=(16, 7))
# Pros: Direct left-right comparison
# Cons: Smaller individual panels

# Option B: Stacked vertically
fig, axes = plt.subplots(2, 1, figsize=(10, 14))
# Pros: Larger individual panels, easier to read details
# Cons: Harder to compare across panels

# Decision: Let user feedback guide choice
# Generate both, ask which is clearer

5. Optimize Element Positioning Ensure all annotations fit within plot bounds:

# Calculate safe positioning
y_max = max([d.max() for d in data_list])
y_min = min([d.min() for d in data_list])

# Position annotations WITHIN bounds
y_pos = y_max * 0.92  # 92%, not 105% (which goes outside)

# Set explicit limits with headroom
ax.set_ylim(y_min * 0.95 if y_min > 0 else y_min - 5,
            y_max * 1.05)

Checklist for Publication Figures

Use this checklist before finalizing figures:

• [ ] Plot type appropriate for data distribution (no violin on log scale)
• [ ] All text readable at publication size (8-10 pt minimum)
• [ ] Statistical annotations visible and within plot bounds
• [ ] Legend clear and doesn't obscure data
• [ ] Axis labels descriptive with units
• [ ] Color scheme colorblind-friendly
• [ ] Line weights balanced (not too thick or thin)
• [ ] Point sizes optimized (visible but not overlapping)
• [ ] DPI adequate for publication (300 minimum)
• [ ] Layout tested (try both horizontal and vertical if applicable)
• [ ] File format publication-ready (PNG, PDF, or SVG)

Common Refinement Patterns

Pattern 1: Decluttering Dense Plots

Problem: Too many visual elements competing for attention

Solution sequence: 1. Reduce point size (60 -> 25) 2. Thin line widths (2.5 -> 1.5) 3. Increase transparency (alpha=0.8 -> 0.5) 4. Reduce font sizes (10 -> 8) 5. Remove grid or make it lighter (alpha=0.3)

Before/After test: Generate both versions, compare

Pattern 2: Fixing Overflow Issues

Problem: Annotations, legends, or labels cut off

Solutions:

# 1. Adjust annotation positions
y_pos = y_max * 0.92  # Within bounds

# 2. Use bbox_inches='tight' when saving
plt.savefig('figure.png', dpi=300, bbox_inches='tight')

# 3. Explicitly set limits
ax.set_ylim(min_val * 0.95, max_val * 1.05)

# 4. Move legend outside plot area
ax.legend(bbox_to_anchor=(1.05, 1), loc='upper left')

# 5. Reduce text size
ax.text(..., fontsize=8)  # Down from 10

Pattern 3: Multi-Panel Layout Optimization

Try both orientations:

# Version 1: Horizontal (side-by-side)
fig, axes = plt.subplots(1, 2, figsize=(16, 7))
plt.savefig('fig_horizontal.png', dpi=300, bbox_inches='tight')

# Version 2: Vertical (stacked)
fig, axes = plt.subplots(2, 1, figsize=(10, 14))
plt.savefig('fig_vertical.png', dpi=300, bbox_inches='tight')

# Present both to user, ask which is clearer

Decision criteria:

• Horizontal: Better for direct comparison between panels
• Vertical: Better when each panel needs more space
• User context: Journal column width, presentation slides, etc.

Pattern 4: Iterative Statistical Annotation

Common issue: P-values positioned outside plot or overlapping with data

Solution:

# Calculate data range first
all_data = [data_dual, data_prialt]  # All datasets in plot
y_max = max([d.max() for d in all_data if len(d) > 0])

# Position relative to actual data, not theoretical maximum
for i, (x_pos, comparison) in enumerate(comparisons):
    stat, pval = stats.mannwhitneyu(...)

    # Safe positioning
    y_annotation = y_max * 0.92  # Below the top

    # Format text
    if pval < 0.001:
        text = 'p < 0.001***'
    elif pval < 0.01:
        text = 'p < 0.01**'
    elif pval < 0.05:
        text = 'p < 0.05*'
    else:
        text = f'p = {pval:.3f} ns'

    ax.text(x_pos, y_annotation, text, ha='center', fontsize=9)

# Set explicit limits to ensure annotations fit
ax.set_ylim(0, y_max * 1.05)

Refinement Workflow Example

Real case: VGP Figure 5 improvement sequence

1. Initial version: 4 categories, violin plots on log scale

• Issue: Violin distortion, too complex

2. V1 refinement: Remove violin plots, keep boxplots

• Better, but still issues

3. V2 refinement: Simplify to 3 categories

• Clearer interpretation

4. V3 refinement: Reduce point sizes (60->25), thin lines (2.5->1.5)

• Less clutter

5. V4 refinement: Test vertical vs horizontal layout

• Horizontal clearer for this case

6. V5 refinement: Fix p-value positioning (105%->92% of y_max)

• All elements now visible

7. Final: Smaller text in statistics box (10->8)

• Publication ready

Total iterations: 7 versions over refinement process Result: Clear, accurate, publication-quality figure

Best Practices

1. Version Your Refinements

Keep working versions during major changes:

scripts/
  plot_figure.py          # Original
  plot_figure_v2.py       # After major change (layout)
  plot_figure_final.py    # Publication version

2. Generate Alternatives in Parallel

When testing layout options:

# Save both versions
layouts = [
    ((1, 2), (16, 7), 'horizontal'),
    ((2, 1), (10, 14), 'vertical')
]

for (nrows, ncols), figsize, name in layouts:
    fig, axes = plt.subplots(nrows, ncols, figsize=figsize)
    # ... plot data ...
    plt.savefig(f'figure_{name}.png', dpi=300, bbox_inches='tight')

3. Document Each Refinement

"""
Figure 5 - Terminal Telomere Presence

Version history:
- v1: Initial 4-category version with violin plots
- v2: Removed violin plots (distortion on log scale)
- v3: Simplified to 3 categories (terminal only)
- v4: Reduced point/line sizes for clarity
- v5: Fixed p-value positioning
- final: Publication ready

Changes from v4 -> v5:
- P-value y-position: 1.05 * y_max -> 0.92 * y_max
- Added explicit y-axis limits: (y_min*0.95, y_max*1.05)
- Ensures all annotations visible within plot bounds
"""

4. Get Feedback at Key Milestones

Don't over-iterate without input:

• After fixing major issues (wrong plot type): Show user
• After layout changes (horizontal vs vertical): Show user
• After final polish: Show user

5. Maintain Consistency Across Figure Set

If refining one figure, check if same improvements apply to others:

# Applied violin->boxplot fix to Figures 2, 7, 10, 11
# Applied size reductions consistently across all figures
# Used same color scheme throughout

Summary

Systematic refinement workflow: 1. Identify issue -> 2. Fix visualization -> 3. Improve clarity -> 4. Test layouts -> 5. Optimize positioning

Key principles:

• Iterate based on user feedback
• Test alternatives (show options)
• Document changes
• Apply lessons across figure set
• Meet publication standards

Common adjustments:

• Point sizes: 60 -> 25
• Line widths: 2.5 -> 1.5
• Font sizes: 10 -> 8
• Annotation positions: 105% -> 92% of max
• Always set explicit axis limits

For additional guidance, see the supporting files:

• Publication standards (DPI, formats, sizes): publication-standards.md
• Multi-study result writing: multi-study-results.md
• Methodological transparency: methodological-transparency.md
• Overleaf package creation: overleaf-packages.md

Methodological Transparency in Figures and Statistics

The Dual Approach Pattern

When presenting statistical analyses, you often need to balance:

• Visual clarity: Clean figures that communicate patterns effectively
• Statistical rigor: Conservative tests using all available data

Best Practice: Explicit Documentation

If figures show cleaned/filtered data while statistical tests use the full dataset:

1. Figure captions must include explicit notes:

   \textbf{Note}: Statistical tests in Table X use [test name] on the full
   dataset (n=XXX) for conservative assessment; this figure shows cleaned
   data (outliers removed) for visual clarity only.

2. Methods section must explain the dual approach:

   \textbf{Visualization approach}: Figures show [cleaned data description]
   for visual clarity. All statistical tests use [full dataset description]
   for conservative and robust assessment. This dual approach ensures:
   (1) clear visual communication, and (2) statistically conservative
   hypothesis testing using all available data.

   \textbf{Outlier definition}: [Specific criterion, e.g., "Points beyond
   1.5x IQR from quartiles"]. Outlier removal applied only to visualization,
   not to statistical testing.

3. Why this matters:

• Prevents accusations of cherry-picking or p-hacking
• Shows scientific integrity and transparency
• Helps reviewers understand your methodology
• Demonstrates you're using conservative statistical practices

Example Use Case: Temporal Trends

Scenario: Scatter plots with many outliers obscure temporal trends

Solution:

• Figures: Remove outliers beyond 1.5x IQR for clean visualization
• Statistics: Use Spearman correlation on full dataset (all points)
• Documentation: Explicit notes in captions + Methods explanation

Template:

% In figure caption
Outliers removed for clarity (points beyond 1.5x IQR from quartiles).
\textbf{Note}: Statistical tests in Table S3 use Spearman correlation (rho)
on the full dataset for conservative assessment; this figure shows cleaned
data for visual clarity only.

% In Methods section
\textbf{Visualization approach}: Figures show scatter plots with regression
lines using cleaned data (outliers beyond 1.5x interquartile range removed)
for visual clarity. All statistical tests reported use the complete dataset
(including outliers) for conservative and robust assessment.

When to Use This Pattern

Use when:

• Outliers obscure visual patterns but should be included in tests
• You want both clear communication and rigorous statistics
• Submitting to high-impact journals (Nature, Science, etc.)
• Anticipating reviewer questions about data filtering

Don't use when:

• Outliers are actual data errors (remove from both)
• Sample size is too small to justify removal
• The outliers ARE the interesting pattern
• Methods would be simpler without this complexity

Writing Integrated Results from Multi-Study Analyses

Challenge: When you have multiple parallel analyses (e.g., same metrics across 5 different populations/clades/conditions), how to present findings coherently without overwhelming readers.

Solution: Organize by pattern type first, then by study

Structure Pattern

1. Universal Patterns Section

• Present findings consistent across ALL studies first
• This establishes the "baseline truth" readers can rely on
• Use strong language: "consistently," "across all," "universal"
• Provide statistical evidence from multiple studies

2. Study-Specific Patterns Section

• Present deviations and unique findings by study
• Explicitly contrast with universal patterns
• Explain why this study differs (biological/technical context)

3. Cross-Study Comparisons Section

• Tables comparing effect sizes across studies
• Discussion of what drives variation
• Statistical power considerations

Example Structure (from clade-specific genome analysis):

## Universal Patterns Across All Vertebrates

### Gap Density: Architecture Dominates Curation
- Finding: [Universal pattern]
- Evidence: [Stats from all 5 clades]
- Interpretation: [Why this is universal]

### Telomere Detection: Technology-Limited
- [Similar structure]

## Clade-Specific Patterns

### Mammals: Dual Curation Provides Benefits
- Finding: [Unique to this clade]
- Contrast: [How this differs from universal]
- Interpretation: [Biological context]

### Birds: No N50 Benefit from Dual Curation
- [Unique pattern and explanation]

## Cross-Clade Comparisons
- [Table of effect sizes]
- [Discussion of variation]

Benefits of This Structure

1. Readers get reliable findings first: Universal patterns are established before introducing complexity 2. Reduces cognitive load: Don't jump between studies repeatedly 3. Highlights what's generalizable: Universal section shows what works everywhere 4. Explains variation: Study-specific section explains why some results differ 5. Facilitates recommendations: Can give universal advice plus context-specific guidance

Writing Tips

For Universal Patterns:

• Lead with the finding, then provide evidence from multiple studies
• Use consistent statistical reporting across all supporting evidence
• Emphasize the consistency: "across all," "in every," "universal"

For Study-Specific Patterns:

• Explicitly state how this differs from universal patterns
• Provide biological/technical context for why this study is unique
• Don't just report statistics - explain the mechanism

For Statistical Power:

• Be explicit about which studies have sufficient power
• Note limitations in smaller studies
• Don't over-interpret null results from underpowered studies

Common Pitfalls to Avoid

• Don't: Report each study sequentially (Study 1 all results, Study 2 all results...)
• Do: Report by finding type (Finding A across all studies, Finding B across all studies...)

• Don't: Hide that some patterns aren't universal
• Do: Explicitly highlight when a pattern is study-specific and explain why

• Don't: Give equal weight to all findings
• Do: Emphasize universal patterns; note study-specific as "interesting variations"

Application Beyond Clade Analysis

This pattern works for any multi-study synthesis:

• Clinical trials across different populations
• Experimental treatments across multiple cell lines
• Algorithm performance across different datasets
• Policy interventions across different regions

Key principle: Organize by what readers need to know (universal vs specific) rather than by how you conducted the studies (study-by-study).

---

Providing Practical Recommendations from Complex Trade-offs

Challenge: When different methods excel at different outcomes, how to give clear guidance?

Pattern: "Depends on priority" recommendations with decision tree

Structure:

### For [Population/Context]

**Recommended**: [Method A]
- [Metric 1]: [Performance with stats]
- [Metric 2]: [Performance with stats]
- Use when: [Priority/constraint]

**Alternative**: [Method B]
- [Metric 1]: [Performance with stats]
- [Metric 2]: [Performance with stats]
- Use when: [Different priority/constraint]

**Note**: [Important caveat or key difference from other contexts]

Example (from avian genome assemblies):

### For Avian Genomes
**Depends on priority**:

**For gap density minimization**: Phased assembly (dual or single curation)
- Dramatic 75-100x reduction in gaps vs Pri/alt
- Strong significance (p=1.87e-10)

**For chromosome assignment**: Pri/alt + Single curation
- Best assignment (98.93% median)
- Significantly better than phased approaches (p<0.001)

**Note**: Dual curation does NOT improve scaffold N50 in birds (p=0.378),
unlike mammals. Initial assemblies are already near-optimal due to
favorable genome characteristics.

Benefits:

• Acknowledges trade-offs honestly
• Provides clear decision criteria
• Gives actionable guidance despite complexity
• Explains when different approaches are optimal

Creating Production-Ready Overleaf Packages

Package Structure

A complete Overleaf upload package should include:

Supplementary_Overleaf.zip
├── Supplementary_Information.tex      # Main LaTeX document
├── figures/                           # All figures in subdirectories
│   └── category_name/
│       ├── figure1.png               # 300 DPI minimum
│       └── figure2.png
└── README.txt                         # Compilation instructions

README.txt Template

Include comprehensive documentation for collaborators/reviewers:

SUPPLEMENTARY INFORMATION - OVERLEAF PACKAGE
============================================

VERSION HISTORY:
---------------
- v1.0 (date): Initial version with [description]
- v2.0 (date): Added [features]
- v2.1 (CURRENT): [Latest changes]

FILES INCLUDED:
--------------
1. Supplementary_Information.tex - Main LaTeX document
2. figures/category/*.png - Figure files (list each)

COMPILATION INSTRUCTIONS:
------------------------
1. Upload this entire zip file to Overleaf (New Project → Upload Project)
2. Overleaf will automatically detect Supplementary_Information.tex
3. Click "Recompile" to generate the PDF
4. First compile: 60-90 seconds (figures at 300 DPI)
5. Subsequent compiles: 20-30 seconds

EXPECTED OUTPUT:
---------------
- ~XX-YY page PDF document
- N figures embedded at 300 DPI
- M statistical tables with real data

NATURE METHODS COMPLIANCE:
-------------------------
- Font: Helvetica/Arial (sans-serif)
- Figure resolution: 300 DPI minimum
- Figure format: PNG (acceptable, PDF preferred for final)
- All figures with detailed captions
- Statistical reporting complete and accurate

TROUBLESHOOTING:
---------------
If compilation fails:
1. Check compiler: Menu → Settings → Compiler = "pdfLaTeX"
2. Verify all PNG files uploaded successfully
3. Check directory structure: figures/category/*.png
4. Try compiling twice (needed for table of contents)

Created: [date]
For: [journal name] submission
Purpose: [brief description]

Creation Workflow

1. Prepare directory structure:

   mkdir -p Supplementary_Overleaf/figures/category_name

2. Copy files:

   cp Supplementary_Information.tex Supplementary_Overleaf/
   cp figures/category/*.png Supplementary_Overleaf/figures/category/

3. Create README.txt with compilation instructions

4. Create zip package:

   cd Supplementary_Overleaf
   zip -r ../Supplementary_Overleaf.zip .
   cd ..
   ls -lh Supplementary_Overleaf.zip

5. Verify package contents:

   unzip -l Supplementary_Overleaf.zip

6. Test in Overleaf:

• Upload zip to new Overleaf project
• Verify compilation succeeds
• Check all figures render at correct resolution
• Verify all cross-references work

Version Control Documentation

Create a FINAL_PACKAGE_*.md file to track versions:

# Final Overleaf Package - [Description]

**Date**: YYYY-MM-DD (vX.X FINAL)
**Status**: READY FOR UPLOAD

## Version X.X - What Changed

### From vX.0 to vX.X:
**Changed**: [Description]
- **Was**: [Previous state]
- **Now**: [Current state]

**Why**:
- [Reason 1]
- [Reason 2]

## Package Contents (Final)

[List all files with sizes and purposes]

## Figure Details

[Description of each figure with key findings]

## Upload to Overleaf

### Quick Steps:
1. Go to https://www.overleaf.com
2. New Project → Upload Project
3. Select: `Package_Name.zip`
4. Click "Recompile"
5. Verify PDF appears

### Success Indicators:
- All figures visible
- All tables populated
- Cross-references work
- Professional appearance

Quality Checklist

Before finalizing package:

LaTeX document:

• All figures embedded with correct paths
• All tables populated (no placeholder values)
• All cross-references working (\ref{} commands)
• Methodological notes included where needed
• Professional formatting

Figures:

• All at 300 DPI minimum
• Correct file format (PNG/PDF)
• Clear, publication-quality appearance
• Proper directory structure

Documentation:

• README.txt with compilation instructions
• Version history documented
• Troubleshooting guide included
• Journal compliance noted

Package integrity:

• Reasonable file size (< 10 MB if possible)
• All files in correct locations
• Zip extracts cleanly
• No broken paths

Testing:

• Compiled successfully in Overleaf
• All figures render correctly
• PDF page count as expected
• No LaTeX errors or warnings

Publication Standards

DPI Requirements

• Screen/web: 150 DPI
• Print (standard): 300 DPI
• High-quality print: 600 DPI

File Formats

• Raster: PNG at 300 DPI (most journals accept)
• Vector: PDF or SVG (preferred for line plots, smaller file size, infinite zoom)
• Avoid: JPG (lossy compression, poor for scientific data)

Size Specifications

Check journal requirements:

• Single column: Usually 3.5 inches (89 mm) wide
• Double column: Usually 7 inches (178 mm) wide
• Height: Typically max 9-10 inches

Plan figsize accordingly:

# Single column figure
fig, ax = plt.subplots(figsize=(3.5, 4))

# Double column figure
fig, axes = plt.subplots(1, 2, figsize=(7, 3.5))

Color Accessibility Requirements

Many journals now require accessibility statements for figures, including:

• Confirmation that color schemes are colorblind-safe
• Use of validated palettes (Okabe-Ito, Paul Tol)
• Alternative distinguishing features (patterns, shapes, labels)

Nature journals specifically recommend:

• Okabe-Ito palette for categorical data
• Avoiding red-green combinations
• Testing figures with colorblindness simulators

In Methods section, document your color choices:

"All figures use the Okabe-Ito colorblind-safe palette (Okabe & Ito, 2008) to ensure accessibility for readers with color vision deficiencies."

Reference: Okabe, M. and Ito, K. (2008) Color Universal Design (CUD): How to make figures and presentations that are friendly to colorblind people. https://jfly.uni-koeln.de/color/