Scientific Figure Analysis Pipeline

Problem

Scientific figures in papers are not just visuals. Volcano plots, UMAPs, heatmaps, Western blots, pathway diagrams, and microscopy panels often carry thresholds, scale bars, cluster labels, significance markers, and software-specific layout clues. Text-only extraction may capture titles and surrounding prose but misses in-figure values, panel relationships, and reproducible workflow signals.

How it works

The skill organizes PDF extraction, visual AI analysis, scientific interpretation, and reverse engineering into a pipeline:
- Stage 1: Parse scientific PDFs, detect and split a/b/c/d panels, and preserve original resolution, axis labels, legends, scale bars, and embedded annotations.
- Stage 2: Recognize chart types such as volcano plots, PCA, heatmaps, UMAP/t-SNE, phylogenetic trees, and Western blots, then combine image content, captions, nearby paragraphs, and scientific terminology for conservative reasoning.
- Stage 2B: When the current model lacks vision support or precise values are needed, call the Kimi K2.6 vision API to extract information that is hard to obtain from text alone.
- Stage 3: Infer likely software from fonts, palettes, layout, arrows, and annotation style, including GraphPad Prism, ggplot2, matplotlib/seaborn, Seurat/Scanpy, and ImageJ/Fiji, and reconstruct plausible workflow clues for clustering, dimensionality reduction, differential expression, and sequencing preprocessing.

Boundaries

It suits dense multi-panel figures, microscopy, sequencing, single-cell, and spatial transcriptomics visuals in Nature, Cell, Science, and bioRxiv papers. Outputs can be JSON or Markdown tables. The key constraint is conservative reasoning: unreadable labels, gene names, statistics, or method claims must be marked as uncertain rather than filled in as facts.