LongVerse figures

The figures from LongVerse: one sentence to the allele-specific methylome, verified multi-agent AI for Nanopore and PacBio long-read sequencing, with the code that draws them and the numbers they are drawn from.

Click the badge and a JupyterLab opens in your browser with everything installed. Open a notebook and run it. Nothing to download, no account. The first launch builds the image and takes ten to fifteen minutes; after that it starts in about half a minute.

---

Contents

notebooks/     one per figure: the code in the cell, the panel underneath
scripts/paper/ the panel scripts, as they are in the paper's repository
data/          what each panel draws, 128 MB
figures/all/   the batch render, each image named for the figure and panel it is
figures/       where a run of one notebook writes its PDFs and PNGs

notebook	figure
01_figure1	Fig. 1e, f - three callers and their consensus
02_figure2	Fig. 2a, b, c - one sentence reproduces the expert, ONT
03_extended_data_fig1	ED Fig. 1a, b, c - the remaining ONT panels
04_extended_data_fig2	ED Fig. 2a, b - PacBio calling
05_extended_data_fig3	ED Fig. 3a, b - PacBio phasing
06_extended_data_fig4	ED Fig. 4a-e - ONT, PacBio and WGBS across the genome
07_extended_data_fig5	ED Fig. 5a-h - coverage, imprinting and regional agreement
08_gnas_region	the region plot behind Fig. 2c - GNAS, per haplotype
00_all_figures	draws everything in one go

The notebooks are committed with their figures, so they can be read on GitHub without running anything.

The scripts

Each cell holds the script that drew that panel, taken from the paper's repository. Edit a line and re-run and you get your edit.

The four R panels are R cells: rpy2's %%R magic runs them in the same kernel as the Python ones. Each differs from its file by one line, marked where it happens, because the scripts read their arguments from a command line and a notebook has none.

Cells are generated from scripts/paper/, and python scripts/build_notebooks.py --check confirms they still match it.

Two panels do not run their script: ED Fig. 4a-c and ED Fig. 5h are drawn from the hexagons their own hexbin() call produced and exported, because the scripts for them read 450 MB and 890 MB inputs. figures/all/RENDER_LOG.tsv says which panel was drawn which way. Where a command line is shown, absolute paths on the shared filesystem read $LONGVERSE and $RESULTS.

The location

A panel script names its output after what it draws, because that is all it knows: panel_f_consensus.png is Figure 1e, and the novel imprinting heatmap is written by a script under fig_ed4 and is Extended Data Fig. 5d. Both names are useful and they are kept apart:

figures/all/       named for the figure: fig1e_consensus_vs_wgbs, ed5d_icr_novel_heatmap
                   FIGURE_PANELS.tsv maps each one back to the name the script gave it
figures/notebook/  named by the script, because the cell above it is that script and
                   the point of the notebook is that the two agree

figures/all/other_renders/ holds what the scripts produce that no figure places: the versions of a panel with both arms in one plot, where the paper draws one arm at a time.

The data

Whole-genome inputs are not needed to draw a figure, so what is here is what each panel reads, 196 MB in all.

per-run evaluation records	23 to 33 KB each
consensus and imprinting tables	0.6 to 19 KB
phased reads at the imprinting regions	116 MB
phased reads at GNAS, one pair per arm	2.9 MB
chromosome 20 per-site calls, four arms, and the bisulfite slice	66 MB
T2T annotation over those regions	777 KB
exported hexagons and DMRs	2.1 MB

The read slice was cut for this repository. The ninety imprinting control regions span 1.02 Mb once padded, so the phased BAMs go from 85 GB and 53 GB to 38 and 19 MB per haplotype. Base qualities are dropped, since NanoMethViz reads the MM and ML tags.

Running it elsewhere

git clone https://github.com/LabShengLi/longverse-figures
cd longverse-figures
conda env create -f environment.yml
conda activate longverse-figures
bash postBuild
jupyter lab notebooks

R here is 4.4.3 and NanoMethViz is 3.2.0, the versions the paper's figures were drawn on. postBuild compiles NanoMethViz rather than installing it prebuilt: the newest conda build is 2.8.1, and 2.8.1's plot_grange() does not take the arguments the paper's script passes.

R runs with --vanilla and R_LIBS pinned to the environment, so a personal R library cannot change what the figures look like.

From reads instead

The published chromosome 20 GNAS test set is 33 MB and runs the pipeline end to end in minutes: Zenodo record 23090404, pipeline at https://github.com/LabShengLi/longverse.