# Deterministic scientific figures — Phase 5

These are **offline figure exports generated from the actual Atlas datasets**.
They are not browser screenshots or evidence that browser file saving completed.

| Figure | Contents |
| --- | --- |
| A | Original unrooted support-collapsed tree, exact SH-aLRT/UFBoot labels and observed-core PDB counts |
| B | The same tree aligned with canonical protein architecture, analyzed PWWP cores and selected construct/observed-coordinate intervals |
| C | The same tree aligned with current domain-specific structural and chemical evidence counts |
| D | The supported HDGF/PSIP1 clade, original aligned residues and inference mask, descriptive conservation/occupancy, directly sourced cage annotations and frozen coordinate-contact markers |
| E | All 30 human PWWP domains, with disjoint exact-bound, measured-bioactivity-only and association-only compound/preparation profile counts |
| F | NSD2 PWWP1 versus NSD3 PWWP1: actual frozen Cα coordinates under the original least-squares transform, domain-specific counts and six separate source assay observations |

Each figure has light and dark versions in standalone SVG and high-resolution
PNG. SVG files contain source versions, source hashes, interpretation rules and
track definitions in their metadata. The 3× PNGs are rendered from those SVGs,
with no external image assets or browser automation. The preview PNGs are smaller
copies for inspection; use the files without `_preview` for exported graphics.

| Figure | SVG dimensions | PNG dimensions at 3× |
| --- | --- | --- |
| A | 1050 × 2084 | 3150 × 6252 |
| B | 1660 × 2441 | 4980 × 7323 |
| C | 1560 × 2149 | 4680 × 6447 |
| D | 1460 × 1760 | 4380 × 5280 |
| E | 1540 × 1770 | 4620 × 5310 |
| F | 1540 × 1810 | 4620 × 5430 |

| Figure | Light exports | Dark exports |
| --- | --- | --- |
| A | [SVG](Figure_A_light.svg) · [PNG](Figure_A_light.png) | [SVG](Figure_A_dark.svg) · [PNG](Figure_A_dark.png) |
| B | [SVG](Figure_B_light.svg) · [PNG](Figure_B_light.png) | [SVG](Figure_B_dark.svg) · [PNG](Figure_B_dark.png) |
| C | [SVG](Figure_C_light.svg) · [PNG](Figure_C_light.png) | [SVG](Figure_C_dark.svg) · [PNG](Figure_C_dark.png) |
| D | [SVG](Figure_D_light.svg) · [PNG](Figure_D_light.png) | [SVG](Figure_D_dark.svg) · [PNG](Figure_D_dark.png) |
| E | [SVG](Figure_E_light.svg) · [PNG](Figure_E_light.png) | [SVG](Figure_E_dark.svg) · [PNG](Figure_E_dark.png) |
| F | [SVG](Figure_F_light.svg) · [PNG](Figure_F_light.png) | [SVG](Figure_F_dark.svg) · [PNG](Figure_F_dark.png) |

The original topology, tip set, sequences, source coordinates, domain intervals,
branch supports and analysis files are unchanged. No new phylogenetic analysis
is implied. Geometric edge spacing has no evolutionary length scale. The 39
tips remain distinct, including NSD-family PWWP1 and PWWP2.

Architecture diagrams use one absolute amino-acid scale. Canonical source
intervals are 1-based and closed; drawn rectangle widths include both endpoint
residues. Annotation cores are not automatically certified complete folded
domains. Construct tags without a canonical mapping are omitted. Every domain
tip and its corresponding track use the same row coordinate.

Evidence counts come from the same read-only `/api/phase5/domain-evidence`
implementation as the application, evaluated locally without an HTTP service.
They preserve resolved identities versus unresolved preparation profiles and
keep domain-specific observations separate from protein-level outcomes.
Bioactivity-only counts include control or negative observations; they are not a
count of confirmed positive binders. Geometric contact counts are separate from
binding affinity. No clinical or pharmacological equivalence is inferred.

## Conservation and residue evidence — Figure D

The named HDGF/PSIP1 comparison group has seven source sequences: human HDGF,
HDGFL1, HDGFL2, HDGFL3 and PSIP1, rat Hdgf and mouse Hdgfl3. Its original
SH-aLRT/UFBoot values are 99.3/97 for the primary mask and 99.1/100 for the
sensitivity mask. No group or support value is newly inferred.

The display keeps the 115 columns that contain a residue in at least one of
these seven rows. The 142 omitted columns are all gaps in this selected clade;
their omission is a display projection, not a new inference mask. All original
1-based alignment column IDs and canonical residue positions remain available.
The frozen 78-site primary mask is preserved: 77 of its columns occur in this
display, and one is all-gap within the selected subgroup. Conservation is the
unweighted nongap canonical-20 Shannon score, `1 − H / ln(20)`. Occupancy is
nongap count divided by seven. These descriptive statistics do not establish
shared histone recognition, ligand binding or selectivity.

Only the source-verified HDGFL2 residues Y18, W21 and F44 receive the cage
outline, from Wu et al. (2011), Figure 4D–F,
https://doi.org/10.1371/journal.pone.0018919 . The historical source name is
HDGF2. Assignments are not transferred to other rows. Contact dots reuse the
frozen residue evidence with its original geometry scope and 5 Å threshold;
they are not affinity measurements. Unmarked residues are unannotated rather
than experimentally negative.

## Profile denominators — Figure E

Every bar partitions domain-specific profiles into mutually exclusive sets:
exact-identity target-specific PWWP-proximal bound evidence; measured evidence
without that verified structure; or remaining association-only evidence. One
profile represents a resolved identity or an unresolved preparation. A source
identity may contribute to more than one target-domain row. Bound PDB entries,
assay observations and typed negative observations are displayed separately
because they have different denominators. Negative/censored observations are
included in bioactivity-only profiles where appropriate. The figure retains all
30 human domain rows, including zero-record rows, and excludes protein-level
outcomes that lack domain attribution.

## Source coordinate projection — Figure F

The figure uses the original annotation-core Cα crops from 9FOC chain A (NSD2,
65 Cα atoms, canonical 222–286) and 6G25 chain A (NSD3, 64 Cα atoms, canonical
270–333). It applies the frozen row-vector least-squares transform to NSD2 and
keeps NSD3 fixed. No fit is recomputed. The source correspondence has 63 pairs.
Applying that transform to the frozen coordinate files reproduces the stored
RMSD of **0.9824190939280065 Å** to numerical precision; the exporter asserts
an absolute difference below `1e-10 Å`.

All three coordinate panels use the same fixed-frame Cartesian X/Y orthographic
projection and the same 8.2 pixels/Å scale. Cα trace segments require consecutive
canonical residues separated by at most 5 Å, so coordinate gaps are not joined.
The source RMSD is a three-dimensional metric. The panels show annotation cores,
not certified complete folded domains or ligand pockets. Structural similarity
does not establish chemical or biological equivalence.

Six assay examples retain separate source IDs, source units, uncertainty labels,
constructs and conditions. In particular, UNC6934's NSD2 binding observations
remain separate studies, and its NSD3 counter-screen retains the censored
`>30 µM` result with target-specific bound status. BI-9321 SPR and ITC results
remain separate measurements. Source links are embedded in the SVG. No potency
ranking across those different experiments is implied.

## Rebuild after a verified curation update

From the release root, in the dedicated Python environment:

```bash
python -m pip install -r scripts/export_phase5_figures.requirements.txt
python scripts/export_phase5_figures.py --scale 3
```

Node.js must be available. CairoSVG also requires the system Cairo library; use
the appropriate supported environment package if it is not already installed.
SVG generation uses the application's pure `evolution-layout.js` module. The
original analysis and the prepared chemical overlay must already be present.
This command never fetches upstream scientific data or promotes a candidate.

`figure_source_snapshot.json` contains the exact input projection. Source hashes
in `figure_manifest.json` identify the original analysis, original architecture
file, scientific database, compound overlay, shared geometry module and exporter
scripts, residue annotation sources, both coordinate crops and the frozen
transform. `figure_geometry.json` gives every displayed domain ID and row
position for independent alignment checking. The six A–C drawings must have
exactly the 39 source domain IDs, and each architecture/evidence row must share
its tip's y coordinate. D additionally checks its seven selected rows, 115
displayed columns, unchanged 78-site source mask and three cage assignments.
E checks all 30 human-domain partitions against API totals. F checks 129 Cα
atoms, 63 exact source correspondence pairs and the original transform RMSD.
SVG XML and PNG dimensions are validated during export for all twelve drawings.
The manifest's `supporting_files` records also hash this README, the exact source
snapshot and the geometry/validation JSON. These records support verified
downloads without claiming an impossible self-hash for the manifest itself.

Rasterizer: CairoSVG 2.9.1. Official API documentation was checked on
2026-10-26: https://cairosvg.org/documentation/ . The figures contain no copied
source imagery, article text or molecular cartoons.