NMGUI2

A standalone desktop application for NONMEM pharmacometric modelling workflows

NMGUI2 is a PyQt6 desktop application that brings together everything a pharmacometrician needs in one window: browse and compare NONMEM models, evaluate goodness-of-fit, run PsN tools, analyse bootstrap and SIR results, visualise model lineage, and read .lst output — without switching between terminals, text editors and R.

It runs entirely offline on macOS, Windows and Linux. No browser. No server. No internet connection required during use.

Features
User guide — full reference for every tab, dialog, and required R/PsN package
Dependencies overview
Quick start
Installation — macOS
Installation — Windows
Installation — Linux
First use
Updating NMGUI2
Keyboard shortcuts
Configuration files
Troubleshooting
Contributing
Author
Acknowledgements
License
Changelog

Features

Models tab

Scans a directory for .mod / .ctl files and displays all models in a sortable, filterable table
Columns: OFV, ΔOFV (relative to user-selected reference), minimisation status, covariance step, condition number, estimation method, individuals, observations, parameters, AIC, runtime
Colour-coded rows: green = converged, red = failed/terminated, orange = stale or boundary warning
Filter buttons: All, Completed, Failed — plus free-text search
New model… (Ctrl+N / ⌘N) — create a blank NONMEM control file from one of 17 built-in templates (analytical ADVAN1/2/3/4 and ODE-based ADVAN6 including Michaelis-Menten, QE-TMDD, time-varying CL, dual/mixed absorption, urine compartment); preview before writing; auto-selects the new file after creation
Right-click context menu: run, toggle star, duplicate, set/clear reference, compare, copy paths, open folder, view .lst, view .ext, QC report, run report, view run record, NM-TRAN messages, workbench, delete
Keyboard: ↑↓ move rows, Space toggles star, Enter jumps to Output

Model detail panel

Parameters — full THETA/OMEGA/SIGMA table with names (parsed from control stream comments), estimates, SE, RSE%, 95% CI, SD for variance parameters. Handles BLOCK(n) and BLOCK(n) SAME. Export to CSV and HTML.

Editor — syntax-highlighted .mod editor with save functionality.

Run — launch any PsN tool (execute, vpc, bootstrap, scm, sir, cdd, npc, sse). Each run opens its own floating popup with a live console, iteration/OFV progress, elapsed timer, and Stop button (SIGTERM or SIGKILL). Multiple models run simultaneously. Run detached (Linux/macOS) launches under nohup so the job survives SSH disconnection or app closure — auto-checked when running over SSH. The Active & Recent Runs table shows live, detached, and historical runs; history persists per project across restarts.

Info — status tag (base/candidate/final/reject), comment, free-form notes, and a dataset integrity card (missing file, non-monotonic TIME, duplicate doses, extreme DV, high BLQ).

Output — structured HTML rendering of the .lst in-panel: summary card, estimation steps table (per-step OFV/ΔOFV/runtime for chained $EST), NM-TRAN warnings, convergence table, parameter estimates with SEs, ETABAR p-values, shrinkage, correlation/covariance matrices, eigenvalues, condition number. "Open in browser" exports full HTML.

Files tab (Ctrl+2)

A built-in file browser with subfolder navigation and a content viewer — no external app needed.

Navigation bar — ← back button, breadcrumb path, and extension filter pills (All · .mod · .ctl · .lst · .tab · .csv · .ext · .cov · .cor · .phi · [+]). All files shown by default; selecting pills filters files while keeping folders always visible.
File list — folders appear first (bold, 📁 prefix); files below. Sortable by name, size, or date.
Double-click folder → navigate into it; breadcrumb and back button update.
Double-click file → open with the OS default application.
Content viewer — syntax-highlighted text for .mod/.ctl; inline find bar; virtualised spreadsheet view for .csv/.tab (no row cap); Table / Plot toggle showing the full Data Explorer (scatter, LOESS, colour-by, filters); Edit / Save / Discard for text and CSV files.

Tree tab (Ctrl+3)

Interactive force-directed node graph of model lineage based on PsN metadata or manually set parent. Zoom, pan, double-click any node to jump to that model. Gold border = starred; green border = final.

Evaluation tab (Ctrl+4)

Native diagnostic plots — no R required:

GOF 2×2 — DV vs PRED, DV vs IPRED, CWRES vs TIME, CWRES vs PRED with reference lines
CWRES histogram — normal overlay, Save PNG
QQ plot — with Shapiro-Wilk test and 95% confidence band, Save PNG
NPDE distribution — shown when NPDE column is present, Save PNG
ETA vs covariate — scatter with LOESS overlay; recognises truncated NONMEM ETA column names (ET12, ET13 …)
Individual fits — paginated DV/IPRED/PRED vs TIME per subject
OFV waterfall — ranked ΔOFV across all completed models
Convergence traces — parameter trajectories from .ext

Auto-loads sdtab on model selection.

VPC tab (Ctrl+5)

Generate Visual Predictive Checks via R:

vpc (Ronny Keizer) and xpose (tidyverse) backends — availability shown with ✓/✗ at startup
"Use PsN settings" mode inherits binning, stratification and pred-corr from the PsN output folder
Configurable PI/CI percentiles, LLOQ, bins, log Y, pcVPC
Editable R script — modify before running
Live console, inline PNG preview, Save high-res PNG, Save PDF

Uncertainty tab (Ctrl+6)

Bootstrap — run via PsN or load existing results; parameter table with 95% CI from percentiles, forest plot, diagnostic checks (completion rate, bias, correlations, CI validity, boundary proximity), overall PASSED/ACCEPTABLE/WARNING/FAILED verdict.

SIR — same layout; ESS, resampling efficiency, degeneracy detection, weighted percentile CIs, overall assessment.

Sim Plot tab (Ctrl+7)

Prediction interval visualisation from NONMEM Monte Carlo simulation output:

Load any NONMEM table file or CSV — no row cap
Auto-detects replicate column (REP, IREP, SIM, SIMNO, …)
Up to 4 configurable PI band ribbons (percentile pair, colour, alpha, visibility), 4 presets
Median line, log/linear Y axis, LOESS smoothing
Up to 6 column filters (AND), MDV=1 exclusion
Optional observed data overlay
Background quantile computation — UI stays responsive
Save 300 DPI PNG

History tab (Ctrl+8)

Chronological log of all PsN runs (last 200 entries). Columns: status, model, tool, command, timestamps. Double-click any row to view the full run record.

Run Records (audit trail)

Every run creates an immutable record: UUID, SHA-256 hashes of control stream and dataset, NONMEM/PsN/NMGUI versions, timestamps, OFV, minimisation/covariance status, full command. Accessible from right-click → "View run record". Export JSON.

Settings tab (Ctrl+9)

Dark / light / follow-system theme toggle
Path configuration for NONMEM, PsN, R, RStudio
Directory bookmarks management
GitHub update check (numeric version comparison)
Debug logging to ~/.nmgui/nmgui_debug.log

Additional features

Model comparison — side-by-side parameter table with ΔOFV, ΔAIC, ΔBIC, LRT p-value
Model workbench — sortable table of all completed models for quick multi-model review
QC report — PASS/WARN/FAIL HTML checklist per model
Duplicate model — copy with incremented run number
NM-TRAN messages — compilation warnings/errors at a glance
Stale detection — orange highlight when .mod or data file is newer than .lst
GitHub update check — optional notification on new version
Cross-platform — macOS, Windows, Linux (including X11/MobaXterm over SSH)

Dependencies overview

Python packages

All Python dependencies are listed in requirements.txt and installed automatically:

Package	Purpose
PyQt6	GUI framework
pyqtgraph	Interactive plots
numpy	Numerical operations
matplotlib	CWRES histogram, QQ plot
scipy	Shapiro-Wilk test, confidence bands (optional but recommended)

External tools for running models

Dependency	Minimum version	Notes
NONMEM	7.4	Must be installed and licensed
PsN (Perl-speaks-NONMEM)	5.0	Must be on system PATH
Perl	5.16	Required by PsN

External tools for VPC generation

Dependency	Notes
R ≥ 4.0	Must be on system PATH (`Rscript` command must work)
RStudio	Optional but recommended
R package: vpc	`install.packages("vpc")`
R package: xpose	`install.packages("xpose")`
R package: tidyverse	`install.packages("tidyverse")` — required by xpose
R package: ggplot2	`install.packages("ggplot2")` — required by vpc/xpose

NMGUI2 (browsing, evaluating, comparing results) works without NONMEM, PsN, R or RStudio. These are only needed if you want to run models or generate VPCs from within the app.

Quick start

If you already have Python 3.10+, Git, and optionally R/PsN installed:

git clone https://github.com/robterheine/nmgui2.git
cd nmgui2
pip install -r requirements.txt
python3 nmgui2.py

For full installation instructions including NONMEM, PsN, and R, see the platform-specific sections below.

Installation — macOS

1. Install Xcode Command Line Tools

xcode-select --install

2. Install Homebrew (if not already installed)

/bin/bash -c "$(curl -fsSL https://raw.githubusercontent.com/Homebrew/install/HEAD/install.sh)"

3. Install Python 3

brew install python
python3 --version

4. Install R (required for VPC only)

brew install r

Or download from https://cran.r-project.org/bin/macosx/

Verify:

Rscript --version

5. Install RStudio (optional)

Download from https://posit.co/download/rstudio-desktop/ and add to /Applications.

6. Install required R packages

Rscript -e 'install.packages(c("vpc","xpose","tidyverse","ggplot2"), repos="https://cran.r-project.org")'

7. Install Perl (required by PsN — usually pre-installed on macOS)

perl --version

If missing:

brew install perl

8. Install PsN

Follow the official guide at https://uupharmacometrics.github.io/PsN/install.html

Verify:

execute --version

9. Clone and install NMGUI2

git clone https://github.com/robterheine/nmgui2.git
cd nmgui2
pip3 install -r requirements.txt

10. Run NMGUI2

python3 nmgui2.py

Optional: desktop launcher

Create ~/Desktop/NMGUI2.command:

#!/bin/bash
cd /path/to/nmgui2
python3 nmgui2.py

Make it executable:

chmod +x ~/Desktop/NMGUI2.command

Double-click in Finder to launch.

Installation — Windows

1. Install Python 3.10 or newer

Download from https://www.python.org/downloads/windows/

During installation, tick:

✅ Add Python to PATH
✅ Install pip

Verify in Command Prompt:

python --version
pip --version

2. Install R (required for VPC only)

Download from https://cran.r-project.org/bin/windows/base/

During installation, tick "Add R to system PATH" (or add manually: C:\Program Files\R\R-4.x.x\bin to PATH).

Verify:

Rscript --version

3. Install RStudio (optional)

Download from https://posit.co/download/rstudio-desktop/

4. Install required R packages

Open Command Prompt or RStudio and run:

Rscript -e "install.packages(c('vpc','xpose','tidyverse','ggplot2'), repos='https://cran.r-project.org')"

5. Install Perl (required by PsN)

Download Strawberry Perl from https://strawberryperl.com/ — this includes all required modules.

Verify:

perl --version

6. Install PsN

Follow the official guide at https://uupharmacometrics.github.io/PsN/install.html

Verify:

execute --version

7. Install Git

Download from https://git-scm.com/download/win

8. Clone and install NMGUI2

Open Command Prompt:

git clone https://github.com/robterheine/nmgui2.git
cd nmgui2
pip install -r requirements.txt

9. Run NMGUI2

python nmgui2.py

Optional: desktop launcher

Create NMGUI2.bat in the nmgui2 folder:

@echo off
cd /d %~dp0
python nmgui2.py

Double-click to launch.

Installation — Linux

Shown for Ubuntu/Debian. For Fedora replace apt with dnf; for Arch replace with pacman.

1. Install system dependencies

sudo apt update
sudo apt install -y python3 python3-pip python3-venv git curl perl
python3 --version

2. Install Qt system libraries (required by PyQt6)

sudo apt install -y libxcb-xinerama0 libxcb-cursor0 libxkbcommon-x11-0 \
    libgl1-mesa-glx libegl1

3. Install R (required for VPC only)

sudo apt install -y r-base r-base-dev

For the latest R version, follow https://cran.r-project.org/bin/linux/ubuntu/

Verify:

Rscript --version

4. Install RStudio (optional)

Download the .deb installer from https://posit.co/download/rstudio-desktop/

sudo dpkg -i rstudio-*.deb
sudo apt --fix-broken install

5. Install required R packages

Rscript -e 'install.packages(c("vpc","xpose","tidyverse","ggplot2"), repos="https://cran.r-project.org")'

6. Install PsN

Follow the official guide at https://uupharmacometrics.github.io/PsN/install.html

Verify:

execute --version

7. Clone and install NMGUI2

git clone https://github.com/robterheine/nmgui2.git
cd nmgui2
pip3 install -r requirements.txt

Or with a virtual environment (recommended on systems with externally-managed Python):

git clone https://github.com/robterheine/nmgui2.git
cd nmgui2
python3 -m venv venv
source venv/bin/activate
pip install -r requirements.txt

8. Run NMGUI2

python3 nmgui2.py

Or if using a virtual environment:

source venv/bin/activate
python3 nmgui2.py

Optional: desktop launcher

Create ~/.local/share/applications/nmgui2.desktop:

[Desktop Entry]
Type=Application
Name=NMGUI2
Exec=/bin/bash -c "cd /path/to/nmgui2 && python3 nmgui2.py"
Terminal=false
Categories=Science;

Updating NMGUI2

On all platforms, from the nmgui2 directory:

git pull
pip install -r requirements.txt

The second command ensures any new dependencies are installed. If no new packages were added, it completes instantly.

If using a virtual environment on Linux:

source venv/bin/activate
git pull
pip install -r requirements.txt

Previous versions

All tagged releases from v2.5.0 onward are available on the Releases page. To run an older version, check out the tag into a separate directory:

git clone https://github.com/robterheine/nmgui2.git nmgui2-old
cd nmgui2-old
git checkout v2.8.0
pip install -r requirements.txt
python3 nmgui2.py

First use

Launch the app (python3 nmgui2.py)
Click Browse… or use ⌘O / Ctrl+O and navigate to a folder containing .mod files
Click + Bookmark to save the directory for quick access
Select a model row to view its parameters, output, and diagnostic plots
Go to Settings (⌘9 / Ctrl+9) to configure paths if PsN/NONMEM/RStudio are not auto-detected

Keyboard shortcuts

Action	macOS	Windows / Linux
Models tab	⌘1	Ctrl+1
Files tab	⌘2	Ctrl+2
Tree tab	⌘3	Ctrl+3
Evaluation tab	⌘4	Ctrl+4
VPC tab	⌘5	Ctrl+5
Uncertainty tab	⌘6	Ctrl+6
Sim Plot tab	⌘7	Ctrl+7
History tab	⌘8	Ctrl+8
Settings tab	⌘9	Ctrl+9
Open directory	⌘O	Ctrl+O
Rescan directory	⌘R	Ctrl+R
New model	⌘N	Ctrl+N
Navigate model table	↑ / ↓	↑ / ↓
Jump to Output panel	Enter	Enter
Toggle star	Space	Space

Configuration files

All settings are stored in ~/.nmgui/ (created automatically on first run):

File	Contents
`settings.json`	Theme, paths, window geometry, splitter sizes
`model_meta.json`	Stars, comments, status tags, notes, parent model
`bookmarks.json`	Directory bookmarks
`runs.json`	Run history (last 200 entries)
`nmgui_debug.log`	Debug log for troubleshooting

These files are written inside each project folder (not in ~/.nmgui/):

File	Contents
`nmgui_run_records.json`	Immutable audit trail of runs for that project (last 500 entries), used by the Active & Recent Runs table
`<run_id>.nmgui.log`	stdout/stderr of a detached run; tailed by the Watch Log window
`<run_id>.nmgui.pid`	PID and metadata of a running detached job; removed automatically on completion

To reset all settings: delete the ~/.nmgui/ folder. Run records in individual project folders are not affected.

Troubleshooting

App fails to launch — ModuleNotFoundError: No module named 'PyQt6' Dependencies did not install. Rerun pip install -r requirements.txt (or pip3 on macOS/Linux). If you see "externally-managed environment" on Linux/macOS, use a virtual environment (see the Linux section).

App launches but window is blank or scrollbars are native-grey Your Qt style is overriding the stylesheet. NMGUI2 forces Fusion at startup; if you see this, confirm you are on v2.5.0 (git pull && pip install -r requirements.txt).

"R: vpc ✗ xpose ✗" in the VPC tab Rscript is not on PATH or the R packages are not installed. Verify Rscript --version in a shell, then install packages with

Rscript -e 'install.packages(c("vpc","xpose","tidyverse","ggplot2"), repos="https://cran.r-project.org")'

execute --version not found PsN is not on PATH. On Linux/macOS, add the PsN bin directory to your shell profile (~/.zshrc, ~/.bashrc). On Windows, add it to the system PATH via Environment Variables.

NM-TRAN messages show "parser.py not available" Your nmgui2 repository is outdated. Run git pull.

Models tab shows models but OFV column is empty The .lst file is missing or incomplete — the run never finished or was deleted. Check the model's run directory.

Dark mode colours look wrong after toggling theme Fixed in v2.5.0. If you still see stale colours, restart the app.

Everything else — enable debug logging and share ~/.nmgui/nmgui_debug.log when filing an issue.

Contributing

Contributions are very welcome. See CONTRIBUTING.md for details.

Areas particularly in need of help:

Windows and Linux testing and bug reports
Additional NONMEM output parsing (BAYES, mixture models)
New diagnostic plot types
Documentation and tutorials
Translations

Author

Rob ter Heine Hospital pharmacist – clinical pharmacologist Radboud Applied Pharmacometrics · Radboudumc, Nijmegen, the Netherlands

Acknowledgements

Developed with Anthropic Claude.

License

MIT License — free to use, modify and distribute.

Changelog

v2.9.25 — Fix: axis change did not update plot (LOESS blocking main thread)

Two fixes for the Plot view in the Files tab and data explorer:

LOESS blocked the UI: loess() is O(n·k) per output point where k ≈ frac × n. With 50 000 rows (raised in v2.9.24) and the default frac=0.4, each of 80 output points ran np.argsort on 50 000 elements — ~30–60 s of blocking. The Qt event loop was frozen; the user saw no response to axis changes. Fix: subsample the input to at most 2 000 evenly-spaced points before fitting. A LOESS trend curve is indistinguishable between 2 000 and 50 000 points. The sim_canvas LOESS (operating on 20–100 quantile values) is unaffected.
View range stuck after axis change: added pw.autoRange() at the end of _plot() so the viewport resets to fit the new data when columns are switched.

v2.9.24 — Fix: Plot tab blank in Files tab viewer; smarter axis auto-selection

Two fixes for the Plot view in the Files tab:

Blank plot on load: _load_data() only auto-selected axes when columns were named exactly PRED and DV. Tables with TIME/IPRED (typical sim output) left both dropdowns empty and _plot() was never called. Fixed: extended auto-selection with ordered fallback lists TIME→TAD→PRED→IPRED (X) and DV→IPRED→PRED→CWRES→... (Y), plus an explicit _plot() call at the end of _load_data() to guarantee a render regardless of which columns fired the signal.
Row cap too low for plot: v2.9.23 introduced a 5 000-row cap for the table viewer; the same cap was passed to the plot, making a 2M-row sim table visible as only 0.25% of the data. Raised to 50 000 rows — meaningful for any typical run.

v2.9.23 — Fix: "Convert to CSV" in Files tab truncated multi-subproblem tables

Bug: Converting a NONMEM table file to CSV (right-click → Convert to CSV) produced an output containing only the first subproblem / first set of individuals. Root cause: _read_nonmem_table() used break when it encountered an embedded TABLE NO. banner, stopping at the end of the first replicate block.

Fix: Changed break → continue so the parser skips embedded banners and reads all sections. Repeated column-header rows between sections are also filtered out. The table viewer now passes max_rows=5000 to keep the interactive display responsive; the converter still reads the full file with no row cap.

Affects any multi-subproblem $SIMULATE output or FIRSTONLY table file. Single-table files were unaffected.

v2.9.22 — Sim Plot performance: 10-30× faster file loading for large simulations

Five targeted performance fixes for the Sim Plot tab, most impactful on Linux with large simulation datasets (1 000 virtual subjects, many replicates):

Fix 1 — Fast file loading (parser.py): new read_table_df() function delegates to pd.read_csv() with the C engine after format detection. For a 1 M-row sim table this drops load time from ~20–30 s to ~1–2 s. Automatic fallback to the row-by-row parser on any error (D-notation, encoding edge cases, old pandas).
Fix 2 — Slim DataFrame copy (sim_canvas.py): _SimWorker now copies only the 3–5 columns actually needed instead of the full 30–50 column DataFrame, reducing memory allocation ~10×.
Fix 3 — Faster pivot (sim_canvas.py): replaced pivot_table(aggfunc='mean') with groupby().mean().unstack(), which avoids pivot_table's generic group-key overhead (~3× faster for typical sim data).
Fix 4 — Batched quantiles (sim_canvas.py): all required quantile levels (lo/med/hi across all bands) are now computed in one np.nanquantile() call instead of three per band.
Fix 5 — LOESS weight matrix (format.py): removed np.diag(w) (allocated a k×k matrix just for a diagonal multiply); replaced with element-wise A * w[:, None].

v2.9.21 — Init→Final visualization redesign: log-scale ratio bar

A pharmacometrician review of the v2.9.18 bullet bar identified four UX failures:

Wide-bound collapse — $THETA (0, init, 1e6) (the NONMEM default) mapped both the initial tick and final marker into the leftmost ~1px of the track. Most parameters showed a dot on the left edge of empty wasteland.
FIXED diamond didn't read as "fixed" — looked like a stray marker that lost its track.
Initial tick invisible when init ≈ final — the 1px tick was occluded by the 6px marker exactly in the common "well-behaved fit" case.
OMEGA/SIGMA marker landed near the middle regardless of movement.

v2.9.21 replaces the bullet bar with a log-scale ratio bar centered on the initial estimate. Same delegate, same SVG helper, same payload — only the paint logic changed.

What you see now

For a normal parameter:

[ 0.1×    1× •─●    10× ]   ×2.6
        center    marker     inline label

Track always spans 0.1× → 10× initial (2 decades, log scale) — same scale for every parameter, independent of bound width. The wide-bound collapse is impossible by construction.
Center tick at 1× (the initial estimate as reference).
Filled circle at log10(final/initial), clamped to track ends.
Chevron at the clamped end when the movement is off-scale (>10× or <0.1×).
Inline ×N numeric label (×1.6, ×0.50, ×12) removes the visual-translation step. Compact (Latin-1 × symbol only — no glyphs that fail on older Windows fonts).

FIXED parameters

Render as a thin uniform grey horizontal line across the full track width, no marker, no tick, no label. Reads visually as "frozen — nothing to show." Combined with the existing FIX italic on the parameter name, FIXED rows are quiet and unambiguous.

Boundary proximity

When the final estimate is within 1% of a bound, the marker turns red AND a red wall-line is drawn at the corresponding track edge. This signal is independent of the log-ratio scale, so wide bounds no longer suppress it.

Color thresholds

Same conceptual tiers as v2.9.18, now driven by |log10(final/initial)|:

< 0.04 (within ~10%): subtle grey
< 0.18 (within ~50%): accent blue
>= 0.18: orange
At bound: red

Edge cases (rare in PMx but handled correctly)

Sign change (init < 0, final > 0 or vice versa): log scale undefined. Falls back to a signed delta badge (+0.8 / -0.5) with no bar.
Initial = 0: ratio undefined. Falls back to the bare final value (0.5) with no bar. The absence of a bar IS the visual signal.
Both negative: log scale of final/initial works (positive ratio), marker positioned normally.

Cross-OS

Pure QPainter primitives on the desktop and pure SVG primitives in the HTML report. All glyphs are Latin-1 only (× U+00D7) — no symbols that may render as boxes on older Windows fonts. Verified by encoding every SVG variant as Latin-1.

Sizing

Desktop column width bumped from 70 to 110px to accommodate the bar + inline label.
HTML SVG width bumped from 80 to 110px; CSS .viz rule widened to 120px.

Verified with 30 inline checks across 10 acceptance criteria

Wide-bound case now positions marker meaningfully (not collapsed). FIXED rows render as quiet lines (no spurious markers). No-movement cases show marker at center with ×1.0 label. Off-scale (>10×) cases render a chevron. At-bound cases render red marker + red wall. Negative same-sign ratios work correctly. Sign-change and zero-initial cases fall back gracefully to ASCII numeric badges. Run19.lst regression: all 25 viz cells render meaningfully. Desktop and HTML formatters agree byte-for-byte. All 8 SVG variants encode as Latin-1.

v2.9.20 — Fix: Init→Final viz cells were blank (wiring gap in ScanWorker)

The Init→Final visualization shipped in v2.9.18 (desktop) and v2.9.19 (HTML) was silently invisible in real usage. parser.parse_lst() populated the five new fields (theta_initials, theta_lowers, theta_uppers, omega_initials, sigma_initials) correctly, but ScanWorker in app/workers.py iterates over a hardcoded allowlist of field names to copy from the parser result into each model dict — and the allowlist wasn't extended in v2.9.18.

Result: the field-copy loop dropped all five fields on the floor; the parameter table and HTML report saw empty lists; every viz cell rendered blank with no error message. The bug only manifested in the actual app flow; my v2.9.18 test suite called parse_lst() directly, bypassing the worker.

Fix: five new field names added to the ScanWorker allowlist. End-to-end regression test now simulates the worker → model-dict → ParameterTable → HTML report pipeline, confirming all 25 viz cells in a real run19.lst model are populated correctly (18 diamonds for FIXED params, 7 tracks for non-FIXED).

v2.9.19 — Init→Final visualization in the HTML report

Mirrors the v2.9.18 desktop Init→Final column into the HTML run report. Inline SVG primitives — the report stays a single self-contained file with no external assets.

New "Init→Final" column in the parameter table of generated HTML reports, between Estimate and SE.
Same visualization logic as the desktop delegate: track + initial tick + final marker for bounded params, diamond for FIXED, auto-scaled-with-dashed-right for unbounded (OMEGA/SIGMA).
Same color thresholds: subtle grey (<10% movement), accent blue (10-50%), orange (>50%), red marker + red wall-line at bound (within 1%).
Tooltip on hover: Initial: X • Final: Y (+Z%) • Bounds: [lo, hi] via the HTML title attribute.
Implementation: a single _init_final_svg() helper produces an inline <svg> per row using <rect>, <line>, <circle>, <polygon> primitives. Mirrors the QPainter rendering in widgets/parameter_table.py byte-for-byte. CSS rule .viz {padding:2px 6px; line-height:0; vertical-align:middle;} keeps the SVG aligned in the table.
Cross-browser: tested HTML opens in Safari / Chrome / Firefox without external dependencies. No JavaScript. No web fonts.
Verified with 50 inline checks across SVG helper extremes (14 payload shapes including zero-init/zero-final/lower>upper/None values), full report generation against real run19.lst with extract_param_names + parse_lst merged (18 diamonds for FIXED params, 7 tracks for non-fixed, structural HTML well-formedness), and color-threshold verification across each tier.

v2.9.18 — Init→Final visualization in the parameter table

New column "Init→Final" in the Models-tab parameter table (column index 3, between Estimate and SE). Inspired by Pirana — shows where each parameter's final estimate sits relative to its initial estimate and bounds.
Visualization: a small bullet bar with a tick at the initial estimate and a filled circle at the final estimate. Track spans the user-specified bounds for THETAs that have them; auto-scaled (and dashed on the right) for OMEGAs / SIGMAs that don't have an upper bound.
Diagnostic color cues on the final marker:
- Subtle grey: parameter moved < 10% from initial
- Blue (accent): 10–50% movement
- Orange: > 50% movement (initials may have been off, worth a glance)
- Red + red wall-line at the bound: final is at or beyond a bound (within 1%) — the optimizer is pushing against the constraint; structural model or bounds need re-examination.
FIXED parameters: small filled diamond, no track, no movement (visually distinct from "parameter that happens to land near its initial").
No data (model never run, or .lst lacks the INITIAL ESTIMATE echo): cell stays blank and the rest of the table works normally.
Cross-OS bulletproof: every visual element drawn with QPainter primitives (drawRoundedRect, drawLine, drawEllipse, drawPolygon). No Unicode glyphs, no SVG, no font fallback issues on Windows / Linux / macOS.
Tooltip on each viz cell shows the numbers: Initial / Final (Δ%) / Bounds.
New parser function _parse_initial_estimates() in parser.py extracts initial estimates and bounds from NONMEM's 0INITIAL ESTIMATE OF THETA/OMEGA/SIGMA: echo block. Authoritative — it's exactly what NONMEM ran with, regardless of post-run edits to the .mod.
Tested against 51 cases: real run19.lst (15 THETAs / 8 OMEGAs / 2 SIGMAs), synthetic models with extreme values (1e-12, 1e10, negative bounds, equal bounds), D-notation Fortran doubles, OMEGA block matrices (3×3 with 6 elements), missing INITIAL echo, models with no .lst, large models (200 THETAs load in ~7ms), and a delegate-stress matrix of 16 payload shapes including zero-initial, zero-final, lower>upper bad data, and None-valued fields. All 51 pass.

v2.9.17 — Line-number gutter in the model editor

Line numbers now appear in a left-edge gutter alongside the Models-tab editor. Removes the manual counting step when NONMEM reports an error like ERROR ENCOUNTERED IN LINE 47 OF THE CONTROL STATEMENTS.
Cursor-position indicator Ln X, Col Y shown right-aligned in the editor's toolbar row.
Implemented as a new CodeEditor (subclass of QPlainTextEdit) using Qt's canonical line-number-area pattern. The existing NMHighlighter syntax highlighting, palette, font, and all editor operations (Save, View .lst, model load) are unchanged.
Gutter width is dynamic with a 3-digit minimum reservation to prevent jitter on small files. Numbers right-aligned in a muted color; the current line's number is brighter. Theme-aware — repaints correctly on light/dark switch.
Hook for future error-jumping: editor.goto_line(n) is a public method that scrolls to line N and briefly highlights it, plus a lineJumped(int) signal. Click-to-jump from NMTRAN error output is intentionally deferred to a separate release — it needs $INCLUDE handling, error-panel UI design, and lifecycle plumbing that don't belong in the gutter PR.

v2.9.16 — Parser fix: $-in-comment no longer truncates parameter blocks

Fixes a silent-correctness bug in inject_estimates (used by the Duplicate dialog's "Inject final estimates from .lst" option) and a cosmetic issue in extract_table_files.

Root cause. Four regex sites in parser.py captured $THETA / $OMEGA / $SIGMA / $TABLE record bodies using [^\$]* or (.*?)(?=\$|\Z). Both patterns terminate at any literal $ — including $ that appears inside a comment (e.g. ; $5 paper Smith et al.). When that happened, the captured block was truncated and subsequent parameter values (or table-file entries) were silently missed.
Impact when it fired. In inject_estimates, the duplicated .mod was written with mixed old/new initial estimates — the user got a model that fits with the wrong starting point and no error message. In extract_table_files, the table-file listing for that model was incomplete.
Fix. All four sites now use (.*?)(?=\$[A-Za-z]|\Z) (with re.DOTALL). The lookahead requires the next $ to be followed by a letter — i.e. a real NM-TRAN record like $OMEGA, not a $digit in a comment.
Residual edge case (deliberately not fixed). A comment containing $letter (e.g. RCS keywords like $Id$ , or markers like $Patient) would still mis-scope the block. These are rare in hand-written NONMEM control streams; the bulletproof fix (enumerating all NM-TRAN keywords or stripping comments before scoping) is held back as a separate change if the residual case ever surfaces in practice.
Pre-existing FIX-counter bug (separate issue, not fixed in this release). Three sibling regexes at parser.py:673,682,695 count FIX'd parameters via \$THETA[^;$]*?FIX-style patterns. They have a different structural bug (count at most 1 FIX per block, not per parameter) and would need a per-line state-machine rewrite rather than a regex swap. Filed for a future release.
Verified with 16 inline checks: bug case fires (4 H1 + 2 M2), regression-clean inputs unchanged (5), end-to-end parse_lst against run19.lst returns identical results to v2.9.15 (5).

v2.9.15 — Input-validation hardening (safety, no behavior change)

Three small safety improvements found by a senior-developer code review of the dialogs and run-tracking layer.

Duplicate dialog now validates the new filename with the same regex as the New Model dialog (alphanumerics + underscore + hyphen, must start with a letter, optional .mod/.ctl extension). Previously, a duplicate name like ../foo was accepted and would have written the file outside the project directory. The new validation refuses path separators, .. segments, spaces, and other shell-quoting hazards before the dialog accepts.
Run history IDs now include a 4-hex random suffix (e.g. run104_1715357893_a3f7) so two runs launched in the same wall-clock second on the same source model no longer share an ID. Previously the second run's status update would silently clobber the first run's record on completion.
Duplicate flow no longer reports a "meta save failed" error as if the duplication itself failed. The duplicated .mod is now considered the critical-path output: if the file is written successfully but the meta-save step then fails (rare — disk full, permission change), the user sees the file in the model list and a status-bar warning, instead of a blocking error dialog that incorrectly suggests the duplication didn't happen.

v2.9.14 — Description field in dialogs and automatic $TABLE rename on duplicate

Description field in the New Model dialog: optional one-line field that becomes the model's description (the same field shown in the Models tab "Description" column and editable later via the Annotation panel). Backed by meta['comment'] — no schema change.
Description field in the Duplicate dialog: pre-filled verbatim from the source model's description, with auto-select on focus so a single keystroke replaces the inherited text. Pharmacometricians typically duplicate to test a variant of a base model; pre-filling gives the user the closest-correct answer (and they edit when they need to differentiate) instead of leaving 80% of duplicates with empty descriptions.
Automatic $TABLE / $MSFO rename on duplicate: when duplicating run104.mod → run105.mod, NMGUI2 now rewrites output filenames inside the duplicated control stream so the new run does not silently overwrite the source's output tables and MSF files. Visible checkbox in the Duplicate dialog (default checked) for rare cases where downstream R/PsN scripts hard-code the original filenames.
Two-anchor rewrite algorithm covers both PsN convention and NMGUI's own template:
- Strategy A — stem substring: run104.tab → run105.tab, path/run104/sdtab104 → path/run105/sdtab105, mymodel104_pk.tab → mymodel105_pk.tab.
- Strategy B — trailing-runno digits (PsN): sdtab104 → sdtab105, patab104.tab → patab105.tab, msfb104 → msfb105.
- Strategy C — neither anchor matches: filename is left unchanged and the user is shown a Yes/Cancel modal listing each unanchored filename so they can choose to proceed or edit the .mod manually.
Strict input preservation: $MSFI (input MSF for continuation runs), $DATA, $INCLUDE, $SUBR are NEVER rewritten. The shared-MSF workflow (base → covariate search) continues to work correctly.
Path prefixes, file-name case (SDTAB104 stays uppercase on Linux), quoting (FILE='sdtab104'), and inline ; comments are preserved.
Status bar reports what was renamed: Created run105.mod · renamed sdtab104→sdtab105, patab104→patab105.

v2.9.13 — Parser fix: THETA value/label misalignment when N(thetas) > 12

THETA values shifted by 3 in the HTML output, Models tab parameter table, and QC report whenever a model had more than 12 thetas. Root cause was in _parse_values() (parser.py): NONMEM wraps the TH 1 ... TH12 header onto a continuation line TH13 TH14 TH15 with inconsistent spacing (no space between prefix and digit on the wrap line). The skip-regex ^(TH|ET|EP|EPS|SE)\s required a space, missed the wrap line, and the digit-extraction fallback then treated 13, 14, 15 as estimate values — prepending three phantom values and shifting everything else by 3.
Correlation matrix labels truncated to 12 in any covariance step that produced more than 12 estimated parameters (i.e. nearly all real models). Same root cause in the correlation-matrix parser — ^(TH|OM|SG)\s matched only the first header row and ignored every wrap continuation, so result['cor_labels'] held 12 entries while result['correlation_matrix'] held the full 50+ rows.
Defensive hardening of _parse_matrix_diag and _parse_matrix_full: both now explicitly skip wrap-continuation lines beginning with TH13, OM11, SG12, etc., rather than relying incidentally on the "no decimal point" guard.
Removed dead code in widgets/lst_viewer.py (parse_theta_from_block, parse_omega_sigma_from_block) that had the same wrap bug and was never called.
Verified end-to-end against a real run with 15 thetas and a 54-parameter covariance matrix: parser now returns 15 correctly aligned theta values and 54 matching correlation labels (was 12).

v2.9.12 — VPC types: categorical, TTE, and censored support

New: categorical, TTE, and censored VPCs, in addition to the existing continuous VPC. Type is auto-detected from the PsN output folder (meta.yaml / command.txt) on folder load and on every Generate press.
Categorical VPC: vpc backend uses vpc_cat() (levels inferred from observed DV; stratification not supported, tooltip explains why); xpose backend uses vpc_data(vpc_type="categorical") and supports stratification.
TTE (time-to-event) VPC: detected when PsN was run with -tte=. All quantile/LLOQ/pred-corr controls are greyed out — they don't apply to KM plots. Pre-run check warns and blocks if m1/simtab* files are missing. vpc backend uses vpc_tte(rtte=FALSE); xpose backend uses vpc_data(vpc_type="time-to-event"). The vpc_results.csv absence warning is suppressed for TTE.
Censored VPC: detected when PsN was run with -censor=VAR. The LLOQ field is force-enabled even when "Use PsN settings" is checked, so the user can supply it if PsN didn't record one. Pre-run check blocks with a clear message if LLOQ is still missing. vpc backend uses vpc_cens(lloq=, uloq=); xpose backend uses vpc_data(vpc_type="censored", opt=vpc_opt(lloq=)).
The continuous VPC code path is unchanged — the new routing exits early before reaching it.

v2.9.11 — Startup warning and TypeError fixes

TypeError: invalid argument to sipBadCatcherResult(): The QPainter.drawText(QRect, flags, str) overload is broken in some PyQt6 builds (sip binding bug). The NM logo painter now uses QFontMetrics.horizontalAdvance() to manually centre the text and calls drawText(x, y, str) instead, which has no ambiguous overload.
qt.qpa.fonts: … missing font family "Sans-serif": Qt stylesheets do not understand CSS generic family names like sans-serif — Qt treats them as literal font names and spends ~80 ms searching before giving up. Removed the generic fallback from the font stack; Qt applies its own system fallback automatically.

v2.9.10 — Custom application icon

Custom app icon: The NM monogram (blue rounded rectangle, white bold "NM") now appears in the macOS Dock and Windows taskbar instead of the generic Python icon. Built at eight sizes (16–256 px) from the existing QPainter logo, so it is sharp at every resolution including Retina/HiDPI displays. No image files required.
Windows taskbar grouping: On Windows, SetCurrentProcessExplicitAppUserModelID is called before QApplication starts, which tells the taskbar to treat NMGUI as its own application entry rather than grouping it under pythonw.exe.

v2.9.9 — Individual fit plots: legend and observation count

Improvements

Shared legend bar: A 22 px legend bar now appears below the toolbar on the individual fit plot, identifying the three series: blue filled circles (DV — Observed), solid line (IPRED — Individual prediction), and red dashed line (PRED — Population prediction). The legend carries a tooltip explaining the diagnostic meaning of each comparison, including the ETA shrinkage warning (DV ≈ PRED ≠ IPRED).
Observation count in subplot titles: Each subplot now shows ID 101 (n=12) — the number of informative observations plotted — so subjects with very sparse data (n=2–3) are immediately obvious without counting dots.
The legend repaints correctly when switching between dark and light themes.

v2.9.8 — iOFV waterfall: normalised mode and chi-squared threshold

New features

Normalised iOFV view: The waterfall plot now offers an Absolute / Normalised toggle. Normalised mode divides each subject's iOFV by their number of informative observations (EVID=0, MDV=0), making subjects with very rich or very sparse sampling directly comparable.
Chi-squared outlier threshold: The previous mean+2 SD horizontal line has been replaced by a statistically sound χ²(df=n_obs, α=0.001) threshold. In Absolute mode this renders as a per-subject step function; in Normalised mode as a smooth decreasing curve (because variance of χ²(k)/k = 2/k, so sparse subjects have a higher per-unit threshold).
Amber highlighting for sparse subjects: Subjects with fewer than 5 observations are highlighted in amber — the chi-squared approximation is unreliable at that sample size.
Info bar: A one-line explanation beneath the toolbar reads "iOFV scales with the number of observations per subject · Absolute: raw iOFV with subject-specific χ² threshold · Normalised: iOFV / n_obs corrects for unequal sampling richness", so the reason two options exist is always visible.
Rich tooltips: Each toggle button carries a detailed tooltip explaining when to use it and how the threshold is computed.
Colour legend: A legend bar at the bottom of the plot identifies normal (blue–purple gradient), outlier (red), sparse (amber), and no-data (grey) bars.

How it works

Observation counts are read from the dataset file specified in $DATA in the background using a QThread worker. Rows with EVID≠0 or MDV=1 are excluded.
If scipy is not available the plot falls back to a mean+2 SD threshold with a visible warning.

v2.9.7 — Code quality and stability audit

Reliability fixes

Run popup no longer freezes the UI on launch: NONMEM and PsN version detection (_detect_nonmem_version, _detect_psn_version) was called synchronously on the main thread, blocking the GUI for up to 10 s when PsN was not on $PATH. Detection now runs in a background thread and patches the run record after the popup opens.
No more crash after closing the main window while a background check is in progress: The R-availability check and version-update check post their results back to the UI via QTimer.singleShot. If the window is already closed when the callback fires, the call previously raised RuntimeError: wrapped C++ object deleted. Both callbacks are now guarded with try/except RuntimeError.
VPC Stop button now actually kills the Rscript process: Clicking Stop terminated the Python QThread wrapper but left the Rscript subprocess running (it was started with start_new_session=True in its own process group). A new stop_subprocess() method on VPCWorker explicitly kills the child process. The same fix is applied to export workers (PNG/PDF). An export in progress now also correctly blocks a new VPC run from starting.
Stale scan-error callbacks no longer accumulate: When a directory scan was cancelled and immediately replaced by a new scan, the error signal of the old worker was not disconnected. Each rapid rescan leaked one error handler that could fire misleading status messages. The error signal is now disconnected alongside result.
Parameter CSV export status message no longer silently fails: ParameterTable used a fragile parent().parent() widget-tree walk to emit a status-bar message after CSV export. It now emits its own export_done signal, connected at the callsite in models.py.

Code quality

Bootstrap / SIR sample, resample, and thread spinboxes in the Uncertainty tab changed from QDoubleSpinBox(decimals=0) to QSpinBox — avoids showing a spurious decimal point on some platforms and removes unnecessary int() casts in the command builder.
Per-model parse errors in ScanWorker are now logged at WARNING level instead of being silently swallowed, so problems with individual .mod files appear in diagnostic output.

v2.9.6 — VPC coverage hardening

Bug fix (critical)

xpose VPC failed with "Model file run3.lst not found" when the model that generated the VPC had been renamed or moved. The xpose backend now auto-detects the correct .lst file from command.txt in the vpc folder (xpose_data(file=…) instead of xpose_data(runno=…, dir=…)). A new Model .lst field in the xpose settings allows manual override when auto-detection fails.

VPC correctness fixes (previously silent failures)

pcVPCs now plot correctly when PsN was run with -predcorr: pred_corr=TRUE is auto-injected into vpc::vpc() and xpose::vpc_opt() from meta.yaml.
Stratified VPCs now work with "Use PsN settings": stratify_on from meta.yaml is forwarded as stratify= (vpc package) or stratify_on= (xpose).
Categorical and TTE VPCs: NMGUI2 now detects -categorical and -tte in meta.yaml and shows a clear error instead of silently producing a wrong continuous-style plot.
vpc package v1.x compatibility: the vpc() return value is unwrapped from $plot when needed, preventing ggsave() failures on vpc ≥ 1.0.
Non-default DV column (-dv=LNDV etc.) now forwarded to obs_cols/sim_cols.
Log-scale VPCs: -lnDV=1 is detected from meta.yaml and the Log Y axis is auto-enabled.

Robustness

Multiple simulation directories (m2.zip, m3.zip, …) from PsN -n_simulation_models > 1 are now all extracted before the R script runs.
command.txt fallback now handles both = and space-separated PsN arguments (-idv TAD).
meta.yaml parser fixed for values containing colons (e.g., stratify_on: DOSE:BIN).
xpose pred_corr injection is version-gated for xpose < 0.4 compatibility.
LLOQ method selector (M1/M2/M3) added to VPC settings.

v2.9.5 — VPC tab: non-default IDV support and regex fix

Two bug fixes for the VPC tab. Both surfaced when running psn vpc -idv=TAD (or any non-TIME IDV).

Bug fixes

xpose VPC failed at the parse step with '\.' is an unrecognized escape in character string. A regex pattern in the xpose error-diagnostic block (added in v2.9.4) used Python f-string escaping that rendered as \. in R source. R 4.x rejects this. The two patterns now use glob2rx("*.lst") and glob2rx("*.mod"), letting R construct the regex itself.
vpc and xpose backends failed with "no idv column" when PsN was run with -idv=TAD. Neither vpc::vpc(psn_folder=...) nor xpose::vpc_data(psn_folder=...) auto-detect the IDV column from PsN output, so non-default IDVs caused the run to abort with No column for indepentent variable indicator found. NMGUI2 now parses PsN's meta.yaml (preferred) or command.txt (fallback) to recover the IDV, DV, predcorr flag, samples, stratification column, and LLOQ that PsN actually used. The IDV is forwarded as obs_cols=list(idv="…") / sim_cols=list(idv="…") to vpc::vpc() and as idv=… inside vpc_opt() for xpose. The detected options are logged to the run console so users can see exactly what was recovered.

New functionality

IDV override field — a new IDV field in the VPC settings panel (Row 5: IDV / LLOQ / ULOQ). Leave blank to use auto-detection; type a column name to override. Always available, regardless of "Use PsN settings", because the IDV column-name issue is independent of which numerical settings are used.

Internal

New _parse_psn_meta(vpc_folder) helper reads meta.yaml's tool_options: block and command.txt flags without adding a YAML dependency.
New _resolve_idv(vpc_folder) centralises the resolution order: manual override → meta.yaml → command.txt → package default.

v2.9.4 — VPC tab: hardening and new options

Seven targeted fixes and improvements to the VPC tab, covering silent failures, missing UI controls, and R script robustness.

Bug fixes

Stale m1/ after a PsN re-run — previously, if PsN's vpc was re-run with different settings and produced a fresh m1.zip, the app kept using the old extracted m1/ folder. The extractor now compares the modification time of m1.zip against the newest file in m1/ and re-extracts automatically when the zip is newer.
vpc_results.csv absence not detected — PsN run with -no_results=1 or an aborted VPC produces no vpc_results.csv, causing the R backend to fail with a confusing message. The app now detects this before running and prints a clear warning to the console.
xpose4 confused with xpose — users who have the legacy xpose4 package installed (common in older NONMEM groups) but not the newer xpose package got a cryptic R error. The R-package check now explicitly detects this and shows: "xpose ✗ (xpose4 found — run install.packages('xpose'))".
R locale mismatch — on systems with a non-English LC_NUMERIC locale (e.g. Dutch/German), read.csv would misparse decimal separators in vpc_results.csv. Both generated R scripts now begin with Sys.setlocale("LC_NUMERIC", "C").
xpose error messages lacked run-file diagnostics — the xpose error handler now also lists .lst and .mod files in the run directory alongside sdtab files, so the cause of a "run not found" failure is immediately visible in the console.

New functionality

ULOQ field — an upper limit of quantification field is now available next to LLOQ in the settings panel. Passed as uloq= to both vpc::vpc() and xpose::vpc_opt(). Disabled when "Use PsN settings" is checked.
Configurable timeout — the hard-coded 5-minute R script timeout is replaced by a user-configurable Timeout field (1–120 min, default 30 min) in the settings panel. VPCs with 1000+ simulations routinely need 10–30 minutes.

UX

Stratify field now has a tooltip warning that combining it with PsN's -stratify_on option produces incorrect results.

v2.9.3 — VPC tab: auto-extract m1.zip

A focused bug-fix release for the VPC tab.

Bug fix

VPC failed silently when PsN had zipped the m1/ folder. PsN's vpc command produces m1.zip (instead of an unpacked m1/) when run with -clean=2 or higher. Both VPC backends used by NMGUI2 — PsN's bundled vpc.R and the xpose-based plotter — read simulation tables from m1/ directly, so they would either fail with "no simulation tables found" or silently produce an empty plot. The VPC tab now detects m1.zip on Run, extracts it in place if m1/ is missing or empty, and logs the action to the run console. Existing unpacked m1/ folders are left alone (no re-extract). Corrupt zips, permission errors, and unexpected layouts are reported with a clear error instead of failing silently. The folder hint in the VPC settings panel was updated to reflect that either m1/ or m1.zip is acceptable.

v2.9.2 — Files tab improvements

Focused on the Files tab: new file-type handling, usability fixes, and a conversion utility.

New functionality

Convert .tab to CSV — right-click any .tab file in the file browser and choose Convert to CSV. A new .csv file is created in the same folder using comma-separated format. If the target file already exists a warning is shown and nothing is overwritten. The file list reloads automatically and the new CSV is selected.
NONMEM output files open as tables — .ext, .cov, .cor, .phi, .cnv, .coi are now parsed by the NONMEM TABLE reader and open in the Table / Plot viewer, exactly like .tab files. Filter-bar pills for .cnv and .coi added.
Multi-section CSV files display correctly — PsN summary reports such as sir_results.csv (multiple sections with repeated headers) are now merged into a single flat table with section-label rows (e.g. ── Quantiles (R type=2) ──) instead of showing a broken single-column table.
Binary files show a friendly placeholder — clicking a .pdf, .png, .docx, .zip or similar file now shows "No preview available — double-click to open with the system application" instead of garbled content.

Usability fixes

Resizable file-browser columns — all three columns (Name, Size, Modified) are now freely draggable. Default widths: Name 220 px, Size 72 px, Modified 130 px.
Duplicate Table/Plot pills removed — when a table file was open in Plot mode, two identical Table/Plot pill rows appeared. The inner row (from DataExplorerWidget) is now hidden when the widget is embedded in the Files tab.
Button text clipping on Linux — setFixedWidth on the Remove selected and Save settings buttons in the Settings tab, and Discard in the Files toolbar, replaced with setMinimumWidth so text is never clipped on Linux.

v2.9.1 — bug fixes and performance

A bug-fix release from a multi-lens code audit. No new features, no UI changes. Every fix was independently reviewed before merge.

Bug fixes

Editor and Notes panels stayed dark after theme switch. The Editor and Notes panels in the Models tab now correctly update their palette colours when the theme is toggled (previously they remained at the original theme's background).
Open run popups stayed at the old theme. RunPopup and WatchLogPopup are tracked on a popup-list and re-themed on theme change. Previously, popups opened in dark mode stayed dark after switching to light.
Run-output console font on Linux. The console now uses the existing monospace_font() fallback chain (SF NS Mono → Consolas → DejaVu Sans Mono → …) instead of hardcoding "Menlo" / "Consolas". Linux gets a defined font instead of Qt's silent system fallback.
Stop-button race during run teardown. RunWorker._send_signal() now also checks isRunning() alongside the existing _proc None-check, hardening the stop path against teardown races.
Dead code in dataset parser. A no-op nested function inside the dataset-row parsing loop was discarded (it created an unused closure on every row).

Performance

Models tab — faster bulk fill. setUpdatesEnabled(False) now wraps both the scan and filter setItem loops, eliminating per-cell repaints. Visible flicker on directories with 30+ models is gone, especially noticeable over X11/MobaXterm. The wrapper sits inside a try/finally so updates always re-enable, even if a row raises mid-fill.
Models scan — fewer file reads. ScanWorker now caches check_dataset() results within a single scan, so a directory of 10 models that all reference the same dataset reads it once instead of ten times. Regex patterns used per-model ($PROBLEM, $DATA, Based on:) are compiled at module level instead of inline.
Detached runs — _boot_time() cached. On Linux, the system boot time was being read from /proc/stat on every is_alive() check (≈1,200 redundant reads per 2-hour bootstrap with 5 active runs). Cached on first call now.
set_reference() no longer triggers a full model reset. Only the affected columns (Name, dOFV) emit dataChanged.
Nav-icon cache. Sidebar nav icons are now cached per (name, size, color); theme toggling no longer redraws every icon from scratch.

v2.9.0

New Model — create a blank NONMEM control file directly from the Models tab. Click New model… in the filter row (or press Ctrl+N / ⌘N) to open a dialog that lets you choose a model name, pick a structural-model template (Blank, 1-CMT oral, 1-CMT IV, 2-CMT oral, 2-CMT IV), set the $DATA path with a file picker, and preview the rendered file before writing it. After creation the directory is rescanned and the new model is auto-selected with the Editor tab open.

v2.8.0

Files tab now supports plotting — selecting a .tab or .csv file shows a [ Table | Plot ] pill pair in the toolbar. Plot view is the full DataExplorer (scatter, LOESS, colour-by column, filters) that previously lived in the Evaluation tab. Column deduplication (from v2.7.9) now also applies in the Files tab viewer.
Evaluation tab simplified — "Data Explorer" pill removed; the tab is now a clean 4-pill diagnostics panel: GOF | Individual Fits | OFV Waterfall | Convergence.
Auto-select sdtab in Files tab — navigating to the Files tab with a model selected automatically highlights the model's sdtab in the file list.

v2.7.9

Sim Plot — fix crash on duplicate column names — NONMEM sdtab files that list the same column twice (e.g. IPRED appearing as both column 3 and 18) caused pd.to_numeric to raise arg must be a list, tuple, 1-d array, or Series because selecting a duplicate column returns a DataFrame instead of a Series. Duplicates are now renamed at load time (IPRED → IPRED, IPRED_2) and a warning is shown in the Data card label.

v2.7.8

Sim Plot — PI band row redesign — each band is now a single-line card (rounded, bordered) instead of a two-line layout; Alpha moved onto the same row as Lo/Hi; colour picker is now a circular 22×22 swatch; remove button is a round ghost pill that turns red on hover; "Add band" and "Add filter" buttons use a dashed-border style to communicate additive intent; band card spacing increased for breathing room; preset dropdown gains a tooltip warning it resets all bands. Changes are fully cross-platform (macOS/Windows/Linux).

v2.7.7

Eliminated -apple-system startup warning — Qt cannot resolve CSS-level font aliases like -apple-system; replaced with "Helvetica Neue", "Segoe UI", Arial, sans-serif which Qt can find on all platforms without the 71 ms alias-population penalty.

v2.7.6

SF Mono actually renders now — Qt registers SFNSMono.ttf under the internal family name .SF NS Mono (not SF Mono). The font priority list now uses the correct name, so macOS editors show SF Mono instead of Courier.
Removed system-ui from base stylesheet — Qt does not understand the CSS system-ui keyword and emitted a 74 ms warning on every startup. Replaced with -apple-system, "Segoe UI", "Helvetica Neue", Arial, sans-serif.

v2.7.5

SF Mono now loads on macOS — QFontDatabase.addApplicationFont() explicitly registers /System/Library/Fonts/SFNSMono.ttf (and the italic variant) at startup. Qt's normal font enumeration cannot see SF Mono; this one-time registration makes the SF Mono name resolvable so editors render in SF Mono instead of falling back to Menlo.

v2.7.4

Virtualised table viewer — CSV and NONMEM TABLE files in the Files tab are now backed by QAbstractTableModel + QTableView instead of QTableWidget. Qt only fetches data for visible rows, so files of any size (100k+ rows) open instantly and scroll smoothly. The 5 000-row cap and the "Showing first N rows" notice have been removed. Column sorting tries numeric comparison first, falling back to case-insensitive string sort.

v2.7.3

Font priority reordered — SF Mono is now first in the editor font chain: SF Mono → Consolas → DejaVu Sans Mono → JetBrains Mono → Cascadia Code → Liberation Mono → Menlo → Ubuntu Mono → Courier New. macOS users will see SF Mono by default.

v2.7.2

File list panel wider — Files tab middle pane default width increased from 240 → 320 px so filenames are no longer truncated on first open.
SF Mono added to the monospace font priority list (macOS 10.12+, the font used by Xcode and Terminal). On a standard Mac without JetBrains Mono installed the editors now use SF Mono instead of Menlo.

v2.7.1

Consistent monospace font — both the Models tab editor and the Files tab viewer now use a shared monospace_font() helper that calls QFont.setFamilies() with a priority list: JetBrains Mono → SF Mono → Cascadia Code → Cascadia Mono → Consolas → DejaVu Sans Mono → Liberation Mono → Menlo → Ubuntu Mono → Courier New. StyleHint.TypeWriter ensures a monospace fallback on minimal installs. Both editors render at 11 pt.

v2.7.0

New Files tab and Models tab right-click additions.

Files tab (Ctrl+2) — three-pane file browser: extension filter checkboxes (.mod, .ctl, .lst, .tab, .csv, .ext, .cov, .cor, .phi pre-populated; custom extensions addable), sortable file list (name / size / modified), and content viewer with syntax highlighting for .mod/.ctl, inline find, and spreadsheet view for .csv/.tab with NONMEM TABLE NO. header handling. Row cap of 5 000 with notice. Edit/Save/Discard for text and CSV files.
Run added to the top of the Models right-click menu — runs the selected model with current panel settings.
Open folder added to the right-click menu — opens the model's directory in Finder / Explorer / xdg-open.
View .ext added to the right-click menu (shown when a run exists) — opens the parameter-estimates-by-iteration file.
Delete… added to the bottom of the right-click menu — removes .mod and all paired output files; parses $DATA to identify and exclude the dataset. Confirmation dialog lists every file before deletion.
Navigation shortcuts shifted: Ctrl+2 = Files, Ctrl+3 = Tree, Ctrl+4 = Evaluation, Ctrl+5 = VPC, Ctrl+6 = Uncertainty, Ctrl+7 = Sim Plot, Ctrl+8 = History, Ctrl+9 = Settings.

v2.6.13

Run detached defaults to OFF — the "Run detached" checkbox now always starts unchecked, regardless of whether an SSH session is detected. Detached mode must be opted in to explicitly.

v2.6.12

Sim Plot default panel width increased to 600 px.

Splitter default — 380 → 600 px so the panel opens at the width where all spinbox values display fully without the user needing to drag the splitter
Panel max width — 500 → 700 px to allow further widening if desired

v2.6.11

Sim Plot spinboxes — adaptive width instead of fixed pixels.

Stop guessing font metrics — fixed-width spinboxes (64 px) still clipped values because macOS font size varies with display scaling; switched Lo%, Hi% and Alpha spinboxes to setMinimumWidth(62) + setMaximumWidth(110/90) with addWidget(stretch=1) so they expand to fill whatever space the panel provides
Result: spinboxes now use the full available panel width (typically 80–110 px each), guaranteeing values like "95.0" and "0.25" always display completely regardless of system font size or display scaling

v2.6.10

Sim Plot band row — fix checkbox expansion and spinbox clipping.

Checkbox fixed width — setFixedWidth(20) added to the visibility checkbox; without it, QCheckBox has Preferred size policy and silently absorbs all spare horizontal space, pushing the colour swatch and × button off-screen
row1.addStretch() — explicit stretch added at the end of the band row so any remaining panel width goes to blank space on the right, not to widgets
Spinboxes widened 58 → 64 px — on macOS the arrow buttons consume ~28 px, leaving 30 px text area at 58 px; "95.0" (4 chars) needs ~32 px and clipped to "9". At 64 px the text area is 36 px, enough for all expected values
Alpha row spacer uses literal 20 px — previously used sizeHint() which would not match the now-fixed checkbox width

v2.6.9

Sim Plot band row redesign — two-line layout with inline labels.

Root cause fixed — macOS QDoubleSpinBox has a minimum intrinsic rendering width (~58 px); forcing 46 px clipped the value to nothing. Pixel-squeezing six controls onto one line was the wrong architecture
Two-line band rows — each band now uses two compact lines: line 1 holds the visibility checkbox, Lo%, Hi%, colour swatch and remove button; line 2 holds Alpha. Inline labels ("Lo%", "Hi%", "Alpha") replace the old separate column-header row
Spinboxes now 58 px — wide enough for macOS to render "49.9" and "0.25" without clipping
Column-header row removed — no longer needed; inline labels provide the same context without alignment fragility
Colour swatch restored to 32 px — previous shrink to 26 px hit macOS button minimums

v2.6.8

Sim Plot band row layout overhaul — all controls now fully visible.

Removed % suffix from Lo/Hi spinboxes — the column header "Lo%" / "Hi%" already communicates the unit; dropping the suffix saves ~12 px per spinbox and lets the values render fully
Narrowed Lo/Hi spinboxes — 60 → 46 px (fits "49.9" at 1 decimal place)
Narrowed colour swatch — 32 → 26 px (still clearly a colour picker)
Narrowed alpha spinbox — 58 → 52 px
Narrowed remove button — 22 → 20 px
Tightened row spacing — 4 → 3 px between widgets
New band row total: 221 px (was 270 px) — fits comfortably in the available content width with ~40 px to spare, even accounting for the vertical scrollbar
Filter header fixed — Op and Value column labels now have explicit fixed widths (52 / 72 px) matching the data row, so "Op" is never clipped

v2.6.7

Sim Plot panel default width correction.

Wider default panel — splitter default increased from 300 → 380 px and max-width from 400 → 500 px; the vertical scrollbar was consuming ~15 px of horizontal space, causing the band row columns to clip even at 300 px; 380 px gives ~333 px usable content width, well clear of the 270 px band row

v2.6.6

Sim Plot left panel layout fixes.

Narrower band rows — Lo%/Hi% spinboxes reduced from 68 → 60 px, Alpha from 72 → 58 px so all columns (including the × button) fit without clipping
Narrower panel — default splitter position reduced from 360 → 300 px; max panel width reduced from 480 → 400 px
Span spinbox — widened from 62 → 70 px so "0.30" is never clipped

v2.6.5

Sim Plot LOESS smoothing option.

Smooth curves — new checkbox in the Appearance card applies LOESS smoothing to all PI ribbon boundaries and the median line before rendering; a Span spinbox (0.05–1.0, default 0.30) controls bandwidth — lower follows data closely, higher gives a smoother curve

v2.6.4

Sim Plot MDV filter accessibility fix.

MDV filter moved to Filters card — the "Exclude MDV=1 rows" checkbox is now in the Filters card (expanded by default) instead of the collapsed Appearance card, making it immediately visible and easy to toggle

v2.6.3

Sim Plot usability and layout fixes.

Workflow hint — a note in the Data card explains that the Sim Plot tab requires a NONMEM simulation output file with a replicate column (REP, IREP, SIM, SIMNO…), not a standard estimation sdtab
Replicate column validation — a status bar warning is shown when no recognised replicate column is found in the loaded file, preventing meaningless plots from being silently generated
Median colour default — now initialises to the current theme foreground colour instead of white (which was invisible on light backgrounds)
PI band header alignment — column labels (Vis / Lo% / Hi% / Colour / Alpha) now use exact fixed widths matching the data row widgets so they align correctly
Alpha spinbox — widened from 58 px to 72 px; values such as "0.25" no longer clip to "0"
Y-axis label overflow — replaced tight_layout with explicit subplots_adjust so the Y-axis label no longer overflows beyond the right edge of the canvas

v2.6.2

Performance and audit fixes.

Threaded table loading — the Evaluation tab now parses sdtab files in a background QThread; the Load button shows "Loading…" while parsing and the UI stays fully responsive on large files or slow network drives
IndFit row bounds guard — ragged rows (fewer columns than the header) in the Individual Fits widget no longer raise IndexError; they are silently skipped
GOF replot error logging — the silent except Exception: pass in the GOF 2×2 replot is replaced with a debug-level log entry so errors are visible with --debug without crashing

v2.6.1

Bug fixes for the Sim Plot tab and the Models tab METHOD column.

SIM method label — simulation-only runs ($SIM without $EST) now display SIM in the METHOD column of the Models tab instead of the incorrect default FO
Sim Plot layout — PI band rows redesigned: label-free compact layout with a column header row, % suffix on spinboxes, wider panel (360 px default), compact Browse/Load buttons; filter rows similarly updated

v2.6.0

New Simulation Plot tab for Monte Carlo prediction interval visualisation.

Sim Plot tab — a dedicated tab (Ctrl+6) for plotting prediction intervals from NONMEM Monte Carlo simulation output. Load any NONMEM table file or CSV (no row-count cap), configure PI bands, and generate publication-ready plots in seconds
Unlimited file loading — the table parser now accepts max_rows=None so large simulation files (50k–500k rows typical for 1 000 replicates × 500 time points) load fully without truncation
Replicate auto-detection — the tab recognises explicit replicate columns (REP, IREP, SIM, SIMNO, REP_NO, SIM_NUM, etc.) and also detects ID-cycling automatically when no explicit column is present
Configurable PI bands — up to 4 simultaneous prediction interval ribbons, each with its own percentile pair (Lo/Hi %), colour (colour picker), alpha and visibility toggle. Four presets cover the most common pharmacometric conventions (5/95 + 25/75, 2.5/97.5 + 10/90, etc.)
Multiple filters — up to 6 independent column filters (==, !=, >, <, >=, <=), ANDed together, to subset by compartment, dose group, sex, or any other column before plotting
Median line — configurable colour and line width; computed as the 50th percentile across replicates
Log/linear Y-axis — one checkbox toggles between linear and logarithmic scale, essential for PK concentration-time plots
MDV=1 exclusion — dosing-only rows filtered out by default before quantile computation
Observed data overlay — optional second file load; observed DV points are overlaid as a semi-transparent scatter on the simulated ribbons, with independent X/Y column selectors
Background computation — all quantile calculations run in a QThread so the UI stays responsive on large datasets
Save PNG — 300 DPI export via matplotlib savefig
Model context — selecting a model in the Models tab sets the browse directory for the Sim Plot file dialog, consistent with Evaluation and VPC tabs

v2.5.8

Bug fix for the ETA vs Covariate plot.

ET\d+ columns now recognised as ETAs — NONMEM truncates ETA(12) to ET12 in TABLE output. The ETA dropdown previously required the full ETA prefix, so ET12, ET13, ET14 etc. were silently placed in the covariates list and the dropdown stayed empty. Fixed by also accepting ET\d+ and PHI\d+ column names as ETAs
Residual columns excluded from covariates — NPDE, IWRES and WRES added to the covariate skip-list (alongside the existing CWRES, PRED, IPRED etc.) so they no longer appear as candidate covariates

v2.5.7

New GOF features.

NPDE distribution plot — a new "NPDE Dist" pill appears in the GOF sub-strip whenever the loaded table file contains an NPDE column. Shows a histogram with a normal density overlay and mean/SD statistics, identical in style to the CWRES Hist panel. When NPDE is absent the button is hidden automatically
PNG export on all matplotlib GOF panels — CWRES Hist, QQ Plot and NPDE Dist each now have a "Save PNG…" button that saves a 300 DPI PNG via matplotlib's savefig; the button is disabled until a plot has been rendered

v2.5.6

Bug fix for the Ron Keizer vpc backend.

Three CI bands now visible — pi_as_area was incorrectly set to TRUE, which collapsed the three separate confidence ribbons (around the 5th, 50th and 95th simulated percentiles) into a single filled slab. Changed to FALSE (the vpc package default) so a proper three-band VPC is displayed
Readr parsing warnings suppressed — NONMEM simulation files embed a TABLE NO. header row every N records; readr emits ~2000 parsing-failure warnings per VPC run (one per simulation replicate). These are harmless but were alarming. The vpc() call is now wrapped in withCallingHandlers to muffle only those specific warnings while keeping real R errors visible

v2.5.5

VPC tab overhaul — fixes implausible plots from the Ron Keizer vpc backend and removes the outdated xpose4 backend.

xpose4 backend removed — xpose4 is no longer maintained and produced implausible VPCs; the modern xpose package is a strict improvement
"Use PsN settings" mode (default on) — when checked, binning, stratification, pred-corr and LLOQ are inherited directly from the PsN output folder for both the vpc and xpose backends. Previously the tab always forced bins="auto" and explicit pred_corr/stratify args that overrode PsN's values, misaligning observations and simulated prediction intervals
Correct bins argument — when overriding manually, bins="jenks" is used instead of the invalid "auto" string
Workflow hint — a one-line note at the top of the VPC settings panel explains that a PsN vpc run must be completed first
Validation gate fix — stratification column validation is now skipped when "Use PsN settings" is checked (the column is inherited from PsN and may not appear literally in the widget)
R status bar now shows only the two current backends: vpc and xpose

v2.5.4

Bug-fix release (Tier 1 audit items + R availability check).

Python 3.9 compatibility — detached_runs.py used int | None, list[dict] and tuple[list, list] type annotations introduced in Python 3.10. Added from __future__ import annotations so the file imports cleanly on Python 3.9 (common on HPC clusters running CentOS/RHEL 8)
VPC column name sanitisation — stratification column names are now cleaned with _r_col() before embedding in the generated R script; control characters and embedded quotes that would break the R string literal are stripped. A separate _sanitize_r improvement also strips newlines and other control characters from path strings
Config directory creation — CONFIG_DIR.mkdir() is now wrapped in try/except OSError; on HPC systems with a read-only home directory the app previously raised PermissionError at import time and refused to start
R availability check at startup — if Rscript is not found on PATH, a status bar message is shown immediately on startup explaining that VPC and RStudio features are unavailable; previously the first indication was an opaque error when actually trying to use those features

v2.5.3

Bug-fix release (continued audit follow-up).

Parameter export status message — the "Parameters exported" status bar message no longer crashes when the widget tree is not in the expected shape; the hasattr(self,'parent') guard was always True and gave false safety
GOF axis column fallback — when a previously-selected X-axis column (e.g. CWRES) is absent from a newly-loaded table, the plot now falls back to the panel default (PRED or TIME) instead of silently selecting column 0 (which was typically ID or TIME at random)
GOF file-open imports — __import__('os') / __import__('subprocess') inline hacks in the "Open exported PNG?" handler replaced with normal top-level imports
History tab run I/O — _load_runs / _save_runs local duplicates removed; history tab now calls the canonical locked versions from config.py, so thread-safety fixes and future bug fixes apply consistently
Unused import removed — APP_VERSION was imported but never used in detached_runs.py
RStudio .Rproj creation — write_text() is now wrapped in a try/except OSError; read-only project directories (common on shared drives) now return a user-visible error instead of raising an unhandled exception

v2.5.2

Bug-fix release addressing findings from a thorough internal audit.

Force-kill fixed — the "Force kill (SIGKILL)" stop option in run popups now works correctly; previously it called a non-existent method and silently did nothing
Boundary warning restored — "Parameter near boundary" warnings are now correctly recorded in run history; a key-name mismatch was preventing this flag from ever being stored
Median statistics corrected — bootstrap and SIR results (bias, RSE, median parameter, dOFV median) now use a proper median calculation; the previous index-based approach produced a systematic upward bias for even-sized samples (7 call sites fixed)
Splitter layout persists — panel splitter sizes are now correctly restored on startup; an early-return in geometry restore was skipping the splitter step
Detached run reconciliation fixed (macOS) — _boot_time() previously returned 0.0 on non-Linux platforms, causing every detached run to be marked finished immediately after restart; it now returns None and the start-time cross-check is skipped when unavailable
Config write thread-safety — all config file writers (save_settings, save_bookmarks, save_runs) now hold the same lock as save_meta, preventing rare corruption from concurrent worker and main-thread writes
Subprocess session handling — RunWorker now uses start_new_session=True (was preexec_fn=os.setsid, which is unsafe with Python threads and can deadlock on macOS); import signal moved to module scope
LRT method-mismatch guard — the workbench comparison table now shows "N/A" with a tooltip instead of a p-value when the reference and candidate models used different estimation methods (e.g. FOCE vs SAEM); exported CSV emits method_mismatch in that column
Waterfall plot at OBJ=0 — subjects with an individual OBJ of exactly 0.0 were excluded from the waterfall plot because the value is falsy; corrected to is not None
ETA–covariate plot filters MDV rows — dosing records (MDV=1) are now excluded from ETA vs covariate scatter plots, consistent with all other diagnostic plots

v2.5.1

Detached runs (SSH/MobaXterm) — a new "Run detached" checkbox in the Run panel launches PsN under nohup in a new session so the job keeps running even if NMGUI2 is closed or the SSH connection drops. Automatically pre-checked when an SSH session is detected. Output is saved to a per-run .nmgui.log file in the project folder. Click the detached row in the Active & Recent Runs table to tail the log in a Watch Log window. On the next NMGUI2 startup, finished detached runs are automatically reconciled (status, timestamps, OFV) without relying on any shutdown hooks. Linux and macOS only.

v2.5.0

Concurrent popup runs — clicking Run now opens a dedicated floating window per model run; multiple models run simultaneously in independent windows, each with its own live console, iteration/OFV progress indicator, elapsed timer, and gentle/force stop controls; on completion the window title and status bar update with the termination result
Active & Recent Runs panel — the Run sub-tab now shows a persistent table of active and recent runs for the current project folder; click a live row to raise its popup window; historical rows are loaded from nmgui_run_records.json in the project directory and survive app restarts; interrupted runs are shown as "? Interrupted"
Theme overhaul — forced Fusion style across all platforms for consistent dark-mode behaviour; migrated ~20 widgets from per-widget stylesheets to object-name + global QSS so theme toggle reliably refreshes every label, separator, spin-box, combo-box and highlighter in the app
QC report — right-click any completed model → QC Report… to open a self-contained HTML report with a PASS / WARN / FAIL checklist covering termination, covariance, condition number, %RSE, parameter correlations, shrinkage, ETABAR and omega boundary checks
Model workbench — Workbench… button opens a sortable table of all completed models with ΔOFV, ΔAIC, ΔBIC, LRT p-value and reference-model selector for quick multi-model comparison
Enhanced compare dialog — the two-model comparison dialog now shows a statistics strip with ΔOFV, ΔAIC, ΔBIC, Δ parameters, LRT p-value and a significance verdict
Dataset integrity checks — at scan time, each model's data file is automatically checked for missing file, column-count mismatches, non-monotonic TIME, duplicate doses, extreme DV values and high BLQ proportion; issues appear in the Info tab's Dataset card
Collapsible layout — the Info panel uses collapsible cards (Dataset / Annotation / Notes); the Parameters tab THETA / OMEGA / SIGMA blocks collapse and expand by clicking the section header
Multi-$EST chain parsing — the .lst Output tab now renders one row per estimation step with per-step OFV, ΔOFV, runtime, significant digits and termination status
Run records — immutable audit trail for every run (UUID, SHA-256 hashes of control stream and dataset, NONMEM/PsN/NMGUI versions, duration, final OFV and status) accessible from the right-click menu
Correctness fixes — LRT p-value now correctly signed when the reference model has fewer parameters; VPC "Save PDF…" reliably produces a PDF; run record and workbench display OFV = 0.0 correctly; GitHub update-check uses numeric version comparison instead of lexicographic

Name		Name	Last commit message	Last commit date
Latest commit History 182 Commits
nmgui2		nmgui2
screenshots		screenshots
.gitignore		.gitignore
CONTRIBUTING.md		CONTRIBUTING.md
LICENSE		LICENSE
README.md		README.md
USER_GUIDE.md		USER_GUIDE.md
nmgui2.py		nmgui2.py
requirements.txt		requirements.txt

Folders and files

Latest commit

History

Repository files navigation

NMGUI2

Table of contents

Features

Models tab

Model detail panel

Files tab (Ctrl+2)

Tree tab (Ctrl+3)

Evaluation tab (Ctrl+4)

VPC tab (Ctrl+5)

Uncertainty tab (Ctrl+6)

Sim Plot tab (Ctrl+7)

History tab (Ctrl+8)

Run Records (audit trail)

Settings tab (Ctrl+9)

Additional features

Dependencies overview

Python packages

External tools for running models

External tools for VPC generation

Quick start

Installation — macOS

1. Install Xcode Command Line Tools

2. Install Homebrew (if not already installed)

3. Install Python 3

4. Install R (required for VPC only)

5. Install RStudio (optional)

6. Install required R packages

7. Install Perl (required by PsN — usually pre-installed on macOS)

8. Install PsN

9. Clone and install NMGUI2

10. Run NMGUI2

Optional: desktop launcher

Installation — Windows

1. Install Python 3.10 or newer

2. Install R (required for VPC only)

3. Install RStudio (optional)

4. Install required R packages

5. Install Perl (required by PsN)

6. Install PsN

7. Install Git

8. Clone and install NMGUI2

9. Run NMGUI2

Optional: desktop launcher

Installation — Linux

1. Install system dependencies

2. Install Qt system libraries (required by PyQt6)

3. Install R (required for VPC only)

4. Install RStudio (optional)

5. Install required R packages

6. Install PsN

7. Clone and install NMGUI2

8. Run NMGUI2

Optional: desktop launcher

Updating NMGUI2

Previous versions

First use

Keyboard shortcuts

Configuration files

Troubleshooting

Contributing

Author

Acknowledgements

License

Changelog

v2.9.25 — Fix: axis change did not update plot (LOESS blocking main thread)

v2.9.24 — Fix: Plot tab blank in Files tab viewer; smarter axis auto-selection

v2.9.23 — Fix: "Convert to CSV" in Files tab truncated multi-subproblem tables

v2.9.22 — Sim Plot performance: 10-30× faster file loading for large simulations

v2.9.21 — Init→Final visualization redesign: log-scale ratio bar

What you see now

FIXED parameters

Boundary proximity

Color thresholds

Edge cases (rare in PMx but handled correctly)

Cross-OS

Sizing