prothon-ensembles 2.1.0__tar.gz

prothon_ensembles-2.1.0/.github/workflows/publish.yml

@@ -0,0 +1,63 @@
+name: Publish to PyPI
+
+on:
+  push:
+    tags:
+      - "v*"
+
+jobs:
+  build:
+    name: Build distributions
+    runs-on: ubuntu-latest
+    steps:
+      - uses: actions/checkout@v5
+        with:
+          fetch-depth: 0          # setuptools-scm reads the tag
+
+      - name: Set up Python
+        uses: actions/setup-python@v6
+        with:
+          python-version: "3.11"
+
+      - name: Install build
+        run: python -m pip install --upgrade pip build
+
+      - name: Build
+        run: python -m build
+
+      - name: Verify the built version matches the release tag
+        run: |
+          TAG="${GITHUB_REF_NAME#v}"
+          VER=$(ls dist/*.tar.gz | sed -E 's/.*prothon_ensembles-(.*)\.tar\.gz/\1/')
+          echo "built: $VER   tag: $TAG"
+          if [ "$TAG" != "$VER" ]; then
+            echo "::error::built version ($VER) does not match the tag ($TAG). A dirty tree or a missing tag will do this."
+            exit 1
+          fi
+
+      - name: Check the metadata PyPI will render
+        run: |
+          python -m pip install --upgrade twine
+          python -m twine check dist/*
+
+      - name: Upload artifacts
+        uses: actions/upload-artifact@v5
+        with:
+          name: dist
+          path: dist/
+
+  publish:
+    name: Publish to PyPI
+    needs: build
+    runs-on: ubuntu-latest
+    environment: pypi
+    permissions:
+      id-token: write           # trusted publishing; no API token is stored
+    steps:
+      - uses: actions/download-artifact@v5
+        with:
+          name: dist
+          path: dist/
+      - uses: pypa/gh-action-pypi-publish@release/v1
+        with:
+          skip-existing: true

prothon_ensembles-2.1.0/.github/workflows/tests.yml

@@ -0,0 +1,41 @@
+name: tests
+
+on:
+  push:
+    branches: [main]
+  pull_request:
+  workflow_dispatch:
+
+jobs:
+  test:
+    runs-on: ${{ matrix.os }}
+    strategy:
+      fail-fast: false
+      matrix:
+        os: [ubuntu-latest, macos-latest, windows-latest]
+        python-version: ["3.9", "3.11", "3.13"]
+    steps:
+      - uses: actions/checkout@v4
+        with:
+          fetch-depth: 0          # setuptools-scm needs the tags
+
+      - uses: actions/setup-python@v5
+        with:
+          python-version: ${{ matrix.python-version }}
+
+      - name: Install
+        run: |
+          python -m pip install --upgrade pip
+          python -m pip install -e ".[dev]"
+
+      - name: Lint
+        if: matrix.os == 'ubuntu-latest' && matrix.python-version == '3.11'
+        run: ruff check src tests
+
+      - name: Test
+        run: pytest --cov=prothon --cov-report=term-missing
+
+      - name: Smoke-test the CLI
+        run: |
+          prothon --version
+          prothon --info

prothon_ensembles-2.1.0/.gitignore

@@ -0,0 +1,22 @@
+# Build artifacts
+build/
+dist/
+*.egg-info/
+src/prothon/_version.py
+
+# Python cache
+__pycache__/
+*.py[cod]
+
+# Test and coverage
+.pytest_cache/
+.coverage
+htmlcov/
+.ruff_cache/
+
+# Prothon output
+*_output/
+
+# Environments
+.venv/
+venv/

prothon_ensembles-2.1.0/CHANGELOG.md

@@ -0,0 +1,173 @@
+# Changelog
+
+All notable changes to Prothon are recorded here. This project follows
+[Semantic Versioning](https://semver.org/spec/v2.0.0.html).
+
+## [2.1.0] — unreleased
+
+A correctness and packaging release. The public API is unchanged: code written
+against 2.0 runs without modification. **Numerical results will differ**, and
+the reasons are set out below — the significance test in 2.0 was not sound.
+
+### Fixed — the significance test
+
+Version 2.0 built its null distribution by drawing two bootstrap resamples from
+the *same* ensemble and measuring the Jensen–Shannon distance between them.
+Two resamples of *n* frames drawn with replacement from the same *n* frames
+share roughly 63% of their points, so they resemble each other far more closely
+than two independent samples of the same size. The null was therefore too tight
+by about a factor of two, and any honest between-ensemble distance cleared it.
+
+Measured on a 400-frame Gaussian ensemble:
+
+| quantity | value |
+|---|---|
+| bootstrap null used by 2.0 | 0.046 |
+| two independent samples, same distribution | 0.097 |
+| observed between-ensemble distance | 0.090 |
+
+The consequence, over 40 null replicates in which both ensembles were drawn
+from an identical distribution:
+
+| | features called different | studies with ≥1 false positive |
+|---|---|---|
+| 2.0 bootstrap null | **100%** | **100%** |
+| 2.1 permutation null | 1.2% | 7.5% |
+
+Version 2.0 reported two independent samples of the same distribution as
+differing significantly at every residue.
+
+The replacement is a permutation test: the frames of both ensembles are pooled
+and relabelled at random into two groups of the original sizes, which gives the
+exact distribution of the statistic under the hypothesis that the ensembles are
+the same. Per-feature values are standardised and pooled before correction, so
+100 relabellings give p-value resolution fine enough to survive a
+false-discovery-rate correction over several hundred residues.
+
+`legacy=True` (CLI: `--legacy-statistics`) reproduces 2.0's behaviour exactly,
+for regenerating published figures. It is documented as unsound.
+
+### Fixed — everything else
+
+- **Per-residue significance.** 2.0 computed one pooled p-value and wrote
+  `local_diss[p_value >= 0.05] = 0.0`. With a scalar `p_value`, NumPy reads
+  that as a mask over the whole array, so a single test decided the fate of
+  every residue at once. Each feature is now tested separately, and the
+  resulting p-values are Benjamini–Hochberg corrected — a 300-residue protein
+  tested at α = 0.05 yields fifteen false positives by construction.
+- **Circular densities for torsions.** `cata` values wrap at ±π. 2.0 estimated
+  them with a Gaussian kernel on a linear grid, which splits a population
+  straddling the wraparound across both ends and puts a false trough between
+  them. Circular measures now use a von Mises kernel with Taylor's plug-in
+  bandwidth on a grid spanning a full turn. Each measure declares whether it is
+  circular, so the call site cannot forget.
+- **`scikit-learn` is now declared as a dependency.** 2.0 imported it for PCA,
+  MDS and t-SNE without listing it, so a clean `pip install` failed on any run
+  that reached dimensionality reduction — which was the CLI default.
+- **Constant features no longer crash the run.** A buried residue with zero
+  SASA in every frame gave `gaussian_kde` a singular covariance matrix and took
+  down the whole study. Such columns now get a degenerate density.
+- **Float32 overflow in the contact sigmoid.** `mdtraj` returns float32
+  distances; `exp()` overflows above ~88 in float32 while the clip allowed 700.
+  The result was right by accident, via `1/(1+inf) → 0`, and noisy with
+  warnings. Distances are promoted to float64 first.
+- **Replotting no longer overwrites saved figures**, and the documented
+  `xlabel`, `ylabel`, `title` and `color` arguments now take effect. In 2.0
+  `replot_global_dissimilarity` accepted them, discarded them, and re-saved
+  over the original file.
+- **Reproducibility.** Resampling drew from the global NumPy state, so two runs
+  of one study gave different p-values with nothing recording why. `Prothon`
+  now takes `random_state`.
+- **Empty atom selections are named.** A coarse-grained model with no C-beta
+  atoms produced an inscrutable NumPy error several frames down the stack.
+
+### Added
+
+- **Noise floor on every result.** The distance between two disjoint halves of
+  a single ensemble is the smallest difference the sampling can resolve. It is
+  reported alongside every comparison, drawn on every dissimilarity figure, and
+  `ComparisonResult.resolved` says plainly whether the measurement clears it.
+- **`manifest.json` per measure**, recording inputs, parameters, seed, Prothon
+  version and full results — so a run can be reproduced and not merely admired.
+- **`ComparisonResult`**, a typed result object that still supports dictionary
+  access (`result["global_dissimilarity"]`) so 2.0 code keeps working. Carries
+  the unmasked per-residue values, per-residue p-values, the significance mask
+  and the noise floor.
+- **`Prothon.summary()`** and a readable default CLI output. `--json` restores
+  the 2.0 behaviour of dumping everything to stdout.
+- `prothon --info`, listing measures and detected backends.
+- Backward-compatibility shim: `from Prothon import Prothon` still works,
+  with a `DeprecationWarning`. Removed in 3.0. Shipped as a single module
+  `src/Prothon.py`, not a package — a directory named `Prothon` beside one
+  named `prothon` is a single path on macOS and Windows.
+- 81 tests, a GitHub Actions matrix across operating systems and Python
+  versions, and `ruff` linting.
+
+### Changed
+
+- **Distribution name is `prothon-ensembles`; the import name and the command
+  are both `prothon`.** PyPI's `prothon` was registered in 2020 by an unrelated
+  protobuf generator and names there are permanent. conda-forge, where the name
+  is free, gets `prothon`. Version 2.0 was never published to either index, so
+  nothing that already works breaks.
+- Releases publish to PyPI from a tag via trusted publishing (OpenID Connect),
+  so no long-lived API token exists to leak.
+- **`src/` layout and `pyproject.toml`**, replacing `setup.py`. Versioning via
+  `setuptools-scm`.
+- **Dimensionality reduction is off by default.** It defaulted to
+  `pca,mds,tsne`; MDS builds a dense frame-by-frame distance matrix, so on a
+  real trajectory the default turned a short comparison into an out-of-memory
+  failure. MDS is now refused above 5,000 frames with a message naming the
+  memory it would need, and a refusal no longer discards the comparison that
+  already succeeded.
+- **Contact numbers are computed once rather than once per atom.** 2.0 rebuilt
+  the pair list in Python for every atom and recomputed the same distances.
+  Identical output to float32 precision:
+
+  | residues | 2.0 | 2.1 | speedup |
+  |---|---|---|---|
+  | 50 | 0.21 s | 0.05 s | 4× |
+  | 100 | 1.37 s | 0.48 s | 3× |
+  | 200 | 10.66 s | 1.10 s | 10× |
+  | 300 | 36.03 s | 2.22 s | 16× |
+
+  Pairs are processed in blocks, so a long trajectory of a large protein no
+  longer needs the whole distance matrix resident.
+- Progress reporting moved from `print` to the `logging` module, so an
+  embedding program can silence or redirect it.
+- Ensembles with fewer than 50 frames now emit a warning that the noise floor
+  understates the true uncertainty.
+
+### Licence
+
+Relicensed from GPL-3.0 to **MIT**, matching the rest of the AAI Research Lab
+tooling and removing an adoption barrier: a number of industrial groups have
+blanket policies against GPL dependencies, and they are a large part of the
+audience for ensemble comparison.
+
+The change is not retroactive. Versions up to and including 2.0.0 were
+distributed under GPL-3.0 and copies obtained under it stay governed by it.
+Nothing in the dependency stack required copyleft — MDTraj is LGPL, and NumPy,
+SciPy, Matplotlib and scikit-learn are BSD.
+
+### Known limitations
+
+- The permutation null assumes frames are exchangeable. Frames from a single
+  continuous MD trajectory are correlated in time, so an ensemble holds fewer
+  independent conformations than it has frames and the p-values remain somewhat
+  optimistic. A block permutation over the correlation time is planned for 3.0.
+  The split-half noise floor is measured rather than assumed and is the more
+  trustworthy guide.
+- Ensembles must share a topology. Comparison across differing sequences —
+  wild type against mutant, ortholog against ortholog — is planned for 3.0.
+
+## [2.0.0] — 2025-04-23
+
+- Restructured the single-module version 1 into a package.
+- Added dimensionality reduction (PCA, MDS, t-SNE), matrix heatmaps, combined
+  local dissimilarity plots and a replotting API.
+
+## [1.0.1] — 2023
+
+- Original release accompanying Aina, Hsueh & Plotkin, *J. Chem. Inf. Model.*
+  **2023**, 63 (11), 3453–3461. Preserved under `legacy/`.

prothon_ensembles-2.1.0/CITATION.cff

@@ -0,0 +1,38 @@
+cff-version: 1.2.0
+message: "If you use Prothon, please cite the method paper below."
+title: "Prothon: efficient comparison of protein conformational ensembles"
+abstract: >-
+  Prothon represents a protein conformational ensemble as a vector of
+  probability distributions over local order parameters and quantifies the
+  dissimilarity between ensembles by the Jensen-Shannon distance between
+  corresponding distributions. The local representation requires no structural
+  superposition, so the cost is linear rather than quadratic in ensemble size.
+type: software
+license: MIT
+repository-code: "https://github.com/aai-research-lab/Prothon"
+authors:
+  - family-names: Aina
+    given-names: Adekunle
+  - family-names: Hsueh
+    given-names: Shawn C. C.
+  - family-names: Plotkin
+    given-names: Steven S.
+preferred-citation:
+  type: article
+  title: >-
+    PROTHON: A Local Order Parameter-Based Method for Efficient Comparison of
+    Protein Ensembles
+  authors:
+    - family-names: Aina
+      given-names: Adekunle
+    - family-names: Hsueh
+      given-names: Shawn C. C.
+    - family-names: Plotkin
+      given-names: Steven S.
+  journal: "Journal of Chemical Information and Modeling"
+  volume: 63
+  issue: 11
+  start: 3453
+  end: 3461
+  year: 2023
+  doi: "10.1021/acs.jcim.3c00145"

prothon_ensembles-2.1.0/CONTRIBUTING.md

@@ -0,0 +1,52 @@
+# Contributing to Prothon
+
+Contributions are welcome — bug reports, new order parameters, better
+estimators, documentation.
+
+## Getting set up
+
+```bash
+git clone https://github.com/aai-research-lab/Prothon.git
+cd Prothon
+pip install -e ".[dev]"
+pytest
+```
+
+## What a change needs
+
+**A test.** Especially for a bug fix: the test should fail before the change
+and pass after it, and be named so that a future regression says which
+behaviour came back.
+
+**A changelog entry**, under `[Unreleased]` in `CHANGELOG.md`. If a change
+alters numerical results, say so plainly and say why the new numbers are the
+right ones.
+
+**Comments that explain why, not what.** The code says what it does. What a
+reader cannot recover is why the bandwidth is chosen that way, or why a null
+distribution is built from permutations rather than bootstrap resamples. That
+reasoning belongs in the source.
+
+## Adding an order parameter
+
+1. Write `compute_<name>(traj)` in `core/representation.py`, returning an
+   `(n_frames, n_features)` array.
+2. Register it in `MEASURES`, declaring `circular` and `per_residue`. Getting
+   `circular` wrong is silent and wrong, not loud and wrong: a linear kernel on
+   circular data produces plausible numbers that understate dissimilarity at
+   the wraparound.
+3. Add it to `_COMPUTE`. A test checks the two stay in step.
+4. Test the shape, the value range, and one case where you know the answer.
+
+## Statistics
+
+Changes to `core/dissimilarity.py` need a calibration check as well as a unit
+test: a null case where both ensembles are drawn from the same distribution,
+asserting that the false-positive rate stays near the nominal level. There is
+one in `tests/test_dissimilarity.py` to copy. This is not optional — the bug
+that motivated the 2.1 release passed every unit test in 2.0 and failed exactly
+this check.
+
+## Style
+
+`ruff check src tests`. Line length 100.

prothon_ensembles-2.1.0/LICENSE

@@ -0,0 +1,21 @@
+MIT License
+
+Copyright (c) 2023-2026 Adekunle Aina
+
+Permission is hereby granted, free of charge, to any person obtaining a copy
+of this software and associated documentation files (the "Software"), to deal
+in the Software without restriction, including without limitation the rights
+to use, copy, modify, merge, publish, distribute, sublicense, and/or sell
+copies of the Software, and to permit persons to whom the Software is
+furnished to do so, subject to the following conditions:
+
+The above copyright notice and this permission notice shall be included in all
+copies or substantial portions of the Software.
+
+THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
+IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
+FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE
+AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
+LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM,
+OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
+SOFTWARE.

prothon_ensembles-2.1.0/PKG-INFO

@@ -0,0 +1,207 @@
+Metadata-Version: 2.4
+Name: prothon-ensembles
+Version: 2.1.0
+Summary: Efficient comparison of protein conformational ensembles using local order parameters
+Author-email: Adekunle Aina <kunleaina@gmail.com>
+Maintainer-email: Adekunle Aina <kunleaina@gmail.com>
+License: MIT
+Project-URL: Homepage, https://github.com/aai-research-lab/Prothon
+Project-URL: Repository, https://github.com/aai-research-lab/Prothon
+Project-URL: Issues, https://github.com/aai-research-lab/Prothon/issues
+Project-URL: Publication, https://doi.org/10.1021/acs.jcim.3c00145
+Keywords: protein-ensembles,conformational-ensembles,molecular-dynamics,intrinsically-disordered-proteins,jensen-shannon,order-parameters,structural-biology,biophysics
+Classifier: Development Status :: 4 - Beta
+Classifier: Intended Audience :: Science/Research
+Classifier: License :: OSI Approved :: MIT License
+Classifier: Operating System :: OS Independent
+Classifier: Programming Language :: Python :: 3
+Classifier: Programming Language :: Python :: 3.9
+Classifier: Programming Language :: Python :: 3.10
+Classifier: Programming Language :: Python :: 3.11
+Classifier: Programming Language :: Python :: 3.12
+Classifier: Programming Language :: Python :: 3.13
+Classifier: Topic :: Scientific/Engineering :: Bio-Informatics
+Classifier: Topic :: Scientific/Engineering :: Chemistry
+Classifier: Topic :: Scientific/Engineering :: Physics
+Requires-Python: <3.14,>=3.9
+Description-Content-Type: text/markdown
+License-File: LICENSE
+Requires-Dist: numpy>=1.22
+Requires-Dist: mdtraj>=1.9.7
+Requires-Dist: scipy>=1.9
+Requires-Dist: matplotlib>=3.5
+Requires-Dist: scikit-learn>=1.0
+Provides-Extra: dev
+Requires-Dist: pytest>=7.0; extra == "dev"
+Requires-Dist: pytest-cov>=4.0; extra == "dev"
+Requires-Dist: ruff>=0.4; extra == "dev"
+Dynamic: license-file
+
+<div align="center">
+
+# Prothon
+
+**How different are two protein ensembles — and is the difference real?**
+
+[![DOI](https://img.shields.io/badge/DOI-10.1021%2Facs.jcim.3c00145-blue)](https://doi.org/10.1021/acs.jcim.3c00145)
+[![Python](https://img.shields.io/badge/python-3.9%2B-blue)](https://pypi.org/project/prothon/)
+[![License: MIT](https://img.shields.io/badge/License-MIT-yellow.svg)](https://opensource.org/licenses/MIT)
+
+</div>
+
+---
+
+```bash
+prothon -traj wild_type.dcd,mutant.dcd -top topology.pdb -m cbcn
+```
+
+```
+CBCN (reference: ensemble 0)
+  ensemble 1: d = 0.2841 (floor 0.0472) — 34/76 residues differ
+```
+
+Prothon represents each conformational ensemble as a vector of probability
+distributions over **local order parameters** — contact numbers, virtual bond
+and torsion angles, solvent accessibility — and measures the Jensen–Shannon
+distance between corresponding distributions. Because the representation is
+local, no structural superposition is needed and the cost is linear in the
+number of frames rather than quadratic, which is what makes ensembles of tens
+of thousands of conformations tractable.
+
+**It reports what it cannot resolve.** Two independent halves of a *single*
+ensemble have a non-zero Jensen–Shannon distance, because a finite sample never
+reproduces a continuous distribution exactly. That self-distance is the
+resolution limit of the comparison, and Prothon measures it, prints it beside
+every result, and draws it on every figure. A difference smaller than the floor
+is reported as unresolvable rather than as a small difference.
+
+## Install
+
+```bash
+conda install -c conda-forge prothon      # preferred
+pip install prothon-ensembles             # the distribution name; see below
+prothon --info
+```
+
+The distribution on PyPI is `prothon-ensembles`, because `prothon` was
+registered in 2020 by an unrelated protobuf generator and PyPI names are
+permanent. The import name and the command are both `prothon`:
+
+```python
+from prothon import Prothon
+```
+
+## Use it
+
+From the command line:
+
+```bash
+prothon -traj a.dcd,b.dcd,c.dcd -top top.pdb -m cbcn,cata -o results --seed 0
+```
+
+| flag | meaning |
+|---|---|
+| `-traj` | Trajectory files, one per ensemble, comma-separated. Never concatenated. |
+| `-top` | Topology (PDB), shared by all of them. |
+| `-m` | Measures: `cbcn`, `cacn`, `caba`, `cata`, `sasa`. |
+| `-r` | Reference ensemble index (default 0). |
+| `-o` | Output root. Each measure writes `<measure>_output/`. |
+| `-d` | Projections: `pca`, `mds`, `tsne`. Off by default. |
+| `--seed` | Set it, and the run is reproducible. |
+
+Or from Python:
+
+```python
+from prothon import Prothon
+
+study = Prothon(["wild_type.dcd", "mutant.dcd"], "topology.pdb", random_state=0)
+results = study.compare_ensembles(methods="cbcn")
+
+comparison = results["cbcn"][0]
+comparison.global_dissimilarity   # 0.2841
+comparison.noise_floor            # 0.0472 — the resolution limit
+comparison.resolved               # True: the difference clears the floor
+comparison.significant            # bool array, one per residue
+comparison.local_dissimilarity    # per residue, zero where not significant
+comparison.raw_local_dissimilarity  # per residue, unmasked
+
+print(study.summary())
+```
+
+Each measure writes a directory containing the representation matrices as CSV,
+heatmaps, global and per-residue dissimilarity figures, and a `manifest.json`
+recording the inputs, parameters, seed and version that produced them.
+
+## The measures
+
+| name | quantity | circular |
+|---|---|---|
+| `cbcn` | C-beta contact number, smooth cutoff | |
+| `cacn` | C-alpha contact number, smooth cutoff | |
+| `caba` | Virtual Cα–Cα–Cα bond angle | |
+| `cata` | Virtual Cα torsion angle | yes |
+| `sasa` | Per-residue solvent accessible surface area | |
+
+Torsions live on a circle, so they are estimated with a von Mises kernel on a
+grid spanning a full turn. Each measure declares this, so the call site cannot
+forget it.
+
+## Upgrading from 2.0
+
+The API is unchanged and existing scripts run without modification, but
+**numbers will differ**, because the significance test in 2.0 was not sound: it
+compared each ensemble against a bootstrap of itself, a null about half as wide
+as the true sampling variability. Over 40 replicates in which both ensembles
+were drawn from an *identical* distribution, 2.0 called 100% of residues
+significantly different. The permutation test that replaces it sits at 1.2%.
+
+`--legacy-statistics` reproduces the old behaviour for regenerating published
+figures. [CHANGELOG.md](CHANGELOG.md) has the full account.
+
+`from Prothon import Prothon` still works and warns; use
+`from prothon import Prothon`.
+
+## Citation
+
+> Aina, A.; Hsueh, S. C. C.; Plotkin, S. S. PROTHON: A Local Order
+> Parameter-Based Method for Efficient Comparison of Protein Ensembles.
+> *J. Chem. Inf. Model.* **2023**, *63* (11), 3453–3461.
+> DOI: [10.1021/acs.jcim.3c00145](https://doi.org/10.1021/acs.jcim.3c00145)
+
+```bibtex
+@article{aina2023prothon,
+  author  = {Aina, Adekunle and Hsueh, Shawn C. C. and Plotkin, Steven S.},
+  title   = {PROTHON: A Local Order Parameter-Based Method for Efficient
+             Comparison of Protein Ensembles},
+  journal = {Journal of Chemical Information and Modeling},
+  volume  = {63},
+  number  = {11},
+  pages   = {3453--3461},
+  year    = {2023},
+  doi     = {10.1021/acs.jcim.3c00145},
+}
+```
+
+## Contributing
+
+See [CONTRIBUTING.md](CONTRIBUTING.md). The original version 1 code accompanying
+the 2023 paper is preserved unchanged under `legacy/`.
+
+## License
+
+MIT. See [LICENSE](LICENSE).
+
+Prothon was distributed under GPL-3.0 up to and including version 2.0.0.
+From 2.1.0 the project is MIT-licensed. Copies already obtained under
+GPL-3.0 remain governed by that licence — relicensing is not retroactive
+and takes nothing away from anyone who has a copy.
+
+---
+
+<div align="center">
+
+Built in the [AAI Research Lab](https://aai-research-lab.github.io) at
+California State University Dominguez Hills, on MDTraj, NumPy, SciPy,
+scikit-learn and Matplotlib.
+
+</div>