ggc-kinetics 0.2.0__tar.gz
This diff represents the content of publicly available package versions that have been released to one of the supported registries. The information contained in this diff is provided for informational purposes only and reflects changes between package versions as they appear in their respective public registries.
- ggc_kinetics-0.2.0/.github/workflows/ci.yml +30 -0
- ggc_kinetics-0.2.0/.github/workflows/publish.yml +67 -0
- ggc_kinetics-0.2.0/ATTRIBUTION.md +18 -0
- ggc_kinetics-0.2.0/CITATION.cff +23 -0
- ggc_kinetics-0.2.0/LICENSE +21 -0
- ggc_kinetics-0.2.0/MANIFEST.in +8 -0
- ggc_kinetics-0.2.0/PKG-INFO +109 -0
- ggc_kinetics-0.2.0/README.md +84 -0
- ggc_kinetics-0.2.0/RELEASE_CHECKLIST.md +95 -0
- ggc_kinetics-0.2.0/app/streamlit_app.py +48 -0
- ggc_kinetics-0.2.0/benchmarks/PHASE4_SUMMARY.md +49 -0
- ggc_kinetics-0.2.0/benchmarks/REAL_DATA_BENCHMARKS.md +113 -0
- ggc_kinetics-0.2.0/benchmarks/V02_RELEASE_NOTES_DRAFT.md +61 -0
- ggc_kinetics-0.2.0/examples/quickstart.py +56 -0
- ggc_kinetics-0.2.0/paper/paper.bib +136 -0
- ggc_kinetics-0.2.0/paper/paper.md +125 -0
- ggc_kinetics-0.2.0/pyproject.toml +32 -0
- ggc_kinetics-0.2.0/setup.cfg +4 -0
- ggc_kinetics-0.2.0/src/ggc_kinetics/__init__.py +7 -0
- ggc_kinetics-0.2.0/src/ggc_kinetics/_plate_parser.py +944 -0
- ggc_kinetics-0.2.0/src/ggc_kinetics/cli.py +49 -0
- ggc_kinetics-0.2.0/src/ggc_kinetics/detectors.py +91 -0
- ggc_kinetics-0.2.0/src/ggc_kinetics/parser.py +34 -0
- ggc_kinetics-0.2.0/src/ggc_kinetics/solvers.py +398 -0
- ggc_kinetics-0.2.0/src/ggc_kinetics.egg-info/PKG-INFO +109 -0
- ggc_kinetics-0.2.0/src/ggc_kinetics.egg-info/SOURCES.txt +39 -0
- ggc_kinetics-0.2.0/src/ggc_kinetics.egg-info/dependency_links.txt +1 -0
- ggc_kinetics-0.2.0/src/ggc_kinetics.egg-info/entry_points.txt +2 -0
- ggc_kinetics-0.2.0/src/ggc_kinetics.egg-info/requires.txt +6 -0
- ggc_kinetics-0.2.0/src/ggc_kinetics.egg-info/top_level.txt +1 -0
- ggc_kinetics-0.2.0/tests/data/bad_empty.csv +0 -0
- ggc_kinetics-0.2.0/tests/data/bad_no_wells.csv +2 -0
- ggc_kinetics-0.2.0/tests/data/bad_time_order.csv +45 -0
- ggc_kinetics-0.2.0/tests/data/k57_guard_ref.json +546 -0
- ggc_kinetics-0.2.0/tests/data/plate_coldC6.report.json +157 -0
- ggc_kinetics-0.2.0/tests/data/plate_coldC6_long_tsv.tsv +1249 -0
- ggc_kinetics-0.2.0/tests/data/plate_dead.report.json +157 -0
- ggc_kinetics-0.2.0/tests/data/plate_dead_tecan_transposed.csv +107 -0
- ggc_kinetics-0.2.0/tests/data/plate_healthy.report.json +161 -0
- ggc_kinetics-0.2.0/tests/data/plate_healthy_biotek_wide.csv +45 -0
- ggc_kinetics-0.2.0/tests/test_ggc_kinetics.py +275 -0
|
@@ -0,0 +1,30 @@
|
|
|
1
|
+
name: CI
|
|
2
|
+
|
|
3
|
+
on:
|
|
4
|
+
push:
|
|
5
|
+
pull_request:
|
|
6
|
+
|
|
7
|
+
jobs:
|
|
8
|
+
test:
|
|
9
|
+
runs-on: ubuntu-latest
|
|
10
|
+
strategy:
|
|
11
|
+
fail-fast: false
|
|
12
|
+
matrix:
|
|
13
|
+
python-version: ["3.10", "3.12"]
|
|
14
|
+
steps:
|
|
15
|
+
- uses: actions/checkout@v4
|
|
16
|
+
- uses: actions/setup-python@v5
|
|
17
|
+
with:
|
|
18
|
+
python-version: ${{ matrix.python-version }}
|
|
19
|
+
- name: Install
|
|
20
|
+
run: |
|
|
21
|
+
python -m pip install --upgrade pip
|
|
22
|
+
pip install -e .
|
|
23
|
+
- name: Unit tests
|
|
24
|
+
run: python -m unittest discover -s tests -v
|
|
25
|
+
- name: ggc-fit CLI smoke check
|
|
26
|
+
run: |
|
|
27
|
+
ggc-fit tests/data/plate_healthy_biotek_wide.csv --evidence MODELED --scale 0.05 --offset 0.040 --out ci_out
|
|
28
|
+
test -s ci_out/plate_healthy_biotek_wide.report.json
|
|
29
|
+
test -s ci_out/plate_healthy_biotek_wide.report.md
|
|
30
|
+
python examples/quickstart.py
|
|
@@ -0,0 +1,67 @@
|
|
|
1
|
+
name: Publish
|
|
2
|
+
|
|
3
|
+
# Manual only. Uses PyPI Trusted Publishing (OIDC): no API token is stored anywhere.
|
|
4
|
+
# Run with target=testpypi first, check the install, then run again with target=pypi.
|
|
5
|
+
|
|
6
|
+
on:
|
|
7
|
+
workflow_dispatch:
|
|
8
|
+
inputs:
|
|
9
|
+
target:
|
|
10
|
+
description: "Where to upload"
|
|
11
|
+
required: true
|
|
12
|
+
type: choice
|
|
13
|
+
options: [testpypi, pypi]
|
|
14
|
+
|
|
15
|
+
jobs:
|
|
16
|
+
build:
|
|
17
|
+
runs-on: ubuntu-latest
|
|
18
|
+
steps:
|
|
19
|
+
- uses: actions/checkout@v4
|
|
20
|
+
- uses: actions/setup-python@v5
|
|
21
|
+
with:
|
|
22
|
+
python-version: "3.12"
|
|
23
|
+
- name: Build and check
|
|
24
|
+
run: |
|
|
25
|
+
python -m pip install --upgrade pip build twine
|
|
26
|
+
python -m build
|
|
27
|
+
python -m twine check --strict dist/*
|
|
28
|
+
test "$(ls dist | wc -l)" -eq 2
|
|
29
|
+
sha256sum dist/*
|
|
30
|
+
- uses: actions/upload-artifact@v4
|
|
31
|
+
with:
|
|
32
|
+
name: dist
|
|
33
|
+
path: dist/
|
|
34
|
+
|
|
35
|
+
publish-testpypi:
|
|
36
|
+
if: inputs.target == 'testpypi'
|
|
37
|
+
needs: build
|
|
38
|
+
runs-on: ubuntu-latest
|
|
39
|
+
environment:
|
|
40
|
+
name: testpypi
|
|
41
|
+
url: https://test.pypi.org/p/ggc-kinetics
|
|
42
|
+
permissions:
|
|
43
|
+
id-token: write
|
|
44
|
+
steps:
|
|
45
|
+
- uses: actions/download-artifact@v4
|
|
46
|
+
with:
|
|
47
|
+
name: dist
|
|
48
|
+
path: dist/
|
|
49
|
+
- uses: pypa/gh-action-pypi-publish@release/v1
|
|
50
|
+
with:
|
|
51
|
+
repository-url: https://test.pypi.org/legacy/
|
|
52
|
+
|
|
53
|
+
publish-pypi:
|
|
54
|
+
if: inputs.target == 'pypi'
|
|
55
|
+
needs: build
|
|
56
|
+
runs-on: ubuntu-latest
|
|
57
|
+
environment:
|
|
58
|
+
name: pypi
|
|
59
|
+
url: https://pypi.org/p/ggc-kinetics
|
|
60
|
+
permissions:
|
|
61
|
+
id-token: write
|
|
62
|
+
steps:
|
|
63
|
+
- uses: actions/download-artifact@v4
|
|
64
|
+
with:
|
|
65
|
+
name: dist
|
|
66
|
+
path: dist/
|
|
67
|
+
- uses: pypa/gh-action-pypi-publish@release/v1
|
|
@@ -0,0 +1,18 @@
|
|
|
1
|
+
# Data attribution
|
|
2
|
+
|
|
3
|
+
This package redistributes **no third-party data files**. The numbers below are typed into the tests or the
|
|
4
|
+
benchmark, with their sources.
|
|
5
|
+
|
|
6
|
+
| Used in | Source | Licence | What is used |
|
|
7
|
+
|---|---|---|---|
|
|
8
|
+
| tests (invertase) | Michaelis & Menten 1913, Table 1 as translated in Johnson & Goody 2011, Biochemistry, PMC3381512 (doi:10.1021/bi201284u) | public domain (1913 data) | 31 time / fraction points |
|
|
9
|
+
| tests (PON1) | Goličnik & Bavec 2020, J. Enzyme Inhib. Med. Chem., PMC6896510 (doi:10.1080/14756366.2019.1695792), Table 1 | CC BY 4.0 | 12 published % error values |
|
|
10
|
+
| benchmark only (not shipped) | ICEKAT supplements (Olp, Kalous & Smith 2020, BMC Bioinformatics, PMC7222511, doi:10.1186/s12859-020-3513-y) and the ICEKAT `test.csv` example | CC BY 4.0 | parsed and fitted locally; see `benchmarks/REAL_DATA_BENCHMARKS.md` |
|
|
11
|
+
| benchmark only (not shipped) | HRP / ABTS Source Data of Kont et al. 2020, Nat. Commun. (PMC7666214, doi:10.1038/s41467-020-19561-8) | CC BY 4.0 | plain MM fits per series, locally |
|
|
12
|
+
| benchmark only (not shipped) | DaRUS datasets doi:10.18419/darus-3867 (Häussler 2024, SLAC / ABTS) and doi:10.18419/darus-2467 (Ngubane 2022, laccase / ABTS) | CC BY 4.0 | laccase / ABTS plates fitted locally |
|
|
13
|
+
|
|
14
|
+
`tests/data/` holds MODELED (simulated) plate files from GGC-inf Cycle 33 and the reports Cycle 34 stored for them.
|
|
15
|
+
They are not measurements.
|
|
16
|
+
|
|
17
|
+
If you redistribute any CC BY data next to this package, cite the original authors and DOIs above and state the
|
|
18
|
+
licence.
|
|
@@ -0,0 +1,23 @@
|
|
|
1
|
+
cff-version: 1.2.0
|
|
2
|
+
message: "If you use ggc-kinetics, please cite it as below."
|
|
3
|
+
title: "ggc-kinetics: exact enzyme progress-curve solvers, detection statistics and a plate-reader parser"
|
|
4
|
+
version: 0.2.0
|
|
5
|
+
date-released: "2026-11-20"
|
|
6
|
+
license: MIT
|
|
7
|
+
type: software
|
|
8
|
+
repository-code: "https://github.com/bpmcginley/ggc-kinetics"
|
|
9
|
+
authors:
|
|
10
|
+
- family-names: "McGinley"
|
|
11
|
+
given-names: "Bruce"
|
|
12
|
+
affiliation: "University of Massachusetts Amherst"
|
|
13
|
+
orcid: "https://orcid.org/0000-0000-0000-0000"
|
|
14
|
+
keywords:
|
|
15
|
+
- enzyme kinetics
|
|
16
|
+
- progress curves
|
|
17
|
+
- Michaelis-Menten
|
|
18
|
+
- Lambert W
|
|
19
|
+
- plate reader
|
|
20
|
+
abstract: >-
|
|
21
|
+
Closed-form and split-step solvers for enzyme progress curves (product inhibition, enzyme decay, coupled lag,
|
|
22
|
+
slow binding, tight binding, substrate inhibition), two detection statistics, and a parser for plate-reader
|
|
23
|
+
text exports. Detector thresholds are ASSUMED values calibrated on MODELED (simulated) plates only.
|
|
@@ -0,0 +1,21 @@
|
|
|
1
|
+
MIT License
|
|
2
|
+
|
|
3
|
+
Copyright (c) 2026 ggc-kinetics contributors
|
|
4
|
+
|
|
5
|
+
Permission is hereby granted, free of charge, to any person obtaining a copy
|
|
6
|
+
of this software and associated documentation files (the "Software"), to deal
|
|
7
|
+
in the Software without restriction, including without limitation the rights
|
|
8
|
+
to use, copy, modify, merge, publish, distribute, sublicense, and/or sell
|
|
9
|
+
copies of the Software, and to permit persons to whom the Software is
|
|
10
|
+
furnished to do so, subject to the following conditions:
|
|
11
|
+
|
|
12
|
+
The above copyright notice and this permission notice shall be included in all
|
|
13
|
+
copies or substantial portions of the Software.
|
|
14
|
+
|
|
15
|
+
THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
|
|
16
|
+
IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
|
|
17
|
+
FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE
|
|
18
|
+
AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
|
|
19
|
+
LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM,
|
|
20
|
+
OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
|
|
21
|
+
SOFTWARE.
|
|
@@ -0,0 +1,8 @@
|
|
|
1
|
+
include RELEASE_CHECKLIST.md
|
|
2
|
+
include README.md LICENSE ATTRIBUTION.md CITATION.cff
|
|
3
|
+
recursive-include tests *.py *.csv *.tsv *.json
|
|
4
|
+
recursive-include examples *.py
|
|
5
|
+
recursive-include app *.py
|
|
6
|
+
recursive-include benchmarks *.md
|
|
7
|
+
recursive-include paper *.md *.bib
|
|
8
|
+
recursive-include .github *.yml
|
|
@@ -0,0 +1,109 @@
|
|
|
1
|
+
Metadata-Version: 2.4
|
|
2
|
+
Name: ggc-kinetics
|
|
3
|
+
Version: 0.2.0
|
|
4
|
+
Summary: Zero-ODE enzyme progress-curve solvers, detectors and a plate-reader parser (frozen from GGC-inf)
|
|
5
|
+
Author: Bruce McGinley
|
|
6
|
+
License-Expression: MIT
|
|
7
|
+
Project-URL: Homepage, https://github.com/bpmcginley/ggc-kinetics
|
|
8
|
+
Project-URL: Repository, https://github.com/bpmcginley/ggc-kinetics
|
|
9
|
+
Project-URL: Issues, https://github.com/bpmcginley/ggc-kinetics/issues
|
|
10
|
+
Keywords: enzyme kinetics,progress curve,Michaelis-Menten,Lambert W,plate reader
|
|
11
|
+
Classifier: Programming Language :: Python :: 3
|
|
12
|
+
Classifier: Operating System :: OS Independent
|
|
13
|
+
Classifier: Development Status :: 3 - Alpha
|
|
14
|
+
Classifier: Intended Audience :: Science/Research
|
|
15
|
+
Classifier: Topic :: Scientific/Engineering :: Chemistry
|
|
16
|
+
Requires-Python: >=3.9
|
|
17
|
+
Description-Content-Type: text/markdown
|
|
18
|
+
License-File: LICENSE
|
|
19
|
+
Requires-Dist: numpy
|
|
20
|
+
Requires-Dist: scipy
|
|
21
|
+
Provides-Extra: app
|
|
22
|
+
Requires-Dist: streamlit; extra == "app"
|
|
23
|
+
Requires-Dist: matplotlib; extra == "app"
|
|
24
|
+
Dynamic: license-file
|
|
25
|
+
|
|
26
|
+
# ggc-kinetics 0.2.0
|
|
27
|
+
|
|
28
|
+
Exact enzyme progress-curve solvers that need no ODE integration, two detection statistics, and a plate-reader
|
|
29
|
+
parser. Every function is copied without changes to the math from the GGC-inf research archive (66 cycles) and
|
|
30
|
+
checked against the archive by `release_parity_check.py`: 20/20 checks, 0.0 difference for most.
|
|
31
|
+
|
|
32
|
+
**Read this first.** Every number in this project carries a tag: MEASURED (from real data), MODELED (from
|
|
33
|
+
simulation) or ASSUMED (a chosen value). The solvers are exact numerics. The **thresholds and validity limits come
|
|
34
|
+
from MODELED studies only**. They have not yet been confirmed on a real plate. On real published data so far
|
|
35
|
+
(see `benchmarks/REAL_DATA_BENCHMARKS.md`), 3 of 5 pre-registered gates passed.
|
|
36
|
+
|
|
37
|
+
## Install and run
|
|
38
|
+
|
|
39
|
+
```
|
|
40
|
+
pip install . # numpy, scipy
|
|
41
|
+
pip install ".[app]" # adds streamlit, matplotlib for the demo app
|
|
42
|
+
ggc-fit plate.csv --evidence MEASURED --out reports/
|
|
43
|
+
python -m unittest discover -s tests -v # from the repo root, package installed; < 1 s
|
|
44
|
+
python examples/quickstart.py # 1913 invertase fit + one simulated plate; < 1 s
|
|
45
|
+
```
|
|
46
|
+
|
|
47
|
+
`tests/data/plate_healthy_biotek_wide.csv` is a MODELED healthy plate that gives `P1: KILL` (replicate |ln Km ratio| 0.305 > 0.277): it is one draw from the false-reject tail of `P1` v2, whose healthy-plate power is 0.86 on 300 fresh MODELED plates, so about 14% of healthy plates fail.
|
|
48
|
+
It is kept on purpose, not swapped for a passing draw, to show that the replicate check fails loudly.
|
|
49
|
+
|
|
50
|
+
`ggc-fit FILE [--layout L.json] [--mode substrate|product] [--evidence MEASURED|MODELED] [--out DIR]` writes
|
|
51
|
+
`FILE.report.json` and `FILE.report.md`. It exits 0 on OK and 2 on a parse error (for example `NO_WELL_TABLE`).
|
|
52
|
+
|
|
53
|
+
## What is inside
|
|
54
|
+
|
|
55
|
+
| Function | What it does | Limits and evidence |
|
|
56
|
+
|---|---|---|
|
|
57
|
+
| `kp_kd_exact`, `fit_kp_vs_kd` | Lambert-W progress curve with product inhibition (Kp) and enzyme decay (kd); BIC choice between them | 1913 invertase: Km 16.52 mM (+0.11%), Vmax/Km 0.0461 (+0.32%), dBIC −19.29 (MODELED fit of MEASURED data) |
|
|
58
|
+
| `unified_kp_strang` | Kp + kd + non-enzymatic hydrolysis by Strang splitting | exact without hydrolysis; < 0.005 vs RK4 with it (MODELED) |
|
|
59
|
+
| `coupled_lag_decay` | coupled-assay lag with a decaying primary rate | < 1e-11 vs RK4 |
|
|
60
|
+
| `slow_binding_warped` | slow-onset inhibition by time warping | < 1e-8 vs 2-ODE RK4 |
|
|
61
|
+
| `morrison_exact` | tight binding, E0 comparable to Km | residual ≤ 1e-10. **E0 must be known to ±10%** (K66, MODELED) |
|
|
62
|
+
| `k57_map_fit` | substrate-inhibition test with a per-well pipetting prior | called = dBIC > −2.5 and 1/Ksi > 0.045 (ASSUMED). **Valid for A_peak ≤ 2.00, pipetting CV ≤ 5%, and `sig` close to the plate's real readout SD** (all MODELED; see below). Opt-in `dispersion_guard=True` (v0.2.0). Jacobian FD error ≤ 1e-6, solver residual ≤ 1e-8 |
|
|
63
|
+
| `z_rao_orth_robust` | saturation-curvature score for a whole plate | K59: tau 0.056, threshold 1.768; baseline threshold 4.60 (ASSUMED). Needs several S0 levels on one plate: one well gives 0 |
|
|
64
|
+
| `z_dsf_asym_boltz` | melt-transition score for DSF curves | call at ≥ 5.72 (K60, ASSUMED); fixed Tm and width |
|
|
65
|
+
| `parse_file`, `process_bytes` | CSV / TSV / text exports, long or wide, decimal commas, SoftMax blocks; Gen5 and Tecan headers skipped | no .xlsx. Errors come back as codes, not exceptions |
|
|
66
|
+
|
|
67
|
+
## K57 noise scale (new in 0.2.0)
|
|
68
|
+
|
|
69
|
+
`k57_map_fit` treats the readout SD you pass as `sig` as known. Its false-alarm bound (Wilson 90% upper ≤ 0.060,
|
|
70
|
+
MODELED) holds only if `sig` is close to the plate's real readout SD. In the pre-registered Q65 study (7 arms × 450
|
|
71
|
+
fresh MODELED plates, thresholds frozen), the nominal `sig` gave false-call w90 0.034 at the nominal noise floor,
|
|
72
|
+
0.076 at a 2x floor and 0.131 at a 3x floor (all MODELED). Do not use the default `sig` without checking it.
|
|
73
|
+
|
|
74
|
+
Two ways to fix the noise scale, neither of which re-tunes a threshold:
|
|
75
|
+
|
|
76
|
+
1. **Pass an estimated `sig` (preferred when blanks exist).** `sig` is a 1-D array with one SD per well (per S0
|
|
77
|
+
level). Estimate it from ≥ 3 noise-only blank or replicate reads per S0 level, for example
|
|
78
|
+
`sig = 0.02 * first_read + floor_hat`, and call `k57_map_fit(Y, s0n, t_grid, sig)`. In Q65 (variant V1) this
|
|
79
|
+
gave false-call w90 0.020-0.037 across arms and recall 0.991 / 0.982 (MODELED). **Caveat:** V1 used the
|
|
80
|
+
generator's own noise form (a 2% proportional term plus a floor). The 2% term is ASSUMED, which favours V1; on a
|
|
81
|
+
real plate the proportional part must be estimated too.
|
|
82
|
+
2. **`dispersion_guard=True` (fallback with no blanks).** The null fit's chi-square sets
|
|
83
|
+
`birge = max(1, sqrt(chi2_null / dof_null))`, with `dof_null = max(W*M - 4 - W, 1)`, and each plate with
|
|
84
|
+
`birge > 1` is refitted at `sig * birge`. The result carries `"birge"` (all 1.0 when the guard is off). In Q65
|
|
85
|
+
(variant V2) this passed every arm, but with a thin margin at the 2x floor: w90 0.0555 against the 0.060 gate
|
|
86
|
+
(MODELED). The guard is one-sided: it never shrinks `sig`. The default `dispersion_guard=False` gives output
|
|
87
|
+
bit-identical to 0.1.0 on the same platform (checked on Windows); across operating systems the last 2-3 digits
|
|
88
|
+
can differ, so the test suite compares to a relative 1e-10 (plus an absolute 1e-8 for fitted iksi and km).
|
|
89
|
+
|
|
90
|
+
Neither fix has been checked on a real plate.
|
|
91
|
+
|
|
92
|
+
## Not claimed
|
|
93
|
+
|
|
94
|
+
- The detectors and K57 rule have not yet been shown to work on a fresh real plate. On one real, **non-fresh**
|
|
95
|
+
plate (darus-3867 SLAC / ABTS, Release Phase 2) Z_Rao scored enzyme rows 3.48-3.82 and no-enzyme rows 0.32-0.50
|
|
96
|
+
(MODELED fit of MEASURED data): the controls stayed below 4.60, but the enzyme rows did not reach it (G1 FAIL,
|
|
97
|
+
2/3 gates). The K59 statistic put the enzyme rows at 2.26-2.61 and the controls at 0.05-0.07 against 1.768.
|
|
98
|
+
- **K59 has not been validated on a fresh real plate with matched blanks.** K59 (tau 0.056, threshold 1.768,
|
|
99
|
+
ASSUMED) is calibrated on MODELED plates and checked only on the non-fresh darus-3867 plate above. Two
|
|
100
|
+
pre-registered searches screened 12 public candidates and none qualified, so no fresh plate was run
|
|
101
|
+
(`benchmarks/REAL_DATA_BENCHMARKS.md`). Public deposits mostly give initial rates or processed values, not raw
|
|
102
|
+
multi-S0 progress curves with blanks.
|
|
103
|
+
- `ggc-fit` reports the parse and the frozen contract tests; it does not fit every model listed above.
|
|
104
|
+
- The Streamlit app (`app/streamlit_app.py`) passed a headless `AppTest` smoke run (streamlit 1.67.0) on one
|
|
105
|
+
MODELED plate, with the upload widget stubbed because AppTest cannot drive it. It was also checked once by hand in
|
|
106
|
+
Chrome (20 November 2026): upload of the healthy plate, MODELED, scale 0.05, offset 0.040; title, report and trace
|
|
107
|
+
plot rendered.
|
|
108
|
+
|
|
109
|
+
MIT licence. Data sources: `ATTRIBUTION.md`.
|
|
@@ -0,0 +1,84 @@
|
|
|
1
|
+
# ggc-kinetics 0.2.0
|
|
2
|
+
|
|
3
|
+
Exact enzyme progress-curve solvers that need no ODE integration, two detection statistics, and a plate-reader
|
|
4
|
+
parser. Every function is copied without changes to the math from the GGC-inf research archive (66 cycles) and
|
|
5
|
+
checked against the archive by `release_parity_check.py`: 20/20 checks, 0.0 difference for most.
|
|
6
|
+
|
|
7
|
+
**Read this first.** Every number in this project carries a tag: MEASURED (from real data), MODELED (from
|
|
8
|
+
simulation) or ASSUMED (a chosen value). The solvers are exact numerics. The **thresholds and validity limits come
|
|
9
|
+
from MODELED studies only**. They have not yet been confirmed on a real plate. On real published data so far
|
|
10
|
+
(see `benchmarks/REAL_DATA_BENCHMARKS.md`), 3 of 5 pre-registered gates passed.
|
|
11
|
+
|
|
12
|
+
## Install and run
|
|
13
|
+
|
|
14
|
+
```
|
|
15
|
+
pip install . # numpy, scipy
|
|
16
|
+
pip install ".[app]" # adds streamlit, matplotlib for the demo app
|
|
17
|
+
ggc-fit plate.csv --evidence MEASURED --out reports/
|
|
18
|
+
python -m unittest discover -s tests -v # from the repo root, package installed; < 1 s
|
|
19
|
+
python examples/quickstart.py # 1913 invertase fit + one simulated plate; < 1 s
|
|
20
|
+
```
|
|
21
|
+
|
|
22
|
+
`tests/data/plate_healthy_biotek_wide.csv` is a MODELED healthy plate that gives `P1: KILL` (replicate |ln Km ratio| 0.305 > 0.277): it is one draw from the false-reject tail of `P1` v2, whose healthy-plate power is 0.86 on 300 fresh MODELED plates, so about 14% of healthy plates fail.
|
|
23
|
+
It is kept on purpose, not swapped for a passing draw, to show that the replicate check fails loudly.
|
|
24
|
+
|
|
25
|
+
`ggc-fit FILE [--layout L.json] [--mode substrate|product] [--evidence MEASURED|MODELED] [--out DIR]` writes
|
|
26
|
+
`FILE.report.json` and `FILE.report.md`. It exits 0 on OK and 2 on a parse error (for example `NO_WELL_TABLE`).
|
|
27
|
+
|
|
28
|
+
## What is inside
|
|
29
|
+
|
|
30
|
+
| Function | What it does | Limits and evidence |
|
|
31
|
+
|---|---|---|
|
|
32
|
+
| `kp_kd_exact`, `fit_kp_vs_kd` | Lambert-W progress curve with product inhibition (Kp) and enzyme decay (kd); BIC choice between them | 1913 invertase: Km 16.52 mM (+0.11%), Vmax/Km 0.0461 (+0.32%), dBIC −19.29 (MODELED fit of MEASURED data) |
|
|
33
|
+
| `unified_kp_strang` | Kp + kd + non-enzymatic hydrolysis by Strang splitting | exact without hydrolysis; < 0.005 vs RK4 with it (MODELED) |
|
|
34
|
+
| `coupled_lag_decay` | coupled-assay lag with a decaying primary rate | < 1e-11 vs RK4 |
|
|
35
|
+
| `slow_binding_warped` | slow-onset inhibition by time warping | < 1e-8 vs 2-ODE RK4 |
|
|
36
|
+
| `morrison_exact` | tight binding, E0 comparable to Km | residual ≤ 1e-10. **E0 must be known to ±10%** (K66, MODELED) |
|
|
37
|
+
| `k57_map_fit` | substrate-inhibition test with a per-well pipetting prior | called = dBIC > −2.5 and 1/Ksi > 0.045 (ASSUMED). **Valid for A_peak ≤ 2.00, pipetting CV ≤ 5%, and `sig` close to the plate's real readout SD** (all MODELED; see below). Opt-in `dispersion_guard=True` (v0.2.0). Jacobian FD error ≤ 1e-6, solver residual ≤ 1e-8 |
|
|
38
|
+
| `z_rao_orth_robust` | saturation-curvature score for a whole plate | K59: tau 0.056, threshold 1.768; baseline threshold 4.60 (ASSUMED). Needs several S0 levels on one plate: one well gives 0 |
|
|
39
|
+
| `z_dsf_asym_boltz` | melt-transition score for DSF curves | call at ≥ 5.72 (K60, ASSUMED); fixed Tm and width |
|
|
40
|
+
| `parse_file`, `process_bytes` | CSV / TSV / text exports, long or wide, decimal commas, SoftMax blocks; Gen5 and Tecan headers skipped | no .xlsx. Errors come back as codes, not exceptions |
|
|
41
|
+
|
|
42
|
+
## K57 noise scale (new in 0.2.0)
|
|
43
|
+
|
|
44
|
+
`k57_map_fit` treats the readout SD you pass as `sig` as known. Its false-alarm bound (Wilson 90% upper ≤ 0.060,
|
|
45
|
+
MODELED) holds only if `sig` is close to the plate's real readout SD. In the pre-registered Q65 study (7 arms × 450
|
|
46
|
+
fresh MODELED plates, thresholds frozen), the nominal `sig` gave false-call w90 0.034 at the nominal noise floor,
|
|
47
|
+
0.076 at a 2x floor and 0.131 at a 3x floor (all MODELED). Do not use the default `sig` without checking it.
|
|
48
|
+
|
|
49
|
+
Two ways to fix the noise scale, neither of which re-tunes a threshold:
|
|
50
|
+
|
|
51
|
+
1. **Pass an estimated `sig` (preferred when blanks exist).** `sig` is a 1-D array with one SD per well (per S0
|
|
52
|
+
level). Estimate it from ≥ 3 noise-only blank or replicate reads per S0 level, for example
|
|
53
|
+
`sig = 0.02 * first_read + floor_hat`, and call `k57_map_fit(Y, s0n, t_grid, sig)`. In Q65 (variant V1) this
|
|
54
|
+
gave false-call w90 0.020-0.037 across arms and recall 0.991 / 0.982 (MODELED). **Caveat:** V1 used the
|
|
55
|
+
generator's own noise form (a 2% proportional term plus a floor). The 2% term is ASSUMED, which favours V1; on a
|
|
56
|
+
real plate the proportional part must be estimated too.
|
|
57
|
+
2. **`dispersion_guard=True` (fallback with no blanks).** The null fit's chi-square sets
|
|
58
|
+
`birge = max(1, sqrt(chi2_null / dof_null))`, with `dof_null = max(W*M - 4 - W, 1)`, and each plate with
|
|
59
|
+
`birge > 1` is refitted at `sig * birge`. The result carries `"birge"` (all 1.0 when the guard is off). In Q65
|
|
60
|
+
(variant V2) this passed every arm, but with a thin margin at the 2x floor: w90 0.0555 against the 0.060 gate
|
|
61
|
+
(MODELED). The guard is one-sided: it never shrinks `sig`. The default `dispersion_guard=False` gives output
|
|
62
|
+
bit-identical to 0.1.0 on the same platform (checked on Windows); across operating systems the last 2-3 digits
|
|
63
|
+
can differ, so the test suite compares to a relative 1e-10 (plus an absolute 1e-8 for fitted iksi and km).
|
|
64
|
+
|
|
65
|
+
Neither fix has been checked on a real plate.
|
|
66
|
+
|
|
67
|
+
## Not claimed
|
|
68
|
+
|
|
69
|
+
- The detectors and K57 rule have not yet been shown to work on a fresh real plate. On one real, **non-fresh**
|
|
70
|
+
plate (darus-3867 SLAC / ABTS, Release Phase 2) Z_Rao scored enzyme rows 3.48-3.82 and no-enzyme rows 0.32-0.50
|
|
71
|
+
(MODELED fit of MEASURED data): the controls stayed below 4.60, but the enzyme rows did not reach it (G1 FAIL,
|
|
72
|
+
2/3 gates). The K59 statistic put the enzyme rows at 2.26-2.61 and the controls at 0.05-0.07 against 1.768.
|
|
73
|
+
- **K59 has not been validated on a fresh real plate with matched blanks.** K59 (tau 0.056, threshold 1.768,
|
|
74
|
+
ASSUMED) is calibrated on MODELED plates and checked only on the non-fresh darus-3867 plate above. Two
|
|
75
|
+
pre-registered searches screened 12 public candidates and none qualified, so no fresh plate was run
|
|
76
|
+
(`benchmarks/REAL_DATA_BENCHMARKS.md`). Public deposits mostly give initial rates or processed values, not raw
|
|
77
|
+
multi-S0 progress curves with blanks.
|
|
78
|
+
- `ggc-fit` reports the parse and the frozen contract tests; it does not fit every model listed above.
|
|
79
|
+
- The Streamlit app (`app/streamlit_app.py`) passed a headless `AppTest` smoke run (streamlit 1.67.0) on one
|
|
80
|
+
MODELED plate, with the upload widget stubbed because AppTest cannot drive it. It was also checked once by hand in
|
|
81
|
+
Chrome (20 November 2026): upload of the healthy plate, MODELED, scale 0.05, offset 0.040; title, report and trace
|
|
82
|
+
plot rendered.
|
|
83
|
+
|
|
84
|
+
MIT licence. Data sources: `ATTRIBUTION.md`.
|
|
@@ -0,0 +1,95 @@
|
|
|
1
|
+
# Release checklist: ggc-kinetics v0.2.0
|
|
2
|
+
|
|
3
|
+
Everything up to the upload commands has been run and checked (Release Phase 6, v0.2.0-rc1). The steps below are manual.
|
|
4
|
+
|
|
5
|
+
## 0. The only two fields you must confirm
|
|
6
|
+
|
|
7
|
+
| Placeholder | Where | Replace with |
|
|
8
|
+
|---|---|---|
|
|
9
|
+
| `GITHUB_HANDLE` | `pyproject.toml` (`[project.urls]`), `CITATION.cff` (`repository-code`) | your GitHub user or org name |
|
|
10
|
+
| ORCID `0000-0000-0000-0000` | `CITATION.cff`, `paper/paper.md` | your real ORCID iD |
|
|
11
|
+
|
|
12
|
+
Find every occurrence (PowerShell, from this folder):
|
|
13
|
+
|
|
14
|
+
```powershell
|
|
15
|
+
Select-String -Path pyproject.toml, CITATION.cff, paper\paper.md -Pattern "GITHUB_HANDLE", "0000-0000-0000-0000"
|
|
16
|
+
```
|
|
17
|
+
|
|
18
|
+
Also check that the author name and affiliation are right in all three files.
|
|
19
|
+
|
|
20
|
+
## 1. GitHub
|
|
21
|
+
|
|
22
|
+
The repository ships `.github/workflows/ci.yml` (listed in `MANIFEST.in`). On every push and pull request it runs
|
|
23
|
+
on ubuntu-latest with Python 3.10 and 3.12: `pip install -e .`, `python -m unittest discover -s tests`, a `ggc-fit`
|
|
24
|
+
CLI smoke run on `tests/data/plate_healthy_biotek_wide.csv`, and `examples/quickstart.py`. Check that the first
|
|
25
|
+
run after the push is green before going on.
|
|
26
|
+
|
|
27
|
+
```powershell
|
|
28
|
+
git init
|
|
29
|
+
git add .
|
|
30
|
+
git commit -m "ggc-kinetics v0.2.0"
|
|
31
|
+
git branch -M main
|
|
32
|
+
git remote add origin https://github.com/bpmcginley/ggc-kinetics.git
|
|
33
|
+
git push -u origin main
|
|
34
|
+
git tag v0.2.0
|
|
35
|
+
git push origin v0.2.0
|
|
36
|
+
```
|
|
37
|
+
|
|
38
|
+
## 2. Build and check
|
|
39
|
+
|
|
40
|
+
```powershell
|
|
41
|
+
Remove-Item -Recurse -Force dist -ErrorAction SilentlyContinue
|
|
42
|
+
python -m pip install --upgrade build twine
|
|
43
|
+
python -m build
|
|
44
|
+
python -m twine check --strict dist/*
|
|
45
|
+
```
|
|
46
|
+
|
|
47
|
+
`dist/` must hold exactly two files: `ggc_kinetics-0.2.0-py3-none-any.whl` and `ggc_kinetics-0.2.0.tar.gz`.
|
|
48
|
+
|
|
49
|
+
## 3. TestPyPI, then a fresh-venv install
|
|
50
|
+
|
|
51
|
+
v0.2.0 was uploaded with PyPI Trusted Publishing instead of `twine upload`: `.github/workflows/publish.yml`
|
|
52
|
+
(manual `workflow_dispatch`, input `target` = `testpypi` or `pypi`, GitHub environments of the same names). Register
|
|
53
|
+
a pending trusted publisher on test.pypi.org and pypi.org for owner `bpmcginley`, repo `ggc-kinetics`, workflow
|
|
54
|
+
`publish.yml`, and the matching environment. The `twine upload` lines below are the token-based fallback.
|
|
55
|
+
|
|
56
|
+
```powershell
|
|
57
|
+
python -m twine upload --repository testpypi dist/*
|
|
58
|
+
```
|
|
59
|
+
|
|
60
|
+
```powershell
|
|
61
|
+
python -m venv $env:TEMP\ggc_tp; & $env:TEMP\ggc_tp\Scripts\python.exe -m pip install --index-url https://test.pypi.org/simple/ --extra-index-url https://pypi.org/simple/ ggc-kinetics
|
|
62
|
+
```
|
|
63
|
+
|
|
64
|
+
```powershell
|
|
65
|
+
& $env:TEMP\ggc_tp\Scripts\ggc-fit.exe tests\data\plate_healthy_biotek_wide.csv --evidence MODELED --scale 0.05 --offset 0.040 --out $env:TEMP\ggc_tp_out
|
|
66
|
+
```
|
|
67
|
+
|
|
68
|
+
That should exit 0 and print P1 KILL (expected; see README). Then upload to PyPI:
|
|
69
|
+
|
|
70
|
+
```powershell
|
|
71
|
+
python -m twine upload dist/*
|
|
72
|
+
```
|
|
73
|
+
|
|
74
|
+
## 4. Browser check of the Streamlit app
|
|
75
|
+
|
|
76
|
+
```powershell
|
|
77
|
+
pip install ".[app]"
|
|
78
|
+
streamlit run app/streamlit_app.py
|
|
79
|
+
```
|
|
80
|
+
|
|
81
|
+
Upload `tests/data/plate_healthy_biotek_wide.csv`, set evidence MODELED, scale 0.05 and offset 0.040, and check that
|
|
82
|
+
the report and traces appear. Only after this works, remove the sentence "It has not been used in a browser." from
|
|
83
|
+
`README.md`.
|
|
84
|
+
|
|
85
|
+
## 5. Zenodo DOI (before JOSS)
|
|
86
|
+
|
|
87
|
+
- Log in to https://zenodo.org with GitHub and switch on the `ggc-kinetics` repository under GitHub integration.
|
|
88
|
+
- Make a GitHub release from tag `v0.2.0`. Zenodo archives it and mints a DOI.
|
|
89
|
+
- Add the DOI to `CITATION.cff` (`doi:`) and the README, and keep it for the JOSS submission form.
|
|
90
|
+
|
|
91
|
+
## 6. Paper
|
|
92
|
+
|
|
93
|
+
- Submit `paper/paper.md` and `paper/paper.bib` to JOSS (https://joss.theoj.org/papers/new); it needs the public
|
|
94
|
+
GitHub repository from step 1.
|
|
95
|
+
- Post the preprint to bioRxiv.
|
|
@@ -0,0 +1,48 @@
|
|
|
1
|
+
"""ggc-kinetics demo app. Run: pip install "ggc-kinetics[app]"; streamlit run app/streamlit_app.py
|
|
2
|
+
|
|
3
|
+
Upload a plate-reader export, see the contract report, and plot the parsed traces. Every number is shown with the
|
|
4
|
+
evidence tag you choose (MEASURED for a real plate). No fitting knobs are exposed: the rules are frozen.
|
|
5
|
+
"""
|
|
6
|
+
import json
|
|
7
|
+
|
|
8
|
+
import matplotlib.pyplot as plt
|
|
9
|
+
import numpy as np
|
|
10
|
+
import streamlit as st
|
|
11
|
+
|
|
12
|
+
from ggc_kinetics import __version__
|
|
13
|
+
from ggc_kinetics.parser import dumps, process_bytes, report_markdown
|
|
14
|
+
|
|
15
|
+
st.set_page_config(page_title="ggc-kinetics", layout="wide")
|
|
16
|
+
st.title(f"ggc-kinetics v{__version__} plate report")
|
|
17
|
+
st.caption("Frozen GGC-inf rules. Thresholds are ASSUMED values calibrated on MODELED plates.")
|
|
18
|
+
|
|
19
|
+
up = st.file_uploader("Plate export (CSV / TSV / TXT)", type=["csv", "tsv", "txt"])
|
|
20
|
+
lay = st.file_uploader("Optional layout JSON", type=["json"])
|
|
21
|
+
c1, c2, c3, c4 = st.columns(4)
|
|
22
|
+
evidence = c1.selectbox("Evidence", ["MEASURED", "MODELED"])
|
|
23
|
+
mode = c2.selectbox("Readout", ["substrate", "product"])
|
|
24
|
+
scale = c3.number_input("Absorbance per 50 uM (ASSUMED)", value=1.0, format="%.4f")
|
|
25
|
+
offset = c4.number_input("Background absorbance (ASSUMED)", value=0.0, format="%.4f")
|
|
26
|
+
|
|
27
|
+
if up is not None:
|
|
28
|
+
rep, parsed = process_bytes(up.getvalue(), up.name, layout_raw=lay.getvalue() if lay else None,
|
|
29
|
+
layout_name=lay.name if lay else None, scale=scale, offset=offset, mode=mode,
|
|
30
|
+
evidence=evidence)
|
|
31
|
+
if rep["status"] != "OK":
|
|
32
|
+
st.error(f"{rep['error']['code']}: {rep['error']['message']}")
|
|
33
|
+
else:
|
|
34
|
+
st.markdown(report_markdown(rep))
|
|
35
|
+
if parsed is not None:
|
|
36
|
+
fig, ax = plt.subplots(figsize=(7, 4))
|
|
37
|
+
t = np.asarray(parsed["times_s"]) / 60.0
|
|
38
|
+
A, present = np.asarray(parsed["A"]), np.asarray(parsed["present"])
|
|
39
|
+
for r in range(A.shape[0]):
|
|
40
|
+
for c in range(A.shape[1]):
|
|
41
|
+
if present[r, c]:
|
|
42
|
+
ax.plot(t, A[r, c], lw=0.7)
|
|
43
|
+
ax.set_xlabel("time (min)")
|
|
44
|
+
ax.set_ylabel("absorbance")
|
|
45
|
+
st.pyplot(fig)
|
|
46
|
+
st.download_button("Download report JSON", dumps(rep), file_name="report.json")
|
|
47
|
+
with st.expander("Raw report"):
|
|
48
|
+
st.code(json.dumps(rep, indent=1)[:20000])
|
|
@@ -0,0 +1,49 @@
|
|
|
1
|
+
# GGC-inf Release Phase 4 (v0.2.0-dev): summary
|
|
2
|
+
|
|
3
|
+
Tags: MEASURED, MODELED, "MODELED fit of MEASURED data", ASSUMED (L7). Every script ran twice; run1 and run2 logs
|
|
4
|
+
differ only in the tag and wall-time lines.
|
|
5
|
+
|
|
6
|
+
## 1. Fresh real-plate K59 benchmark: NO_QUALIFYING_FRESH_PLATE_FOUND
|
|
7
|
+
|
|
8
|
+
- Pre-registration `benchmark_fresh_real_k59.py` was hashed before any search (sha256 c871c5cd...c7128715).
|
|
9
|
+
- 6 candidates were screened from metadata only (table in `REAL_DATA_BENCHMARKS.md` and `fresh_k59_screening.json`).
|
|
10
|
+
All 6 were rejected. Main reasons: no matched no-enzyme wells on the same plate, fewer than 4 S0 levels on one plate,
|
|
11
|
+
no raw data, simulated data, or already used (darus-3867, darus-2467).
|
|
12
|
+
- Gates G1-G3 were NOT RUN. Nothing was downloaded, nothing simulated was substituted, nothing excluded was reused.
|
|
13
|
+
- So K59 (tau 0.056, threshold 1.768, ASSUMED) still rests on MODELED plates plus one non-fresh real plate.
|
|
14
|
+
|
|
15
|
+
## 2. Q64 confounder check of the frozen K57 MAP rule: FAIL (G2, arm C)
|
|
16
|
+
|
|
17
|
+
Independent mass-action ODE (no code shared with `ggc_kinetics.solvers`); `k57_map_fit` frozen (thr -2.5, iksi 0.045).
|
|
18
|
+
450 null plates per arm (3 seed blocks x 150), 1,350 total. Pre-registration sha256 a9e1da62...9b133352.
|
|
19
|
+
|
|
20
|
+
| Arm | Mechanism (MODELED, rates ASSUMED) | False calls | Rate | Wilson 90% upper | G2 (<= 0.060) |
|
|
21
|
+
|---|---|---|---|---|---|
|
|
22
|
+
| A | product inhibition, Kp = 3 Km | 3/450 | 0.0067 | 0.0166 | PASS |
|
|
23
|
+
| B | ES inactivated at 2 kd, free E at kd | 14/450 | 0.0311 | 0.0476 | PASS |
|
|
24
|
+
| C | MM + decay, noise floor 0.015 | 34/450 | 0.0756 | 0.0987 | FAIL |
|
|
25
|
+
| pooled | | 51/1350 | 0.0378 | 0.0473 | PASS |
|
|
26
|
+
|
|
27
|
+
- G1 numerics (MEASURED): Jacobian vs central FD max rel err 1.03e-11 (<= 1e-6); invariants max |F| 1.07e-13
|
|
28
|
+
(<= 1e-8); tolerance halving max |d readout| 5.21e-09 (<= 1e-6). PASS.
|
|
29
|
+
- G3 (MEASURED): arm invocations 5.07-5.93 s each (< 30 s). PASS.
|
|
30
|
+
- L21: the number that moved is arm C's false-alarm rate, from 10/450 (Cycle 66, floor 0.0075) to 34/450 (floor
|
|
31
|
+
0.015). The mechanism is noise mis-specification: the rule takes its readout SD as known
|
|
32
|
+
(SIG = 0.015 x max(S0, 0.5)), and a noise floor 2x higher than that inflates dBIC on null plates. No threshold was
|
|
33
|
+
re-tuned.
|
|
34
|
+
- My pre-registered predictions were wrong in two places. I expected B to be the most likely to fail and C to pass;
|
|
35
|
+
the opposite happened.
|
|
36
|
+
|
|
37
|
+
## 3. Regression
|
|
38
|
+
|
|
39
|
+
- Archive suite `sprint15_cycle30_regression_suite.py`: 42/42 PASSED, twice (only the wall_time line differs).
|
|
40
|
+
- Unit tests: 14/14 OK, twice, run against a fresh venv with the rc1 wheel installed (sha256 b88157b2...ed9c5f).
|
|
41
|
+
|
|
42
|
+
## Caveats for the v0.2.0 README and paper
|
|
43
|
+
|
|
44
|
+
- The K59 detector has still not been tested on a fresh real plate with matched blanks. Say so plainly.
|
|
45
|
+
- The K57 rule's false-alarm guarantee needs the user's readout SD to be close to the assumed SIG. If the real
|
|
46
|
+
noise floor is about 2x higher, the false-alarm rate rose to 7.6% (w90 0.099) in MODELED data. Advise estimating SD
|
|
47
|
+
from blank or replicate wells, and state this as a validity condition next to CV <= 5% and A_peak <= 2.
|
|
48
|
+
- Product inhibition (Kp = 3 Km) and faster ES inactivation (2 kd) did not raise false alarms above the bound
|
|
49
|
+
(MODELED).
|