ssys — recast ODEs into canonical S‑system (or GMA) form

Source: README.md | v0.7.0 | 2026-07-16

ssys converts a system of ordinary differential equations into exact S‑system or GMA (power‑law) form and verifies the result symbolically and numerically. It reads and writes Antimony, can parse SBML directly, and can generate a Jupyter notebook report comparing the original and recast models.

Install

pip install ssys

Supported on Python 3.10, 3.11, and 3.12, and on Linux, macOS, and Windows (all tested in CI). The install pulls in the SBML‑first parser and ODE simulation stack (libRoadRunner, Antimony, python‑libsbml).

Quickstart

Recast one model

Antimony in, Antimony out — in Python, using only the stable top‑level API:

import ssys

text = open("model.ant").read()
sym = ssys.parse_antimony_via_sbml(text)          # Antimony -> symbolic ODE system
result = ssys.recast_to_ssystem(sym, mode="simplified")
antimony = ssys.ssystem_to_antimony(result, model_name="model_recast", mode="simplified")

open("model_recast.ant", "w").write(antimony)
print("recast as:", ssys.classify_result(result, mode="simplified").value)

Already have SBML? Replace the first two lines with sym = ssys.parse_sbml("model.xml").

From the command line, the ssys-recast tool works on a manifest — a text file with one .ant path per line. To recast a single model, point it at a one‑line manifest:

printf '%s\n' model.ant > models.manifest
ssys-recast --manifest models.manifest --outdir out
# writes out/model_recast.ant and out/recast_report.ipynb

Recast a batch and validate

ssys-recast --manifest test_models1/models.manifest --outdir out \
  --mode simplified --validate --validation-profile strict

--validate runs the correctness checks and writes a *_validation.json next to each recast. strict is the release‑grade profile: a model counts as validated only when every required check passes.

What ssys does

• S‑system — at most two power‑law terms per equation, one production and one degradation: Ẋᵢ = αᵢ · ∏ⱼ Xⱼ^gᵢⱼ − βᵢ · ∏ⱼ Xⱼ^hᵢⱼ.
• GMA — each equation is a sum of power‑law monomials.
• Automatically lifts rational and composite terms (exp, sin, log, 1/(X+Y), …) into auxiliary variables, so a broad class of ODEs becomes exactly recastable — not just those already in power‑law form.
• Two output modes: simplified (default; preserves structure, allows single‑term equations) and canonical (strict two‑term form via ε‑splitting, following Savageau & Voit 1987).

See RECASTING.md for the theory and worked examples, and CORRECTNESS_SPEC.md for the exact supported input/output contract. The stable API surface is documented in PUBLIC_API.md.

Validation

Validation is fail‑closed: overall_pass is true only when every required check for the selected profile returns pass. Unsupported, not‑attempted, timed‑out, inconclusive, and failed checks are never passes. Every result carries a machine‑readable reason and the reports carry a schema_version; load the packaged JSON Schema with ssys.load_validation_report_schema().

• Symbolic evidence is an exact algebraic proof — the Jacobian chain‑rule identity J_Φ(Z)·f_recast(Z) = f_orig(Φ(Z)) simplifies to zero.
• Numerical evidence is pointwise support at deterministic random samples (ε = 10⁻⁵); it can find counterexamples but does not prove global equivalence.
• Trajectory evidence is solver‑backed behavioral support over a time grid (default 3% peak‑scaled threshold); exact claims require symbolic validation.

Supported input

Antimony reactions (A + B -> C; k*A*B), initializations (X = 2.5), explicit rate rules (X' = ...), boundary species ($X), positive parameters, elementary functions (exp, log, sin, cos, tan, sqrt, sinh, cosh, tanh, …), rational functions, and assignment rules (Z := X + Y). Simulation settings come from @SIM comments, e.g. // @SIM T_START=0 T_END=100 N_STEPS=500 EPS_INIT=1e-6.

For SBML, species hasOnlySubstanceUnits, non‑unit and constant compartment volumes, a species/model conversionFactor, and constant reaction stoichiometry (including a stoichiometryMath that folds to a constant) are all honored.

Not yet supported: modules, events/piecewise functions, and non‑positive state variables. At the SBML trust boundary these are rejected with a structured unsupported_feature error rather than mis‑integrated: variable reaction stoichiometry (a non‑constant stoichiometryMath, or a speciesReference id driven by a rule/InitialAssignment) and time‑varying compartment volume (a rate‑rule compartment, or an assignment‑rule compartment that does not fold to a constant, holding a concentration species — its -[S]·(dV/dt)/V dilution term is not modeled). S‑systems require positive states, so a zero initial condition is ε‑regularized to a small positive EPS_INIT (default 1e-6, configurable via @SIM); this solves a slightly perturbed IVP near t = 0.

Scope and trust boundary

ssys is alpha software. Treat the APIs, generated Antimony details, and the validation‑report format as subject to change until the package leaves alpha. Local artifact builds, validation reports, and benchmark evidence are the source of truth for release claims.

Trust boundary: ssys treats Antimony and SBML inputs as trusted local scientific model files, not as safe untrusted uploads. Do not expose the CLI or parser directly to arbitrary user‑submitted model text in a multi‑tenant or security‑sensitive service. See PARSER_TRUST_BOUNDARY.md for the parser audit and threat model.

Development and releases

Contributor setup, the full test suite, and the local release gates are covered in CONTRIBUTING.md and RELEASE_CHECKLIST.md. In brief:

uv sync --extra dev                 # dev environment (add --extra dae for DAE validation)
uv run pytest -m "not slow"         # fast test suite
uv run ssys-recast --help

The four committed model sets (test_models1/ through test_models4/, 117 models in total) are described in TEST_MODELS.md. Releases are published to PyPI automatically by .github/workflows/publish.yml via trusted publishing; see RELEASE_CHECKLIST.md for the procedure.

Troubleshooting

• Missing RoadRunner/CVODE. ODE and assignment‑rule trajectory validation need libRoadRunner. If it is unavailable, required trajectory checks report unsupported (never a pass). Re‑sync with uv sync --extra dev.
• Missing DAE backend. DAE‑required trajectory validation needs the dae extra (uv sync --extra dev --extra dae); missing IDA/SUNDIALS is reported as unsupported.
• Parser errors. ssys-recast parses Antimony through the reference implementation and SBML/libSBML (ssys.parse_antimony_via_sbml); unsupported features (events, delays, unknown functions, malformed input) are rejected before any artifact is written.
• Numerical sampling failures. invalid_sampling_domain means the model metadata gave no usable finite positive domain; nonfinite_sample means a sampled point hit a singular surface. Reports record the seed, sampled ranges, parameters, threshold, and counterexamples needed to reproduce the failure.
• Trajectory mismatches can indicate a recasting error, a domain/parameter issue, or a model that needs symbolic rather than solver‑backed evidence. Confirm with the symbolic or numerical profile before treating it as a bug.

References

• Savageau, M. A., & Voit, E. O. (1987). Recasting nonlinear differential equations as S‑systems: a canonical nonlinear form. Mathematical Biosciences, 87(1), 83–115.
• Smith, L. P., Bergmann, F. T., Chandran, D., & Sauro, H. M. (2009). Antimony: a modular model definition language. Bioinformatics, 25(18), 2452–2454.

License and citation

MIT License — see LICENSE for the full text and the accompanying Triad National Security, LLC / U.S. Government copyright notice. Produced at Los Alamos National Laboratory under U.S. Government contract 89233218CNA000001 and released under LANL reference O5066. © 2026 Triad National Security, LLC.

Please cite ssys using the metadata in CITATION.cff.