.   °                .......                                     °   
       | .° |                :.   .:                                   (  .° )
       | °  |                .'   '.                                    |°  |
       |° . |                |     |                                   /°  ° \
      (______)               `_____'                                  (_______) 
▌▌▌▌▌▌▌▌▌▌▌▌▌▌▌▌▌▌▌▌▌▌▌▌▌▌▌▌▌▌▌▌▌▌▌▌▌▌▌▌▌▌▌▌▌▌▌▌▌▌▌▌▌▌▌▌▌▌▌▌▌▌▌▌▌▌▌▌▌▌▌▌▌▌▌▌▌▌▌▌▌▌▌▌▌▌▌▌▌▌▌▌▌▌▌▌▌▌▌
  ▄▄▄▄▄▄▄▄▄   ▄▄▄▄▄▄▄▄▄▄  ▄▄        ▄  ▄▄▄▄▄▄▄▄▄▄  ▄        ▄  ▄▄▄▄▄▄▄▄▄▄  ▄▄▄▄▄▄▄▄▄▄  ▄▄▄▄▄▄▄▄▄▄ 
 ▐░░░░░░░░░▌ ▐░░░░░░░░░░▌▐░░▌      ▐░▌▐░░░░░░░░░░▌▐░▌      ▐░▌▐░░░░░░░░░░▌▐░░░░░░░░░░▌▐░░░░░░░░░░▌
 ▐░█▀▀▀▀▀▀█░▌▐░█▀▀▀▀▀▀▀▀ ▐░▌░▌     ▐░▌▐░█▀▀▀▀▀▀▀▀ ▐░▌      ▐░▌ ▀▀▀▀█░█▀▀▀ ▐░█▀▀▀▀▀▀█░▌▐░█▀▀▀▀▀▀█░▌
 ▐░█▄▄▄▄▄▄█░▌▐░█▄▄▄▄▄▄▄▄ ▐░▌ ▐░▌   ▐░▌▐░▌         ▐░█▄▄▄▄▄▄█░▌     ▐░▌    ▐░▌      ▐░▌▐░█▄▄▄▄▄▄█░▌
 ▐░░░░░░░░░▌ ▐░░░░░░░░░░▌▐░▌  ▐░▌  ▐░▌▐░▌         ▐░░░░░░░░░░▌     ▐░▌    ▐░▌      ▐░▌▐░░░░░░░░░░▌
 ▐░█▀▀▀▀▀▀█░▌▐░█▀▀▀▀▀▀▀▀ ▐░▌   ▐░▌ ▐░▌▐░▌         ▐░█▀▀▀▀▀▀█░▌     ▐░▌    ▐░▌      ▐░▌▐░█▀▀▀▀▀▀▀▀ 
 ▐░█▄▄▄▄▄▄█░▌▐░█▄▄▄▄▄▄▄▄ ▐░▌     ▐░▐░▌▐░█▄▄▄▄▄▄▄▄ ▐░▌      ▐░▌     ▐░▌    ▐░█▄▄▄▄▄▄█░▌▐░▌          
 ▐░░░░░░░░░▌ ▐░░░░░░░░░░▌▐░▌      ▐░░▌▐░░░░░░░░░░▌▐░▌      ▐░▌     ▐░▌    ▐░░░░░░░░░░▌▐░▌          
  ▀▀▀▀▀▀▀▀▀   ▀▀▀▀▀▀▀▀▀▀  ▀        ▀▀  ▀▀▀▀▀▀▀▀▀▀  ▀        ▀       ▀      ▀▀▀▀▀▀▀▀▀▀  ▀           

Benchtop

Parallel in-silico experiments for biological models

Benchtop runs simulation benchmarks defined in a PEtab-inspired format. Given a YAML configuration and companion TSV/SBML files, it schedules conditions in dependency order, simulates across multiple CPU cores, caches intermediate results, and computes observables for model–data comparison.

Documentation

Full documentation: https://benchtop.readthedocs.io/en/latest/

Sphinx sources live in docs/. To build locally after pip install -e ".[docs]":

python -m sphinx -b html docs docs/_build/html

Features

• PEtab-style benchmarks — Define experiments with YAML plus standard TSV files for conditions, measurements, observables, and parameters.
• Parallel execution — Distributes simulation tasks across processes using a round-robin worker pool.
• Preequilibration support — Topologically sorts conditions so preequilibration runs before dependent simulations.
• Single-cell replication — Run multiple stochastic replicates per condition via cell_count.
• Result caching — Persists per-simulation pickle files and a cache index; supports resuming incomplete runs.
• Observable calculation — Evaluates observable formulas from simulation trajectories and aligns them with experimental measurements.
• Pluggable simulators — Built-in Tellurium wrapper; optional AMICI support via extras.

How it works

flowchart LR
    YAML[Benchmark YAML] --> Loader[FileLoader]
  Loader --> Exp[Experiment]
  Exp --> Org[Organizer]
  Org -->|topological task order| Pool[Worker pool]
  Pool --> Sim[Simulator wrapper]
  Sim --> Cache[ResultCache]
  Cache --> Obs[ObservableCalculator]
  Obs --> Results[results/*.pkl]

1. FileLoader reads the benchmark YAML and loads PEtab TSV/SBML files into memory.
2. Organizer builds a dependency-aware task list (conditions × cell replicates) and assigns work to workers.
3. Workers load the SBML model, apply condition parameters, run preequilibration when needed, and simulate.
4. ResultCache stores raw trajectories as pickle files alongside a JSON index.
5. ObservableCalculator downsamples trajectories, evaluates observable formulas, and packages results next to experimental data.

Requirements

• Python ≥ 3.12
• See pyproject.toml for pinned dependencies

Installation

Benchtop is registered on PyPI and pip installable:

pip install benchtop

To build from source,navigate to the repository root:

pip install -e .

For AMICI support (optional):

pip install -e ".[amici]"

This installs the benchtop CLI entry point.

Quick start

Command line

Run a single benchmark:

benchtop experiment -p src/tests/data/LR-benchmark.yaml -s tellurium -c 4

Run all benchmarks in a directory:

benchtop experiment --run_all path/to/benchmarks/ -s tellurium

Useful flags:

Flag	Description
-p, --path	Path to benchmark YAML
-s, --simulator	Simulator backend (tellurium or amici)
-c, --cores	Number of parallel processes (default: CPU count)
--cache_dir	Directory for simulation cache (default: ./.cache)
--load_index	Resume from an existing cache index
--No_Observables	Skip observable calculation after simulation
-v, --verbose	Enable debug logging

Python API

from benchtop.experiment import Experiment

experiment = Experiment(
    "src/tests/data/LR-benchmark.yaml",
    cores=4,
    cache_dir="./.cache",
    verbose=True,
)

experiment.run("tellurium")
experiment.calculate_observables()

You can also pass a custom simulator class that extends AbstractSimulator instead of a registry name.

A worked example is available in demo/in-silico-experiment-demo.ipynb.

Benchmark configuration

A benchmark is a directory containing a YAML file that references PEtab-style data files. Example (src/tests/data/LR-benchmark.yaml):

format_version: 1
parameter_file: LR Model - Parameters.tsv
problems:
  - name: test-benchmark
    condition_files:
      - conditions.tsv
    measurement_files:
      - measurements.tsv
    observable_files:
      - observables.tsv
    sbml_files:
      - LR-model.xml
    cell_count: 3

Note: The cell_count attribute is only applicable to Benchtop and specifies the number of replicates to simulate per-condition.

Data files

File	Purpose
Parameter TSV	Model parameters (parameterId, parameterScale, nominalValue, …)
conditions.tsv	Simulation conditions — maps conditionId to parameter overrides
measurements.tsv	Links observables to conditions, optional preequilibration, and measurement timepoints
observables.tsv	Observable definitions (observableId, observableFormula, noiseFormula)
SBML	One or more model files (.xml or .sbml)

When preequilibrationConditionId is set in measurements.tsv, Benchtop runs the preequilibration condition first and uses its final state as the initial condition for the simulation.

Results

After observable calculation, results are written as a pickle file in a results/ subdirectory next to the benchmark YAML (named after the problem or today's date). Each entry in the results dictionary contains:

• conditionId and cell replicate index
• Per-observable experiment, simulation, and time arrays

Simulators

Benchtop uses a small registry of simulator wrappers in src/wrappers/. Each wrapper implements AbstractSimulator with load, modify, and simulate methods.

Name	Class	Install
tellurium	TelluriumSimulator	Included by default
amici	AmiciSimulator	pip install -e ".[amici]"

To add a new backend, create a wrapper under src/wrappers/ and register it in src/benchtop/registry.py.

Project structure

Benchtop/
├── demo/                          # Jupyter notebook walkthrough
├── src/
│   ├── benchtop/                  # Core library
│   │   ├── experiment.py          # Main experiment orchestrator
│   │   ├── _organizer.py          # Task scheduling and topological sort
│   │   ├── _worker.py             # Per-simulation worker logic
│   │   ├── _record.py             # Results dictionary management
│   │   ├── _results_cacher.py     # Pickle cache and index
│   │   ├── _observable_calculator.py
│   │   ├── _abstract_simulator.py # Simulator interface
│   │   ├── file_loader.py         # YAML and TSV loading
│   │   ├── registry.py            # Simulator lookup
│   │   ├── launcher.py            # CLI entry point
│   │   └── arguments.py           # CLI argument parsing
│   ├── wrappers/                  # Simulator backends
│   └── tests/                     # Test suite and example data
├── pyproject.toml
└── README.md

Running tests

From the repository root:

cd src/tests
python run_tests.py

The test suite uses the LR receptor benchmark in src/tests/data/ and exercises experiment execution, caching, workers, organizers, records, and observable calculation.

Authors

• Jonah Huggins — JonahRileyHuggins@gmail.com
• Marc Birtwistle — marc.birtwistle@gmail.com

License

See LICENSE.