scinexus 2026.5.18

A composable app infrastructure for scientific computing

scinexus is a framework for rapid development of data processing applications. It enables interoperability between objects through defined data types, allowing development of scientific domain app ecosystems. Just as attrs and dataclasses use type hints to simplify data type definition, scinexus uses them to simplify writing best-practice scientific algorithms.

Many scientific problems require repeating calculations across many files or database records. Such tasks suit data-level parallelism, but writing robust, maintainable code for them is often tedious and quickly becomes complex.

As the Unix philosophy articulates, writing algorithms that do one thing well and can be composed together through piping data of known type is a Very Good Thing™.

scinexus encourages this design pattern and eliminates the boilerplate. We leverage the Python type annotation system to govern the compatibility (composability) of different applications. This enables in-process composition of your applications with validation of the consistency of the pipeline and the consistency of the data being run through it.

scinexus is designed for scientific reproducibility. Scientific computations should record all conditions needed to reproduce an analysis. scinexus reduces the effort by intercepting all arguments (including defaults) used in app construction and logging the resulting app state.

Examples

Developers can choose inheriting from a base class or use the scinexus.define_app decorator to make composable apps. The following examples show simple composition

Loading files so missing data does not cause a crash

from scinexus import define_app


@define_app(app_type="loader")
def read_json(path: str) -> dict:
    import json

    with open(path) as f:
        return json.load(f)


@define_app
def validate(data: dict, required_field: str) -> dict:
    if required_field not in data:
        # this becomes a NotCompleted sentinel object
        # your run doesn't crash!
        raise ValueError(f"missing {required_field!r} field")
    return data


app = read_json() + validate(required_field="name")

You can apply app to a single file path as app(filepath), or operate in parallel (and show a progress bar) on a sequence of file paths as

results = list(app.as_completed(["some_file_path.json", "some_other_file_path.json"], parallel=True, show_progress=True)

A contrived numerical example

from scinexus import define_app


@define_app
def normalise(values: list[float]) -> list[float]:
    lo, hi = min(values), max(values)
    return [(v - lo) / (hi - lo) for v in values]


@define_app
def threshold(values: list[float]) -> list[bool]:
    return [v > 0.5 for v in values]


app = normalise() + threshold()
app([1.0, 5.0, 3.0, 9.0])

A configurable app

from scinexus import define_app


@define_app(app_type="loader")
def load_csv(path: str) -> list[dict]:
    import csv

    with open(path) as f:
        return list(csv.DictReader(f))


@define_app
class summarise:
    def __init__(self, column: str) -> None:
        """column contains the values to produce summary stats for"""
        self.column = column

    def main(self, rows: list[dict]) -> dict[str, float]:
        vals = [float(r[self.column]) for r in rows]
        return {"mean": sum(vals) / len(vals), "min": min(vals), "max": max(vals)}


app = load_csv() + summarise(column="price")

Features

• Type checking at composition time
• Durable computing -- failures recorded as NotCompleted records, not exceptions
• Data-level parallel execution with pluggable backends (stdlib, loky, MPI, or custom)
• Progress bars (tqdm or rich)
• Automated logging and citation tracking
• Checkpointing via data stores (directory, SQLite)

Installation

pip install scinexus

The scinexus origin story

The app framework and utility functions in scinexus incubated inside cogent3 from March 2019, accumulating over seven years of development, testing, and real-world use in computational genomics before being extracted into a standalone package. The design is mature and has underpinned analyses in published studies.

The extraction into scinexus makes the infrastructure available to any scientific Python project, free of the cogent3 dependency. See the changelog for a detailed list of changes from the cogent3 app infrastructure.

We acknowledge here that many members of the cogent3 community contributed to the code that now lives here, including @GavinHuttley, @rmcar17, @Nick-Foto, @KatherineCaley, @fredjaya, and @khiron.