Screenase — DoE tool for IVT reaction optimization

Plan a randomized 2^k full-factorial (or central-composite follow-up) DoE screen for in-vitro transcription reactions, assign wells on a 96-/384-well plate, print a bench-ready pipetting sheet, and analyze the resulting yields — all from a single CLI, a Streamlit web app, or a Benchling-App-shaped Python subpackage.

Live demo: screenase.com — click Generate, download the bench sheet, upload a results CSV to see the Pareto. docs/screenshots/cli_demo.gif shows the generate → bench sheet → analyze loop.

Why

Scientists optimizing in-vitro transcription yields typically plan screens by hand in a spreadsheet: pick factors, pick levels, randomize runs, compute pipetting volumes per well, print, pipette, measure, eyeball the effects. Each step is manual, every step is error-prone, and the bench sheet is usually a one-off file that doesn't round-trip back into analysis.

Screenase automates the whole loop: config.yaml → design + bench sheet HTML → (bench work) → responses CSV → OLS ranked effects + Pareto plot. The design is deterministic given the config hash and seed, so an identical screen can be rerun months later without hunting for the spreadsheet.

Quickstart

pip install -e '.[dev,ui]'
screenase generate --config examples/config.yaml --out-dir out/
open out/ivt_bench_sheet.html

Edit examples/config.yaml to change factors, levels, stocks, or center-point count. The default reproduces a 4-factor IVT screen (NTPs, MgCl₂, T7, PEG8000).

After the bench run, fill in the response column in out/ivt_screen_coded.csv and:

screenase analyze my_results.csv --response yield_ug_per_uL --out-dir out/

This writes out/pareto.png + out/analysis_report.md with ranked effects and a center-point curvature test. If the curvature is significant at α = 0.05, the report suggests a central-composite follow-up with the exact CLI invocation.

Plate layout

For screens that ship to a 96- or 384-well plate, pass --plate:

screenase generate --config examples/config.yaml --out-dir out/ --plate 96 --plate-layout column-major

This writes out/plate_layout.csv + out/plate_map.png and embeds the plate-map table inside the bench sheet. Layout modes: column-major (default, fills A1..H1, then A2..H2, ...), row-major (fills A1..A12, then B1..B12, ...), or randomized (seeded shuffle, reproducible).

Central-composite follow-up

When the curvature test from screenase analyze rejects the main-effects model:

screenase generate --config examples/config.yaml --out-dir out-ccd/ --design ccd --alpha face

Emits a 2ᵏ factorial + 2k axial + N center CCD. --alpha face (default) keeps axial setpoints within low/high; --alpha rotatable uses α = (2ᵏ)^(1/4) for a rotatable design (only use this if your stocks allow the wider range).

Customizing factors

factors:
  - name: NTPs_mM_each
    low: 5
    high: 10
    unit: mM
    reagent: NTPs
    dosing: concentration
  - name: T7_uL
    low: 0.2
    high: 1.2
    unit: uL
    reagent: T7
    dosing: volume        # volume-based dosing (factor value is already µL)

dosing: concentration triggers V = C_target · V_rxn / C_stock; dosing: volume passes the factor value through as a pipetting volume.

Live demo

Try it in the browser: screenase.com (hosted on Hugging Face Spaces; source at github.com/ethanarnold/screenase). The hosted app wraps the same screenase package — two tabs:

• Generate screen: slider-edit factors, get CSV + rendered bench sheet + download buttons.
• Analyze results: upload the coded CSV with your yields filled in, get the Pareto + ranked-effects table.

Worked examples

Pre-rendered artifacts (bench-sheet PDF, plate map, diagnostic plots, narrated analysis report, CCD follow-up, Benchling schema JSON) live in docs/examples/ — browse the look of a real run without cloning. See docs/examples/README.md for the regeneration commands.

Benchling integration

src/screenase/benchling/ shapes Screenase as a Benchling App: a manifest, SDK-style webhook handlers, and entity mapping that converts the design / results / inventory consumption into Benchling Request / Result / Entry / Container-decrement payload shapes.

Handlers (runnable locally against fixtures, no live tenant needed):

• handle_request_created — build the design + bench sheet, return the Request payload.
• handle_results_submitted — run OLS on submitted responses, return the Entry payload.
• handle_reagent_consumed — sum reagent pipetting, return a container-decrement payload.

Or from the CLI:

screenase generate --config examples/config.yaml --out-dir out/ \
    --export benchling --export benchling-inventory --lot-refs lot_refs.json

• What runs locally: webhook handlers against fixture payloads (fixtures/*.json). python -c "from screenase.benchling.app import run_fixture; run_fixture('...')".
• What doesn't: the live deployment. Benchling's Developer Platform is enterprise-gated; the shape is present, the tenant is the gap.

See docs/benchling_mapping.md for the full semantics and src/screenase/benchling/README.md for the subpackage structure.

Design rationale

Why 2⁴ + 3 center points? Why (f1 + f2 + f3 + f4)**2? Why randomize? See docs/design_rationale.md.

Dev

pip install -e '.[dev,ui]'
pytest -q
ruff check src tests streamlit_app.py
mypy --ignore-missing-imports src/screenase
streamlit run streamlit_app.py

Python ≥ 3.11, CI covers 3.11 and 3.12 on Ubuntu.

Author

Ethan Arnold — emailtoethan@gmail.com. MIT-licensed.