Public media-optimisation datasets, one schema

Eight published cell-culture media studies catalogued to a common schema with a dataset card each, a licence register with the URL every claim was read from, and a one-command loader.

What it is

A catalogue repository: one folder per study, a dataset card written from the paper and checked against the files, a loader to a shared schema, and a licence register that records where each licence claim was read

Why it exists

The replay project needed public campaigns with known provenance. Finding them meant reading eight papers’ supplements; I wrote the reading down so that nobody has to do it twice

Studies

He 2021 (CHO, high-throughput screens); Grzesik 2021 (primary T cells, definitive screening design, four donors); Narayanan 2025 (PBMC and K. phaffii); Ozawa 2025 (HeLa-S3); Hashizume and Ying 2025 (CHO-K1 reformulation, 57 components); Gangwar 2024 (CHO, metal ions and charge variants); Borkowski 2020 (E. coli cell-free); Mattert 2026 (primary NK cells)

Tools

Python, pandas; python3 load.py <study> returns a dataframe; a manifest with DOI, licence, landing page and status per study

Status

Eight cards. Six studies mirrored with download URL, size and checksum recorded per file; two carried as fetch instructions only, because their licences forbid redistribution (Ozawa 2025: unlicensed repository, CC BY-NC-ND article; Hashizume and Ying 2025: CC BY-NC-ND). Public snapshot, no history, 20 September 2026

Repo

github.com/ellenmarsh1612/media-opt-datasets

Eight published cell-culture media studies brought to one schema, each with a dataset card and a recorded licence, so that a model fitted to any of them starts from decisions that can be inspected.

Why I built it

Most published comparisons in this field rest on a loader someone wrote in an afternoon. A loader embodies decisions: which readout is the target, what happens to wells that failed, whether replicate variance is carried or zeroed, whether a design code is a concentration. Those decisions move results as much as the choice of model does, and they are usually invisible. A card per study makes them visible and checkable.

It is the same discipline I used for a published data-driven review in Advanced Healthcare Materials, which brought more than 100 wound-healing studies into one structured dataset before any model was fitted to it. The schema work is slower than the modelling and it decides whether the modelling can be believed.

What each card records

• Cell line and readout, as the paper’s methods state them and not as the file name says. Ozawa (2025) is HeLa-S3; its folder inherited a CHO label from upstream and the card corrects it.
• Design: number of formulations, components, units (design code or molar), rounds or generations, and which rows are one-factor-at-a-time versus multi-component.
• Two kinds of statement, kept apart: counts, column names and ranges verified from the mirrored file, and figures taken from the paper before the files arrived, which the card marks as a specification to check rather than a fact.
• Replicate structure: per-well replicates, per-row standard deviations, or neither, and what the paper’s own supplementary loader does with them.
• Licence of the data files, recorded separately from the article’s licence, with the URL it was read from. A CSV bundled into an MIT-licensed repository does not become MIT.
• The loading rules applied, in a decisions section: failures kept at zero, replicate variance carried where it exists, assay noise not pooled with batch drift, no log transform where skew worsens, one-factor and multi-component blocks kept separate. Entries the paper leaves undecided are marked to complete rather than inferred.
• Absences. No study in the collection ships a component cost vector, so any pound-denominated claim on these data is an external assumption and is labelled as one.

Limitations

• Two of eight studies cannot be mirrored; their fetch instructions depend on publisher URLs that may move.
• Cards are written from the papers and supplements; where a paper is silent on units or replicate handling, the card says so rather than guessing.
• The schema is built for growth-response media data. Other assay types will need it extended.
• Licence readings are the project’s own, recorded with their sources; they are not legal advice.

Artefacts

• Repository, with a README that gets a stranger from clone to a loaded dataframe in one command, an attribution file and the licence register at docs/licences.md.
• Dataset cards, one per study, in each study folder.
• The project that uses two of these studies: Model-guided experiment loops for screening campaigns.