Skip to content

4. Open and Sustainable AI (OSAI)

DOME asks how you reported a method. OSAI asks how you share, reproduce and run it.

Overview

Questions:

  • What does OSAI cover that DOME does not?
  • What are the nine OSAI recommendations?
  • How do I go from a recommendation to something I can actually do on Monday?

Learning Objectives

By the end of this chapter, you will be able to:

  1. Position DOME and OSAI as complementary rather than competing
  2. Recall the nine OSAI recommendations
  3. Use the OSAI ecosystem components list to find a concrete tool for a specific gap

Time: 10 minutes

Where OSAI fits

DOME governs the reporting of a supervised ML method in a publication. It does an excellent job of that and deliberately stops there.

But a well-reported method can still be unusable: no packaged environment, no standardised metadata, not deposited anywhere findable, no benchmark others can run against, and no idea what it cost to train. Open and Sustainable AI (OSAI) addresses that surrounding layer — how AI models and datasets are shared, reproduced, evaluated and run responsibly 1.

The two are complementary, and both are in scope for this tutorial:

DOME OSAI
Unit A supervised ML method in a paper An AI model or dataset as a research object
Question Was it reported transparently? Can it be found, reused, reproduced and run sustainably?
Output A structured 21-field transparency report Practices mapped to ecosystem components
Scope Publication Whole model lifecycle, including compute cost

A useful mental model

DOME makes the paper assessable. OSAI makes the model usable.

The nine OSAI recommendations

The OSAI work maps nine practical recommendations to over 300 components of the AI ecosystem, so that each recommendation comes with real tools rather than good intentions 1.

R1 — Generate and share standardised AI metadata for models and datasets

Use community schemas — FAIR4ML for models, Croissant for datasets — so your work can be found and interpreted by machines as well as people.

R2 — Leverage AI registries as central hubs for sharing and discovering reusable AI models and datasets

Deposit in a registry rather than only a personal repository. The DOME Registry is one such hub; there are many others in the ecosystem list.

R3 — Host, promote and share training and guidance on correct deposition of AI research objects for reuse

Deposition practice has to be taught. This tutorial is itself an instance of R3.

R4 — Ensure transparent disclosure, clear documentation and sharing of all model relevant information

The closest overlap with DOME — disclose data, training configuration, model files and evaluation so others can independently verify your results.

R5 — Provide portable code and reproducible environments to facilitate smooth reuse

Containers, lockfiles, declared dependencies. “It works on my machine” is not a reproducibility claim.

R6 — Use standardised AI-ready datasets and benchmarking evaluation protocols to facilitate reproducible model comparisons

Shared benchmarks and protocols are what make two papers’ numbers comparable at all.

R7 — Implement Green AI model development techniques

Efficient architectures, transfer learning over training from scratch, early stopping, right-sized models — much of this improves science and reduces cost.

R8 — Choose and optimise hardware to reduce environmental impact

Hardware selection and utilisation are a first-order lever on the footprint of a training run.

R9 — Measure and report AI model environmental impact

You cannot manage what you do not measure. Report the compute and energy cost of training alongside your performance numbers.

Finding a tool for your gap

The OSAI ecosystem components list is a community-curated inventory of AI-relevant tools, registries, metadata standards and best-practice frameworks, each mapped to the OSAI recommendations it supports 2.

It is also explicitly designed for reuse — the maintainers acknowledge there is no single perfect set of AI best-practice recommendations, so the list is built to be remapped onto other frameworks and other communities’ guidance.

Sources and further reading

Next: Chapter 5 — Agentic AI Skills for DOME

  1. Gavin Farrell, Eleni Adamidi, Rafael Andrade Buono, Mihail Anton, Omar Abdelghani Attafi, Salvador Capella Gutierrez, Emidio Capriotti, Leyla Jael Castro, Davide Cirillo, Fotis Psomopoulos, Silvio C. E. Tosatto, and others. Open and sustainable AI: challenges, opportunities and the road ahead in the life sciences. Nature Methods, 2026. doi:10.1038/s41592-026-03037-6. ↩↩↩

  2. OSAI Ecosystem Contributors and ELIXIR Machine Learning Focus Group. OSAI: open and sustainable AI ecosystem components list. 2025. doi:10.5281/zenodo.15391274. ↩↩

  3. Gavin Farrell, Eleni Adamidi, Rafael Andrade Buono, and others. Open and sustainable AI: challenges, opportunities and the road ahead in the life sciences. arXiv preprint, 2025. doi:10.48550/arXiv.2505.16619. ↩