Best Practices for Reproducible AI Research in Life Sciences – Introducing DOME

In recent years, there has been exponential growth in the use of machine learning (ML) in the life sciences, fueled by increased access to powerful hardware (e.g., GPUs) and high-quality FAIR biodata enabled by the open science movement. One standout example is AlphaFold, which uses ML to predict protein structure from genetic sequences, a breakthrough that contributed to the 2024 Nobel Prize in Chemistry. Other impactful ML applications include drug discovery (e.g., supporting drug design, screening, and chemical synthesis) and bioimaging (e.g., improving diagnostic accuracy, such as in cancer detection, as seen in large EU-funded initiatives like AI4Life).

Despite this momentum, ML methods are often published with insufficient detail, making them difficult to reproduce, interpret, or reuse. This reproducibility crisis stems from two key challenges. At the researcher level, applying ML effectively requires cross-domain expertise, yet there are no unified pathways or harmonized standards to follow. At the journal level, there is often a lack of requirements enforcing methodological transparency, and peer reviewers may be subject-matter experts but not ML specialists, leading to the publication of poorly described methods.

To address this problem, ELIXIR, a European life sciences infrastructure involving over 250 research institutes across 21 member countries, established the ELIXIR Machine Learning Focus Group in October 2019. This group published the DOME recommendations in a 2021 Nature Methods article: “DOME: recommendations for supervised machine learning validation in biology.” These guidelines focus on best practices for publishing supervised ML methods in life sciences.

What is DOME-ML?

DOME-ML (or simply DOME) is an acronym standing for Data, Optimization, Model and Evaluation in Machine Learning. It is a set of community-developed guidelines and checklists spanning these four pillars of the ML lifecycle. The recommendations are formulated as questions for researchers implementing ML methods. Ideally, authors should disclose relevant information for all four areas in the supplementary materials of a publication.

What is the scope of the recommendations?

Each of the four pillars is outlined in Figure 1. In the Data section, researchers should clearly disclose how datasets were split into training, validation, and test sets to prevent data leakage. For Optimization, the algorithm, such as Convolutional Neural Networks (CNNs) or Support Vector Machines (SVMs), and its hyperparameters should be described, with links to code or software to support reuse and validity check. For Model, details about the architecture and availability of pretrained weights should be provided. Under the Evaluation, researchers should compare their model’s performance with prior approaches using the same or similar datasets and report results transparently.

Figure 1: DOME Overview (Reference: Gavin Farrell’s presentation at the RDA AI in Action Workshop, Feb 13, 2025)

The DOME Registry

To promote adoption, the DOME framework is supported by the DOME registry (Figure 2), where researchers can register their ML methods. Each registered entry receives a unique DOME ID and a compliance score (see Figure 3), which can be cited alongside publications. This promotes transparency and builds trust in published ML methods

Figure 2: Dome Registry Homepage Screenshot on 5/20/2025 (DOME registry homepage URL: https://registry.dome-ml.org/intro)

Figure 3: Dome Registry Example showing the compliance DOME Score (Reference: Gavin Farrell’s presentation at the RDA AI in Action Workshop, Feb 13, 2025)

Incentives to Use the DOME Registry

As described in Figure 4, the GigaScience journal adopted the DOME Registry in 2024, requiring ML methods papers to include a registry entry before publication. More journals are expected to follow. In addition, the registry is federated with APICURON, enabling automatic updates to researchers’ ORCID profiles as a form of credit (Figure 4).

Figure 4: Incentives to use the DOME Registry (Reference: Gavin Farrell’s presentation at the RDA AI in Action Workshop, Feb 13, 2025)

Future Directions

The DOME Registry continues to evolve with new features (see Figure 5), including integration with emerging ML metadata standards such as CROISSANT metadata format and FAIR4ML Metadata Schema for ML models. The latter is a newly released framework that applies FAIR principles to ML models and will be covered in a future blog post.

Figure 5: Future of the DOME Registry (Reference: Gavin Farrell’s presentation at the RDA AI in Action Workshop, Feb 13, 2025)

For those preparing NIH Data Management and Sharing Plans (DMSPs) involving ML and AI models, I recommend referencing the CROISSANT metadata format and FAIR4ML metadata schema in Element 1C (Metadata, other relevant data, and associated documentation) and citing DOME in Element 3 (Standards). These resources support transparent and reproducible ML development, an essential step for trustworthy and reproducible AI in life sciences.

I highly recommend visiting the Research Data Alliance (RDA) AI in Action page and reviewing the AI in Action Global Workshops Report to explore how these principles are shaping the future of AI in research.

Resources