Summary
- BioAIrepo is a new database dedicated to machine learning models for life sciences.
- This repository helps researchers to find, share, cite, and build on AI models, and get access to the datasets used to train them.
- As journals increasingly require researchers to publish and cite their AI models and training data, BioAIrepo offers a simple way to do this.
- BioAIrepo as an EMBL-EBI resource will benefit from being run next to many life sciences information resources and established practices of knowledge curation.
Researchers build and use machine learning models to gain insights into biological processes – for example, to analyse DNA sequences, predict disease risk, or map cell structures from images. However, at the moment, finding and reusing these models can be difficult. AI models are often scattered across different websites, buried in supplementary materials, or published without critical information needed for reuse.
BioAIrepo, a new collection within EMBL-EBI’s BioStudies database, aims to change this by making life science AI models FAIR – Findable, Accessible, Interoperable, and Reusable. The BioAIrepo collection provides a central home where researchers can openly access AI models trained on biological data, along with everything needed to understand and build on them. At the same time, researchers can also submit their own AI models to BioAIrepo for citation and reuse.
What is an AI model?
An AI model is a computational system trained on data to recognise patterns, make predictions, or automate complex tasks. Examples of well-known AI models trained on biological data are AlphaFold and RoseTTaFold, which can accurately predict the structure of proteins.
How can I cite AI models for publication?
To ensure research can be reproduced, publishers and journals are increasingly requiring that AI models and their training data be deposited in public, citable repositories. BioAIrepo helps researchers meet these requirements by providing a central hub – currently in a pilot phase – for sharing models and their associated datasets. This also benefits researchers because it makes their models easier to discover, reuse and improve for the community. Models are assigned unique dataset identifiers, and Digital Object Identifier (DOI) registration is offered on demand.
Researchers can submit models to BioAIrepo with support from the EMBL-EBI project team. Submissions are currently handled manually, but future development plans include self-service uploads.
What is available in BioAIrepo?
The pilot catalogue includes a small number of imaging and genomics models, as well as models related to omics data processing and protein structure building. The models will be acquired from the following sources in the first instance:
- BioImage Model Zoo – a collection of AI models to support the analysis of microscopy data, developed by the AI4Life consortium
- Kipoi – a collection of AI models for genomic data interpretation, including the prediction of chromatin accessibility, transcription factor binding, and alternative splicing from DNA sequences.
The catalogue will continue to expand over time, and researchers are encouraged to contribute.
“BioAIrepo solves a real need; it gives machine learning models a proper home,” said Anna Kreshuk, Group Leader at EMBL Heidelberg. “The BioImage Model Zoo is a collaborative platform for sharing AI models for bioimage analysis. BioAIrepo will build upon this work and serve a wider life sciences community. Sharing models in a standardised way helps researchers apply them to new data, build on each other’s work, and make scientific progress more reproducible and efficient. Improving access to models is also a step towards AI democratisation, where a wider life sciences community can benefit from AI advances and has information to steer further developments.”
Each model within BioAIrepo includes links to the model code, weights (the internal parameters a model learns during training), documentation to support the model, links to the data used to train the model, and references to any relevant publications.
“Our future plans for BioAIrepo include linking models more closely with EMBL-EBI’s existing data resources, and allowing users to test or run models directly through the platform,” said Ugis Sarkans, Team Leader at EMBL-EBI. “Overall, we want to make it straightforward to share and reuse machine learning models so that researchers can get the most out of their model-building effort and so that useful models don’t sit idle after publication.
“At the moment only a small set of models is available,” added Sarkans. “We welcome any early feedback so we can refine how we share models to best meet the needs of the scientific community.”
“BioAIrepo will help researchers find not only the documentation but also the benchmarking information they need to run AI models with confidence,” said Julio Saez-Rodriguez, Head of Research at EMBL-EBI. “We plan to align BioAIrepo with community benchmarking efforts, in particular with the recently launched BEACON (Benchmarking, Evaluation, and Assessment Consortium for Science) initiative . This will help ensure that models shared through the platform are reliable and impactful for life science research.”
Edit
