Reuse, Compose, Extend, Standardize, Automate: Two Decades of RSEing Open (Neuro)Science at CON
Year: 2026
Type: article
Abstract: For two decades the team of the Center for Open Neuroscience (CON) has been building an open, largely domain-agnostic research-software stack — first for neuroimaging, then for neuroscience broadly, and now used well beyond. Five actions — Reuse, Compose, Extend, Standardize, and Automate — carried us here, and we expect them to keep us productive as AI agents enter the workflow. The story traces back to 2000, when Yaroslav joined the lab of a Debian developer (Barak A. Pearlmutter) as a graduate student. That immersion introduced open-source practices for reuse of existing tools while developing new neurophysiological analysis methods. Debian’s compositional pattern instilled the discipline of integrating reused components to compose powerful systems, while sharing the maintenance burden with the broader community. Those practices encouraged extending the distribution with our own packages and contributing to upstream projects we depended on daily (e.g., Fail2Ban). Joining forces with like-minded but geographically distant collaborators, notably Michael Hanke, elevated all three actions to the next level. In 2005 we formed a team to package FSL and PyEPL under the pkg-exppsy project, which became NeuroDebian, used by thousands to this day. Recognizing the potential of machine learning for neural data, and the absence of good practice in then-existing implementations, we initiated PyMVPA (2007): an early reproducible-analysis library with a full test suite, CI, and tutorials before any of that was common in scientific Python. To close the same compositional gap in scientific data management, DataLad (2013) was created to provide data versioning and distribution on top of git and git-annex. Today, datasets.datalad.org aggregates thousands of datasets from many sources. National archives we lead or participate in (DANDI, EMBER, and OpenNeuro) are built on shared standards and accessible via DataLad. registry.datalad.org then federates DataLad datasets across institutions, clouds, and projects, providing discovery of petabytes of data without any central platform. In effect, git and git-annex have become our standard for managing all digital research artifacts — code, data, containers — across these venues. Seeing the need to standardize common “data languages” in neuroscience, we co-founded or joined community efforts on BIDS, NWB, and NGFF / OME-Zarr — serving on their advisory boards and bridging to industry standards such as DICOM. Given our team size, scaling to this many projects is impossible without automation. Since the early days we practice software- and data-level unit and integration testing, performance benchmarking, CI/CD with con/tinuous archiving logs and artifacts across PRs and builds, daily testing of git-annex and DataLad extensions, auto-rebuilt ReproNim containers, automated MRI/stimuli/aux data acquisition, and auto-mirrored dandisets, among many others. The “Age of AI” doesn’t make this list optional — it amplifies every entry. AI lets teams move faster than ever while generating an explosion of divergent approaches, drifting styles, and under-reviewed output. Reuse seemingly becomes unnecessary when an agent confidently writes new code instead of finding a robust upstream; Compose appears effortless when AI generates ad-hoc adapters between misaligned components. Such practices are a recipe for fragile systems. Standardize remains the main protection against cacophonies of APIs and data formats, and Automated testing and review the only viable defense at agent-rate output. Formalizations such as our STAMPED principles — encoding what makes a research object reproducible — give both human and AI agents shared vocabulary to reason about the artifacts they produce. AI makes going it alone tempting, you can move fast, you have all the flexibility to do whatever you want, and you have fewer dependencies and supply-chain risks. But even simple code bloats over time to cover all possible cases, security requires many eyes and iterations to harden, collaboration is still more efficient in the long term. The five actions pay off most when invested in common projects, shared standards, and joint maintainership — a hundred isolated agentic projects reinventing the same infrastructure is far worse than one collaborative ecosystem. This talk walks through ongoing CON projects through those five lenses — alongside our AI adoption — to show how a small RSE team can scale productively through the “Age of AI” in service of our long-standing call: "Together we can make neuroscience a better science!"
Institutions: Dartmouth College, Center for Open Neuroscience
Language: English
Subfield: Information Systems and Management
Field: Decision Sciences
Domain: Social Sciences
Open Access status: green
Awards: Collaborative Proposal: CRCNS US-German Data Sharing Proposal: DataLad - a decentralized system for integrated discovery, management, and publication of digital objects of science, Ecosystem for Multi-modal Brain-behavior Experimentation and Research (EMBER), ReproNim: A Center for Reproducible Neuroimaging Computation, DANDI: Distributed Archives for Neurophysiology Data Integration, OpenNeuro: An open archive for analysis and sharing of BRAIN Initiative data
OpenAlex ID: W7213791784