ISMB 2026: BioinfoCoreWorkshop
Session Overview
The bioinfo-core COSI session was scheduled Tuesday, July 14th at the Washington Hilton in the Cabinet room.
The bioinformatics core session is organized by staff and managers of Core Facilities and designed for all members of bioinformatics core facilities or people in related roles.
Main organizers for this year:
- Madelaine Gogol, Stowers Institute, United States
- Alberto Riva, Human Technopole, Italy
- Awtum Brashear, Shriners Hospitals for Children, United States
- Yuvanesh Vedaraju, Houston Methodist Hospital, United States
- Jim Denvir, Marshall University, United States
- Siva Karunanithi, IMB, Germany
- Syed Murtuza Baker, University of Manchester, United Kingdom
Notes from the session
Talks: Stuart Levine - The Payoff of FAIR: From Natural Language Queries to GEO Submissions and Nextflow Samplesheets in a Bioinformatics Core. Data management in some ways is a good fit for something the bioinformatics core should be responsible for. They have developed a wrapper around the SEEK data management project called NExtSEEK - essentially containing samples as spreadsheets and protocols as word docs. Assays connect samples. The metadata is what’s really important to have in the system, the data lives in a lake somewhere and gets pointed to wherever it is. They have added Nessie, an assistant to do easy queries and more complex tasks (GEO submission). They want to add more complex jobs to it eventually (data ingestion).
Nathan Kennedy - An R Package To Simplify The UK Biobank RStudio Experience After an entertaining introduction to the pain users face when using the UK Biobank Research Analysis Platform (a bespoke cloud based computing infrastructure), we were shown the very approachable R package allowing the user to do things like create projects, install packages, download data, and download a project for later use. http://github.com/natedog0027/UKBEasieR
Patricia Carvajal-López - Scaling up support for bioinformatics core facilities training While the previous training course from EMBL-EBI will not be available in 2026, they will be doing versions of the course in the LATAM (ISCB-Latin America) conference in Peru and one hosted by the African Bioinformatics Institute. Previous training materials are also available in the EBI platform and other options for sharing and hosting training materials were discussed, as well as the previous work on the ISCB competency framework for bioinformatics core facility scientists.
Sabiq Chaudhary - A Hybrid-Cloud Nextflow NF-Core Compliant Framework for High-Throughput Bioinformatics Pipeline for Roche SBX Sequencing This talk covered Roche’s transformation of a legacy pipeline into a modular architecture aligned with nf-core and capable of handing diverse assay types over tons of data across multiple environments. They highlighted the internal interface for flexible configuration plus the monitoring and testing that keep it reliable.
Thomas C. Smits - HuBMAP Workspaces: an integrated analysis environment for the Human BioMolecular Atlas Program Data Portal HuBMAP (Human BioMolecular Atlas Program) is a data repository for multi-modal spatial and single-cell from healthy human tissues with > 5k data sets. They have added embedded workspaces to allow users to run Jupyter notebooks (e.g. running both R and python) on the data directly within the system. They also include analysis templates that enable users to get started with different types of analysis easily. http://portal.hubmapconsortium.org/workspaces
Muruj Tukruni - Telomere-to-Telomere Genome Assembly of Native Desert-Adapted Sheep Ovis aries Breed “Najdi” From Saudi Arabia There is an indigenous sheep adapted to extreme arid conditions that had no genome despite its high economic value as an agricultural animal and interesting adaptations. Using pacbio, nanopore, and Hi-C data, they were able to assemble two (male and female) genomes to a T2T level with excellent completeness. They have put their pipeline on github.
Panels: Agents in the core: How your core is using agentic AI (or how it should be) Panelists: Stuart Levine, Philip Freda, Natalie Gill, Yury Bukhman Leaders closer to the cutting edge of AI use answered various questions from the audience about the use of agentic (and AI in general) in cores, covering topics from training to cost to pitfalls and reproducibility.
In with the new: Case studies on how cores bring in a new data type or technology Panelists: Reuben Thomas, Ryan Dale, Lara Ianov, Hua Li People from several different cores spoke about bringing in new technologies: Xenium, single cell, and long reads. They dealt with everything from how to find the resources or time to do so, what the process has looked like for their core (in good and bad situations), decisions they had to make around data retention, addressing the delay to other projects. It seems in most cases cores are treating this as an on demand task but occasionally have the ability to take the time to really explore and bring something in after more robust examination. People expressed that AI will likely mean getting less of the routine work in the future. Will “Second Eye for AI” be a service provided by cores (check my AI code)? Perhaps cores will be creating agents, skills, and MCPs for end users to do their own analysis.
Breakout groups:
We broke into two breakout groups based on what the people in the room were interested in discussing.
Agentic AI: We longed for an AI user with less permissions or other strategies for keeping an agent isolated from stuff. /sandbox and singularity were mentioned as possibilities. As far as training users, we have processes to train people on explosive chemicals, so why not teach them and make them take a test at the end, they don’t get to use it until they at least have been made aware of some risks. Agents were discussed as developing an agent for a particular well defined task - i.e. grant review trained on old grants. Telling Claude to use “Superpowers” makes it do a lot more software engineering type stuff but will take longer and use more tokens. Performance declines with token usage and permission issues might also start cropping up, so besides just using money you want to be judicious about how you’re burning tokens. Different models do sometimes make correlated errors since they were trained on the same corpus essentially. We’re going to post some AI fluency workshop materials to our mailing list and slack.
New Tech: They had a diverse group in terms of industry / academia / different roles. They discussed pricing models for new technology, deciding when to use something new versus something more established, and got into some agricultural genomics where a genotype array can tell you quite accurately how much milk a given cow was going to produce!
