Day 2 — Community, multi-agent demos, enterprise stacks, AI panel¶

Date: 2026-05-01 Source transcript: 2026-05-01-nextflow-summit-day-2

Day 2 was lower-key on product launches and heavier on community structure (training, ambassadors, hackathons), enterprise architecture (Pfizer iFox on Lamin, Microsoft Foundry agent demo, Azure storage tiering), and a closing panel on what AI is doing to the bioinformatician role.

1. Community training update¶

Geraldine "GiGi" (Seqera Community team)

The Nextflow training portal: free, Creative Commons, GitHub-Codespaces-based, available in 11 languages. Two tracks:

Newcomers: Hello Nextflow, Nextflow Run
Experienced: nf-core training + advanced "side quests"

Heavy use of Claude (Anthropic) skills to update code snippets across courses. Phil's AI-driven translation harness diffs changed sections and re-translates only those — fixes are made in the prompt, not the file, so style and pronoun choices propagate to all 11 languages on the next run. Inspired by FastAPI docs community.

Upcoming: Nextflow Training Week (~2 weeks out).

2. Nextflow Ambassador Program¶

Marcel Ribeiro-Dantas (Seqera Community team)

History and stats: started late 2023, grew from 5 → 143 ambassadors, 16 → 40 countries, ~700 activities in 2.5 years. Rolling six-month cohorts (Jan / July intake). Incentives: travel/swag/event funding, early access to Seqera releases (Seqera AI, VS Code extension), private channel.

Top 10 Certificates of Excellence — James, Julia, Matias, Jose, Firas, Evangelos, Maxim, Cubra, Antoine, plus Fides (Tunisia).

Hybrid Hackathon (March): ~70% of local sites hosted by ambassadors.

Why it matters: The deliberate non-monetary recognition model is the program's stated load-bearing design choice (keeps it non-competitive).

3. (Skipped) — Vipava (Grail) cancelled live; recorded talk to be posted on YouTube.¶

4. Downstream RNA-seq analysis pipeline¶

Reema (Northeastern University; work begun at Novo Nordisk)

Custom DSL2 pipeline that starts from the count matrices that nf-core/rnaseq emits.

Sub-workflows: gene filtering → optional ComBat-seq batch correction → DESeq2 or limma-voom differential expression → optional FGSEA / MSigDB pathway analysis. Plus RNASeqPower for power analysis, YAML-driven config, optional API ingestion of metadata/counts.

Live ecosystem moment: Q&A surfaced overlap with nf-core/differentialabundance (already has a built-in Shiny app) — maintainers publicly invited her to merge efforts.

5. Hackathon recap¶

Marcel (returning)

Pitches nf-core hackathons as fun / collaborative / flexible (not competitive). Recap:

March hybrid hackathon — 41 projects, 29 local sites, 600 participants, 19 time zones
Boston in-person — ~30–40 participants

Highlighted projects:

Topic-channels migration — ~30–40 PRs converting nf-core modules to the versions topic
GPU variant callers in nf-core/sarek (Gary, Joe)
nf-core/cellpainting (Ken)
DuckDB integration for Nextflow — "Eider" h5ad reader (Michael, Edmond, Ben)
Harrison's AI-agent project — Antigravity orchestrating multiple agents to auto-fix ~2,000 modules
Single-cell perturbation pipeline ("Cosmic AI")

A ~30-page "host-your-own hackathon" handbook is now available.

6. Lightning talks¶

Marcus Sujanski (MDI Biological Lab, Bar Harbor ME) — sc-samap: Nextflow wrapper around the SAMap algorithm (Tarashansky / Wang, Stanford) for cross-species single-cell comparison (axolotl, turquoise killifish, zebrafish). Adds homolog detection, cross-species DE, cell-type-to-cell-type mapping.
CUNY PhD student — Two pipelines: Nextflow port of an older Galaxy-based lab pipeline + a downstream analysis pipeline that integrates outputs and produces transcript lists for phylogenetic analysis.
Unnamed — Pipeline for circular / non-canonical-splicing RNA quantification; claims top-tier speed and accuracy vs. CD3 benchmark.

7. Multi-agent workflows with Microsoft Foundry + Nextflow¶

Computational biologist at Microsoft (book on O'Reilly forthcoming)

Defines models vs. agents vs. multi-agent workflows (sequential, concurrent, group-chat, human-in-the-loop). Pitches Microsoft Foundry (announced Nov 2025):

Governed model deployment (incl. Hugging Face's ~14k models)
Agent builder, Foundry IQ (RAG over your data)
Foundry Tools including a generic MCP connector
Foundry Local — auto-quantizes models to local hardware

Demo: workflow with a "bioinformatics evaluator" agent + a "Seqera coding" agent connected to Seqera via MCP. Given a BioProject ID, the agents autonomously identified plasmodium RNA-seq, picked an nf-core pipeline, drafted and ran the Nextflow command. Cost note: ~$600 in GPT-4o tokens to run the book examples.

Why it matters: The autonomous BioProject demo is the conference's strongest "agents end-to-end" data point. The agent identified one of the speaker's own six-year-old datasets with no domain hints.

8. Pfizer's iFox — OMICs cross-functions infrastructure¶

Joe (Pfizer, Machine Learning & Computational Sciences in Medicine Design Research Unit)

iFox (Infrastructure for OMICs Cross-functions) — joint R&D + IT initiative to unify storage / compute / pipelines across Pfizer research units (oncology, immunology, vaccines, drug safety).

Stack:

AWS Batch + Seqera Platform as the backbone
Lamin / LaminDB for data lineage, ontology-aware search, artifact versioning (Pfizer's framing: "LaminDB is to data as GitHub is to code")
Lamin-integrated metadata schema tool (Rebecca White)
Seqera Studios for interactive compute
iFox reference-data ingestion pipeline (Airflow)
Evidence visual analytics platform (built with Austrian partner "David Vision")
ML model registry + custom genome-manager pipeline
Bionty ontologies

Why it matters: A relatively rare enterprise adoption story for Lamin. Recruiting note included.

9. Optimizing Nextflow cost & performance on Azure¶

Azure Storage Engineering (HPC storage products)

Covers Nextflow's five copy strategies, executor choice (Azure Batch vs. AKS via Azure CycleCloud), and storage tiering:

Azure Blob (object) for cold/output
Azure Managed Lustre / Azure NetApp Files / Azure Files NFS for hot working dirs
Single-namespace tiered storage: POSIX hot tier in front, object cold tier behind

Networking: multi-region endpoints to avoid egress; co-locate compute with storage.

Mindset shift: optimize "science per compute budget" instead of cores/petabytes purchased.

Mentions Seqera Fusion as the abstraction layer; Azure MCP server can analyze runs and suggest optimizations.

10. Closing panel — AI in bioinformatics¶

Maria Enrique (Seqera, agentic AI), Isha (Seqera scientific engagement lead), a third panelist with microbiology background, Evan moderating.

Themes discussed:

The bioinformatician's role is shifting from hands-on executor to orchestrator/architect
Validating fuzzy biological "ground truth" is the open problem
IT/governance burden of approving AI tools
Need for guardrails + human-in-the-loop + sandboxes
Ownership of AI-generated code (avoiding duplicate solutions)
Obsidian + Notion as "second brain" repositories LLMs can search
MCP servers as the integration backbone

Closing prediction: systems will beat models — long-term winners are agent harnesses with reasoning + domain context + tool/data access, not any single frontier model.

Plugs: upcoming Seqera Co-Scientist webinar, virtual Nextflow Summit in October, Boston subterranean hackathon.

Cross-cutting threads¶

AI/agents everywhere — training maintenance (Claude skills), code generation, autonomous workflow execution, hackathon project automation.
MCP as integration substrate — Seqera, Azure, Foundry connectors all converging on it.
Governance + trust — Lamin lineage, audit logs, human-in-the-loop, sandboxed adoption before enterprise rollout.
Cost-aware cloud architecture — Azure tiered storage, AWS savings plans, GPT-4o token costs all called out explicitly.
Community as the moat — non-competitive hackathons, ambassador recognition, multilingual training.

See Themes & takeaways for cross-day synthesis.