Skip to content

Day 1 AM — Keynote, AI rewrites, Nextflow 26.04

Date: 2026-04-30 (morning) Source transcript: 2026-04-30-nextflow-summit-day-1-am

The opening session framed Nextflow as the substrate for "AI-era science": agentic pipeline authoring, machine-readable modules, and a closed-loop optimization layer above the workflow engine. Anchored by Evan Floden's keynote, then a string of practitioner talks that show what AI-driven pipeline development actually looks like in 2026.

1. Opening keynote — The AI era of science

Evan Floden (CEO/co-founder, Seqera)

State-of-the-union framed around three pillars: Context (data + lineage), Execution (compute), Co-Scientist (agents). Opens with a tribute to Craig Venter. Announces a long slate of platform features and positions Nextflow as the integration layer for agent-driven, closed-loop science.

What was launched:

  • Benchling AI ↔ Seqera MCP — launch RNA-seq runs from a Benchling chat
  • Data lineage as a first-class entity, surfaced in the Data Explorer UI
  • In-browser IGV for BAMs/BEDs/VCFs straight from cloud buckets
  • Studios SSH — connect via VS Code Remote SSH or Cursor
  • Seqera GxP packages — validation docs aligned to FDA/EMA, GLP/GCP/GMP
  • IDP-to-Teams mapping (Okta, Entra), audit logs, fine-grained authz
  • Community module registry + nextflow module run (any module standalone)
  • GPU metrics first-class (NVIDIA partnership; Sarek GPU acceleration)
  • Seqera Intelligent Compute — new scheduler replacing the AWS Batch model: spot optimization, right-sizing, bin-packing. Live demo: $5/sample → $2/sample.
  • Co-Scientist — three modes (assist / build / automate), model-agnostic (BYO key)
  • Seqera for Sciences (Launchpad) — self-serve scientist UI with "Explain failure" debugging
  • Event-driven pipeline chaining — bucket events trigger downstream pipelines
  • Closed-loop optimization layer — autonomous hill-climbing across params, tools, even model architectures (RNA-seq quant + model-training pipeline demos)

Why it matters: Floden is explicit that the bet is on context + agents + a re-architected scheduler. Modules become agent-callable tools; lineage becomes the audit trail; Intelligent Compute admits AWS Batch was the wrong primitive.

2. Building a pipeline factory with AI

Lorena (clinical-hospital postdoc, Boston → returning to Barcelona)

Personal talk on doing everything (sysadmin / dev / analyst) at a clinical lab. Ran the same pipeline-build task three ways: Claude Code, Grok + skills, Seqera AI. Seqera AI won — 5 min plan / 4 min build — primarily on context, not model quality.

Key concepts: Claude Code skills, sub-agents, XML prompting, the Docker registry as agent context, "context is king."

Why it matters: The headline lesson is that the pipeline factory is built from skills and registries, not from a smarter model. Also doubles as a job-search announcement.

3. Rust-QC — AI-driven rewrites of legacy bioinformatics

Phil Ewels (co-founder MultiQC; nf-core veteran)

Origin story of Rust-QC: a Rust rewrite of the nf-core/rnaseq QC sub-workflow, generated in a Slack experiment with Phil + Seqera AI. Consolidates 16 bioinformatics commands into one tool — file reads drop from ~18× to 1×, disk usage from 2.5 TB to <0.1 TB. Verification target: nf-test snapshots (deterministic, byte-comparable). Cited cost: ~$10K in Anthropic API spend.

The reusable pattern: emulate-then-rewrite. Don't ask AI to invent a novel tool; ask it to reproduce a verifiable target. nf-test snapshot diffs are the agent harness. Phil predicts a "gold rush" of AI-driven legacy rewrites.

Why it matters: The first concrete recipe at the summit for getting trustworthy AI-generated tools — and the cost transparency ($10K) is rare and useful.

4. What's new in Nextflow 26.04

Paolo Di Tommaso (co-founder / chief architect, Seqera)

Two themes: language and engine.

Language: - New Syntax Parser v2 — own grammar, no Groovy dependency - Stronger static type system - Typed params block, typed channels, typed process inputs/outputs

Engine: - Nextflow Modules as first-class registry (evolution of plug-ins) - nextflow model CLI — search / view / run / create / publish - Semantic search over modules - Wave-built containers on the fly - Intelligent Compute in preview on AWS — comparison shown: 230 vs ~280 CPU-h, $15 vs $34

Why it matters: Modules — not Docker containers — are pitched as the right packaging primitive, and the typed-everything direction makes pipelines machine-readable for agents.

5. Nextflow for proactive vaccinology

Suriya Ghose (Imperial College London; Kelly / Barclay group)

Building prodvac-seq for proactive (rather than reactive) pandemic preparedness. Replaces single-virus reporter assays with multiplexed barcoded pseudovirus NGS panels (Jesse Bloom-lab paradigm). Pipeline does barcode counting → neutralization curves / IC50 → deep mutational scanning across COVID and flu variants. Validated against Bloom-lab data; collaborators at Cambridge and the Crick.

Why it matters: A subfield going NGS-native. The reporter-assay → barcoded-pseudovirus shift is the kind of methodological pivot Nextflow is well-positioned to standardize.

6. MGTree — microbial genotyping with Nextflow

Scott Norton (Senior Bioinformatics Scientist, Merck, Cambridge MA)

MGTree does viral genotyping (norovirus, HPV) where Kraken2's k-mer approach fails on closely-related references. Pipeline: Bowtie2 with -a (all alignments) → custom dynamic alignment-score filter → reads placed on a Muscle-aligned phylogenetic tree → custom MultiQC report. Validated against PCR/Sanger; in one case caught a co-infection that Sanger missed.

Caveat: Doesn't handle recombination.

Why it matters: A clean example of where the off-the-shelf tool (Kraken2) is the wrong abstraction, and the Nextflow community's MultiQC tooling makes a custom replacement shippable.

7. nf-spatial — reproducible Xenium preprocessing

Lara Ianov (Co-Director, Biological Data Science Core, UAB)

An nf-core-adjacent pipeline for 10x Xenium spatial transcriptomics. Defines a "Tier 1" automated workflow: annotation → multi-stage QC (raw / filtered / normalized / clustered) → cell-shape classification (circular / non-circular / polygonal) as a segmentation-quality proxy → area-based normalization → PCA + spatial clustering with parameter sweeps → consolidated R / RDS outputs and report (UMAP videos, violin plots).

Out of scope: Segmentation (use nf-core/spatialvi), cell-type label transfer (bias concerns), CosMx.

Runtime: 300–500-panel Xenium = ~day for dozens of samples; 5K panel is much heavier.

Why it matters: Core-facility scaling problem solved by treating spatial preprocessing as a standardized pipeline rather than per-project R notebooks.

Cross-cutting threads

  • Context > model — repeated by Floden, Lorena, and Ewels: skills, lineage, and registries matter more than the LLM choice.
  • Modules as the packaging primitive — registry, semantic search, agent-callable; pitched as the layer above containers.
  • Verification harnesses for AI rewrites — nf-test snapshots are Phil's "zero-pass referral" rule.
  • Cloud-scheduler reform — Intelligent Compute openly replacing the HPC-mimicking Batch model.
  • nf-core as default sharing channel — every academic and industry talk in this session targeted it.

See Themes & takeaways for cross-day synthesis.