Skip to content

Day 1 PM — Co-Scientist demos, MLOps, GPUs, reproducibility

Date: 2026-04-30 (afternoon) Source transcript: 2026-04-30-nextflow-summit-day-1-pm

The afternoon swung between vendor demos (Seqera Co-Scientist on protein design, GSK's MCP-per-pipeline architecture) and grounded counter-takes (GPU economics, Bactopia's failed-LLM-rewrite story, Tommy Tang's reproducibility keynote). The honest takeaway: agents work, but only on top of typed pipelines, machine-readable contexts, and human-gated promotion.

1. From Stack Overflow to Co-Scientist — agentic protein design

Tycho (Seqera; collaborator: Rashni)

Walks through a Nipah-virus protein-G binder design competition (hosted by ProteinBased) on Nebius GPU cloud. Contrasts the manual six-tool pipeline of a year ago with today's Seqera Co-Scientist flow:

  • Infrastructure — agent spins up a Kubernetes/compute env
  • Pipeline development — a colleague extended an existing pipeline with RFdiffusion 3 end-to-end in 44 min
  • Interpretation — auto-ranks top 5 binders from run outputs at ai.seqera.io

Tools shown: BoltzGen, ProteinMPNN, RFdiffusion (v3), Nebius GPU cloud, BYOK enterprise deploy, token-scoped guardrails, full lineage/audit context.

Q&A highlights: BYOK so prompts never leave your AI provider (pharma data-egress fear); continuous evals with ground truth; foundation-model knowledge plus paper-fetch tools for biological context.

Why it matters: The 44-min RFdiffusion-3 module-add is the most concrete "agents shipping pipelines" data point of the conference.

2. ADMET property prediction — Nextflow as MLOps engine

Anthony Reyes (background: NASA/JPL ML platforms, SMS scheduling NLP)

ML pipeline that predicts ADMET (Absorption, Distribution, Metabolism, Excretion, Toxicity) from SMILES strings for a bench chemist's CADD tool.

Architecture: internal assay extraction → RDKit feature/descriptor expansion → parallel model training (sklearn classical models on ECS CPU; deep nets on AWS Batch GPU) → MLflow autologging → containerized inference API.

Backstory: SageMaker tried first; team fell back to Nextflow due to familiarity. Human-gated promotion (no auto-deploy to prod).

Why it matters: Nextflow used as a general-purpose MLOps engine on a non-bioinformatics problem (cheminformatics). Useful counterexample to "Nextflow is for genomics."

3. Three GPU case studies in genomics

Jason Warner (Principal Scientific Consultant, Diamond Age Data Science)

Anti-hype GPU talk:

  1. NVIDIA Parabricks as drop-in for GATK — minimal Nextflow changes (accelerator directive + AWS Batch GPU queue), 10× speed/cost only worth it above ~10–20 samples
  2. Custom gene-editing workflow — only the mapping step was GPU-friendly; switching back to CPU Python downstream killed the gains
  3. Single-cell QC — moving to GPU via RAPIDS / scVI

Reusable rule: profile first, refactor only if the whole pipeline can stay on the GPU (FUSE / file-staging overhead eats hybrid pipelines).

Why it matters: A useful counter to the keynote's GPU enthusiasm — for small batch sizes or hybrid Python downstream, GPUs can be a net loss.

4. Lightning round

  • Alejandro (Sanford / Stowers) — regulatory potential analysis across thousands of TFBS in NCBI + newly assembled TOGA genomes; oxygen-sensing pathway shows distinct circuits in marsupials and monotremes; metabolism genes cluster by lifestyle (diving, high metabolic rate).
  • Pediatric leukemia long-read speaker — updates to nf-core/longread for PacBio HiFi: adds Sawfish (SV caller), HiFiCNV (CNV caller), variant annotation, per-CpG methylation in pb-CpG-tools, and fibertools for fiber-seq adenine methylation.
  • Aditi — short pitch on autonomous-lab / microscope vision (mostly garbled in transcript).

5. Avoiding AI slop — Bactopia v4 with strict syntax + custom linters

Robert Petit (Wyoming Public Health Lab)

History of bacterial genomic surveillance post-COVID and the under-resourced public-health-lab user. Tried a wholesale LLM rewrite of Bactopia — failed. Ended up doing it by hand, but with a tight feedback loop on top:

  • Small machine-readable context indices
  • Custom linters for Nextflow code patterns
  • Inline Groovy doc replacing meta YAML
  • Dev-side Claude Code skills (e.g., "add a Bactopia tool")
  • AI-aided migration to Nextflow 26.04 strict syntax + record types

Bactopia v4 stats: fully rewritten, strict syntax, static types, ~250 modules / 70 standalone workflows, plus a Python package and plugin.

Why it matters: Most honest "we tried full LLM auto-port and it didn't work" story at the summit. The win came from building agent guardrails (linters + indices + skills), not from a smarter agent.

6. Pipelines as infrastructure — GSK's Omics Workbench

Janae Adams (GSK Data Automation & Computational Sciences; with Francine Mira & Mike Gonzalez)

GSK's Omics Workbench wraps Nextflow with a standardized pipeline template (params, error handling, Docker catalog, reference configs) so ~100 teams interoperate. Pipelines are versioned, typed, queryable, modular — i.e. machine-readable.

The vision: one MCP server per pipeline, so an LLM agent can only invoke catalog pipelines (no hallucinated ones), bridging to GSK's "Lab in a Loop" automated experiment cycle.

Stack: GCP Batch + BigQuery + Seqera Platform + GCP artifact registry. Project began ~2022.

Org pattern: centralized reference-genome workspace, weekly Nextflow office hours, internal "core developer" cohort.

Why it matters: Concrete blueprint for how a regulated pharma exposes a curated Nextflow catalog to LLMs without hallucination risk — quietly underway since 2022, only now being talked about.

7. Closing keynote — Reproducibility in the age of AI

Tommy Tang (AstraZeneca)

Cites Titus Brown's "95% of bioinformatics analyses can't be reproduced" and the Anil Potti / Keith Baggerly MD Anderson scandal (a row-shifted spreadsheet that halted clinical trials) as motivation.

Six pillars of "good enough" practice:

  1. Tidy data + spreadsheet hygiene (Karl Broman & Karl Woo)
  2. Project organization & file naming (Jenny Bryan)
  3. Git — six commands suffice
  4. Environment management — mamba / uv / renv + Docker / Singularity, Rocker
  5. Literate programming (RMarkdown, Quarto, Marimo) + automation (shell → Snakemake / Nextflow / WDL)
  6. Clean code & functions (purrr, tidyverse, packaging)

LLM caveat (closing): outputs drift even at temperature 0. Use AI as an orchestrator that fills a Quarto template, not as the analysis itself. Live demo: identical prompt to AstraZeneca's internal agent produced different reports/figures across runs.

Q&A: tertiary analysis remains hard to reproduce; suggests journals should require peer reproduction as a submission gate.

Why it matters: The conference's most pointed pushback on agent maximalism — and the LLM-drift demo gives a concrete reason to keep humans in the loop.

Cross-cutting threads

  • Agentic AI on Nextflow converges on three patterns: Co-Scientist (Seqera), MCP-per-pipeline (GSK), template-orchestrator (Tang).
  • Pipeline as data product — typed I/O, versioning, schemas (Nextflow 26.04 output {} syntax came up in Q&A).
  • Guardrails and evals — token-scoped permissions, BYOK, regression eval suites, human-gated promotion.
  • GPU strategy — only worth it at scale; profile bottlenecks first; watch CPU↔GPU transitions.
  • Standardization — pipeline templates, shared Docker registries, central reference genomes, internal core-dev cohorts.

See Themes & takeaways for cross-day synthesis.