● Public preview · 2026

The
reproducibility layer
for bioinformatics.

SciFlow is self-hostable workflow infrastructure for life-sciences labs. Stages are Python or R. Storage is S3. Compute is wherever you run Docker. Data lands, the right pipeline runs, and the result is a citable HTML lab notebook — version-controlled, reproducible, and yours.

LIVE 11 interactive demos · 12 packages · 122 stages
4 trigger types · cron, file-sensor, webhook, chain
BYO infra · your compute, storage, credentials

12 Curated packages

122 Workflow stages

4 Trigger types

02 — Provenance

Every run is a checkpoint. The lab notebook is the record.

Every SciFlow run captures every config, every random seed, every binary SHA. The notebook you publish is the version-controlled artifact, not a screenshot of one. Re-run it next year and you can see exactly what changed in the computational environment — or confirm nothing did.

Reproduce a run from its identifier. Hand someone the run ID; they reconstruct the entire computational environment from the captured manifest. No README, no chat thread, no "what version were you on?"
Compare two runs side by side. Compare runs across time, hosts, parameters, or tool versions — pull up two published notebooks side by side and the software environment, config, and every typed metric are laid out identically, so a difference in mapping rate or a shifted p-value distribution is visible at a glance, not buried in two PDFs' worth of prose.
Cite the URL forever. A run's notebook is a static HTML artifact. Archive it in S3, in Zenodo, in a paper supplement. Nothing to keep running for the citation to resolve.

runs / bulk_transcriptomicsmain

a9f4c2e run-0xA9F4 · all green

2026-05-29 09:48 · cpu-worker-03 · 11 stages

7b1d901 run-0xA9E2 · re-quant with salmon 1.10.3

2026-05-27 14:11 · cpu-worker-02 · 11 stages

3e5b240 run-0xA9D5 · baseline

2026-05-24 22:03 · cpu-worker-03 · 11 stages

diff a9f4c2e ↔ 3e5b240 · illustrative

+ binary: salmon 1.10.2 → 1.10.3

+ result: mapping rate 92.4% → 92.6%

− sample: outlier flagged in interpretation pass

03 — Mechanism

A pipeline that waits for the file, not the other way around.

Most workflow tools assume a human clicks run. SciFlow assumes the trigger is a file landing in S3, a webhook from an instrument, a cron tick, or the completion of an upstream workflow. Describe the workflow in plain language and you get a stage graph back. Inspect any stage and you get the figures, the numbers, and a read of what happened — checked against domain-expert-authored thresholds for that method.

Data lands. A sequencer writes FASTQ to S3, an MD job finishes a trajectory, a caller emits VCF — or a webhook fires, or a cron tick, or an upstream workflow completes.
The right pipeline runs. A trigger watches for the pattern, picks the workflow from the curated catalogue, and hands the stage graph to GPU workers for MD or CPU workers for everything else.
Stage outputs read themselves. Every stage emits typed outputs. A built-in interpretation pass reads the numerical results against domain-expert-authored thresholds for that method, and writes a one-paragraph plain-English read of what's normal and what isn't, alongside the raw artifacts.
You get a lab notebook. A static HTML report renders every stage — figures, tables, files, parameters, binary SHAs. No login to view. Forward the URL. Cite it. Archive it. Hand it to a collaborator or to future-you with nothing else attached.

sciflow · file-sensor · run-0xA9F4

# 09:14:22  cron tick — scan s3://lab-raw/runs/
→ event=file_landed key=RNA-2026-A12.fastq.gz
→ match=bulk_transcriptomics (pattern: *.fastq.gz)
→ workflow=salmon_quant → tximport → deseq2
✓ dispatched · 11 stages · cpu-worker-03 (docker)
# 09:48:51  finished · all stages green
→ interpret=3 anomalies flagged (2 outlier samples, 1 low-mapping)
→ commit=a9f4c2e (config + 14 binary SHAs)
✓ notebook /runs/0xA9F4/index.html 

04 — Catalogue

Twelve packages. One hundred twenty-two stages. All curated.

Every stage is a typed, versioned operation against named file formats. Every package ships the same automation surface: cron, file-sensor, webhook, workflow-chain.

catalogue / v1
122 stages
12 packages

Eleven domains across the sequencing-to-structure arc — genomics (short-read, long-read, bacterial isolate), transcriptomics (bulk & single-cell), microbiome (shotgun metagenomics & amplicon), sequence & phylogenetics, functional annotation, epigenomics, and structural biology (molecular dynamics).

01 / GENOMICS Short-Read Genomics Illumina germline variant calling — fastp, BWA-MEM2, GATK4 BQSR + HaplotypeCaller, bcftools normalize + filter, bcftools stats QC, VEP. short_read_genomicslive demo 02 / GENOMICS Long-Read Genomics Nanopore & HiFi assembly — Flye/hifiasm, Medaka polishing, Sniffles2 SVs, QUAST/BUSCO + Merqury QV scoring. long_read_genomicslive demo 03 / GENOMICS Bacterial Isolate Genomics Isolate assembly to classification — Shovill/SPAdes, QUAST/BUSCO QC, Prokka/Bakta annotation, GTDB-Tk r220 taxonomy, barrnap rRNA. bacterial_isolate_genomicslive demo 04 / TRANSCRIPTOMICS Bulk RNA-seq Reads to differential expression — Salmon, tximport, DESeq2/edgeR, GO & GSEA enrichment. bulk_transcriptomicslive demo 05 / TRANSCRIPTOMICS Single-Cell RNA-seq scRNA QC to annotated atlas — Seurat, Harmony/fastMNN integration, Leiden + UMAP, Azimuth/CellTypist labels. single_celllive demo 06 / MICROBIOME Metagenomics Shotgun profiling & MAG recovery — Kraken2/MetaPhlAn, MEGAHIT, MetaBAT2, CheckM2, HUMAnN function. metagenomicslive demo 07 / MICROBIOME Amplicon Profiling 16S/ITS marker genes — RDP classifier (RDP 2.14), alpha/beta diversity & ordination. ampliconlive demo 08 / SEQUENCE Phylogenetics Sequences to a published tree — MAFFT, trimAl, ModelFinder, IQ-TREE 2, outgroup-rooted ML tree with UFBoot/SH-aLRT support. phylogeneticslive demo 09 / ANNOTATION Functional Annotation Genes to function — InterProScan domains, KEGG module completeness, GO & GSEA enrichment, evidence-tiered annotation merge + report. functional_annotationlive demo 10 / CHROMATIN Epigenomics ChIP-seq chromatin profiling — MACS3 peak calling with input control, deepTools FRiP/fingerprint enrichment QC, HOMER peak-to-gene annotation, and de novo + known-motif enrichment (HOMER / MEME-FIMO vs JASPAR2024). epigenomicslive demo 11 / STRUCTURE Structural Biology MD · SMD · docking Prep, docking & molecular dynamics — PDB/PSF prep, solvation, AutoDock Vina docking, NAMD production & steered MD, MDAnalysis trajectory analysis. structural_biologylive demo

05 — AI, method-grounded

Ask any stage why. Answered from documented, expert-authored conventions.

Every tested package's stages carry domain-expert-authored guidance on what a good result looks like — QC thresholds, mapping-rate norms, dispersion behavior, convergence criteria. When you ask a stage what it did or why, the answer is grounded in that documented guidance, not improvised. Attach your own reference PDFs to a project and the same assistant retrieves and reasons over them too.

Type

"Which genes are differentially expressed between airway smooth muscle cells treated with steroids vs. control?"

→

SciFlow drafts

read_qc · FASTQ quality & adapter check
salmon_quant · transcript quantification
bulk_de · DESeq2 differential expression
go_enrichment · pathway enrichment on hits

edit any node before it runs · trigger: manual / cron / S3

A · Describe

Describe a workflow in words. Get a stage graph.

Type the question you're trying to answer. SciFlow proposes a stage graph from the catalogue, names the inputs it expects, and shows the trigger you'd wire it to. You can edit any node before it runs.

B · Your key

Your key, your model.

ANTHROPIC_API_KEY or a local Ollama endpoint — wire in whichever you run.

C · Grounded

Outputs that explain themselves.

QC metrics, mapping rates, p-value distributions, RMSD trajectories: read against domain-expert-authored thresholds for that method, so the explanation traces to a documented standard, not a guess.

06 — Architecture

Bring your own infrastructure. Bring your own credentials.

Jobs run in Docker containers — that is the entire runtime contract. Beyond that, every component is something you own and operate. Your compute, your storage, your credentials, your data — end to end. Laptop today, homelab tomorrow, K8s when you outgrow it. The notebook URL stays valid the whole way.

Compute

Laptop / homelab / Kubernetes / AWS ECS / GCP / Azure / on-prem

Storage

Any S3-compatible store — AWS S3, Cloudflare R2, MinIO, Backblaze B2, Wasabi

LLM (optional)

Anthropic / local Ollama

Credentials

Yours, end-to-end. No SciFlow-side keychain. No SaaS account.

Licensed tools

Yours, under your own license — NAMD, KEGG, SignalP and other restricted tools: you supply the binary or database, SciFlow runs the stage against it. Never bundled, never redistributed.

Scale

From a single Docker host to a K8s pool.

Workers pull jobs from a queue. Add workers, you get throughput. Remove them, you don't. The control plane is small enough to run on the same laptop that runs the first job; the worker plane scales to whatever you point it at.

GPU

Your GPU pool. Your CUDA. Your drivers.

Structural-biology stages (NAMD, OpenMM, AlphaFold) target your GPU workers directly. No hosted GPU billing. No queue limits set by someone else's pricing tier. Plug into the SLURM allocation you already have.

Licensing

We orchestrate. We don't relicense.

Commercially-restricted tools — NAMD, KEGG, SignalP, TMHMM — run under your license, never ours. You accept the vendor's terms, install the binary or database on your own worker, and SciFlow runs the stage against it. The licensed software is never bundled into SciFlow, copied between tenants, or redistributed. Your compliance stays clean, and the obligation stays where it belongs — with the licensee.

Your data, your compute, your control

No SciFlow-hosted database, no SciFlow-hosted bucket, no SciFlow-hosted billing. Your sequence data and results never leave infrastructure you operate.

07 — Landscape

Where SciFlow sits.

Self-hostable platforms (KBase, Galaxy) wait for a human to click run. General-purpose orchestrators (Airflow, Dagster, Prefect) automate fine, but know nothing about a FASTQ. Hosted clouds (Terra, DNAnexus) automate and know biology — and meter every run on their billing surface. Pipeline engines (Nextflow, nf-core) execute a DAG beautifully but don't watch for events or capture the result. SciFlow runs the same shape of workflow on infrastructure you control: your laptop, your cluster, your VPC, or a managed deployment.

Platform	Manual	Cron	File-sensor	Webhook	Chain	Bio catalogue	Self-host	Result provenance capture
KBase	●	○	○	○	○	●	○	○
Galaxy	●	partial	○	partial	○	●	●	○
Terra (Broad)	●	●	GCS only	●	●	●	○	○
Nextflow (+nf-core)	●	○	○	○	○	●	●	○
Nextflow Tower	●	●	○	○	○	●	paid	run metadata only
DNAnexus	●	●	●	●	paid	●	○	○
Airflow / Dagster / Prefect	●	●	●	●	●	○	●	○
SciFlow	●	●	●	●	●	●	●	●

08 — Published lab notebooks

Open a real run. Real data. No account.

Six packages have a published lab notebook from a real run on a real dataset — a read-only view of the actual workflow, parameters, logs, and outputs.

structural_biology

If you've been waiting for the reproducibility layer to show up, it's here.

Project SciFlow

Get in touch Contact us