CPU compute
Most pipelines
48 vCPU
- Memory
- 240 GB
- Local NVMe scratch
- 750 GB
- Ready in
- 2–4 min
From research question to reproducible result.
Describe an analysis in plain language. The agent prepares the inputs, proposes a run on a curated set of version-pinned Nextflow pipelines, and returns results with a complete record of how they were produced.
Inspect results where they are produced
Structures, alignments, reports and tables open directly from a run or an analysis, with no download and no second tool. Files render in the browser, and HTML reports run in a sandbox with no network access.
GFP_P42212.pdb
colabfold/top_ranked_structures
- C-terminal tailResidues 231–238; with Met1, the only residues below pLDDT 70
- Ser65, Tyr66, Gly67Chromophore-forming residues on the central helix, inside the barrel
- Very high, 90 and above
- Confident, 70–90
- Low, 50–70
- Very low, below 50
N-terminusC-terminus
- Rotate, zoom and recolourSwitch between pLDDT, spectrum and chain colouring; step through multi-model ensembles and docking poses.
- Overlay related filesShow a docked ligand with its receptor, or a model inside its density map at an adjustable σ level.
- Macromolecular structures
- PDB, mmCIF, BinaryCIF, MMTF, GRO, PQR, PDBQT
- Small molecules and maps
- SDF, MOL2, MOL, XYZ, CCP4, MRC, DX, Cube
- Sequence alignments
- FASTA, Stockholm, A3M, Clustal
- Residue-class colouring and a position ruler, up to 2,000 sequences.
- Reports, tables and text
- HTML, PDF, PNG, SVG, TSV, CSV, JSON, GFF, GenBank
- MultiQC and Nextflow reports, figures, tables and logs. Compressed tables, text, alignments, structures and maps open directly.
Every result carries its own methods section
Each pipeline run and each custom analysis records what was executed, on which inputs, with which parameters and software, and on whose decision. The record stays attached to the result and exports as JSON.
- Pinned code and referencesPipeline release and commit, Nextflow version, reference build and annotation.
- Parameters with their originEach value labelled as set by the researcher, the server policy or the reference.
- Inputs, checksums and lineageSHA-256 for sample sheets and for custom-analysis scripts, inputs and outputs; size and timestamp for run input files; links to the runs that produced each input.
- Software and the exact commandContainer images per process, tool versions and the Nextflow command as executed.
- Custom analyses includedThe method in plain language, the full script, package versions, and whether each output is still as recorded.
- Pipeline
- nf-core/rnaseq 3.27.0 at its pinned commit
- Nextflow
- 25.10.4
- Reference
- GRCh38 primary assembly, GENCODE v50
- Where
- CPU compute VM, us-central1-a
- Decision
- Approved by the researcher at 14:02 UTC
| save_trimmed | true | Researcher |
|---|---|---|
| fasta | GRCh38.primary_assembly.fa | Reference |
| aligner | star_salmon | Server |
| input | samplesheet.csv | Server |
Answers grounded in citable sources
When a question calls for literature or documentation, the agent searches a curated list of scientific sources. Every cited page is traceable to a search the agent ran in that conversation.
Which strandedness setting fits a TruSeq Stranded mRNA library?
For dUTP-based kits such as TruSeq Stranded mRNA, read 1 maps to the antisense strand, which nf-core/rnaseq calls reverse (nf-core/rnaseq usage). The setting can also stay at auto, in which case the pipeline infers strandedness with Salmon on a subsample of reads (nf-core/rnaseq usage, Salmon, PubMed). Users report, as community opinion, that inference is reliable at typical sequencing depths (Biostars).
Sources
- nf-core/rnaseq: Usagenf-co.re, nf-core documentation
- Salmon provides fast and bias-aware quantification of transcript expressionncbi.nlm.nih.gov, PubMed
- Inferring library strandedness from RNA-seq readsbiostars.org, Biostars Community opinion
9 other results from this search
- Numbered sources under each answerAssembled by the server from the saved search results, so they are identical after a reload.
- Links only to retrieved pagesA link is clickable only if a search returned that page; any other link is flagged as not verified.
- Community content labelledForum and Q&A material is marked as community opinion, distinct from the literature.
- PubMed and PMC
- Europe PMC
- bioRxiv
- UniProt
- EMBL-EBI
- GTDB
- nf-core and Nextflow docs
- Biostars
A curated, version-pinned pipeline catalogue
The agent runs only pipelines that have been reviewed and configured for this service, each pinned to one release, reference and set of containers. The conversation sets parameters; it does not change the method.
- Pinned releasesOne release and one set of container images per pipeline.
- Validated before submissionSample sheets and parameters are checked against the pipeline schema.
- Approval with a cost estimateIn manual mode, the default, each run waits for the researcher's decision.
- Diagnostics with a next stepFailures report the cause, the affected step and the recommended action.
Transcriptomics
- rnaseqBulk RNA-seq quantification, human GRCh38nf-core 3.27.0
Metagenomics
- taxprofilerTaxonomic profiling of shotgun readsnf-core 2.0.1
- magGenome-resolved metagenomicsnf-core 5.5.0
- funcscanFunctional screen of contigs and binsnf-core 4.0.0
Genomes and proteins
- prokkaProkaryotic genome annotationin-house 1.0.0
- blastpProtein search against Swiss-Protin-house 1.0.0
- diamond-uniref90Fast protein search against UniRef90in-house 1.0.0
- pfamscanPfam domain annotationin-house 1.0.0
- proteinannotatorDomains and secondary structurenf-core 1.1.0
- proteinfamiliesBuild and extend protein familiesnf-core 2.6.0
Structure and phylogeny
- proteinfoldColabFold structure prediction on GPU; sequences go to the public ColabFold MSA servernf-core 2.0.0
- proteinfoldESMFold single-sequence prediction; sequences stay on the lab's machinenf-core 2.0.0
- phyloplacePlace sequences on a reference treenf-core 2.3.0
Open biology models, run like pipelines
Seven open models for structure, protein design, sequence embeddings and chemistry. They are predictors, not chat models: each answers one kind of question, and the agent proposes it as a run on one pinned version of the weights, the same way it proposes a pipeline.
- Open weights, commercial use allowedCode and weights of every model are released under the MIT or Apache-2.0 licence.
- Run on the lab's own computeWeights are staged on the lab's machines. Co-folding can optionally use the public ColabFold MSA server; without it, sequences stay in the lab.
- Confidence with a validity checkPredicted poses are checked for clashes, bond geometry and stereochemistry. Confidence scores are the model's own estimate, not measured accuracy.
- Approved and recorded like any runModel runs use the same approval, cost estimate and provenance record as pipeline runs, including the model version.
Structure and interactions
- Boltz-2Co-folding of proteins with ligands and nucleic acids, with a predicted binding affinity for small moleculesMIT licence
- Chai-1An independent co-folding model for a second opinion; can run on single sequences without an MSAApache-2.0 licence
- PoseBustersPhysical-validity check applied to every predicted poseMIT licence
Protein sequences
- ESM CProtein language model embeddings for clustering, search and downstream classifiersMIT licence
- ProteinMPNN, LigandMPNNSequence design for a fixed backbone; LigandMPNN accounts for bound ligands, nucleic acids and metalsMIT licence
Microbial genomes
- gLM2Genomic language model trained on metagenomic contigs; embeddings that capture gene context across proteins and intergenic DNAApache-2.0 licence
Chemistry
- ADMET-AIMore than 40 absorption, distribution, metabolism, excretion and toxicity endpoints from a list of moleculesMIT licence
- ChempropTrain a property model on the lab's own assay data and apply it to new moleculesMIT licence
Structure and embedding models run on the GPU machine; sequence design, chemistry and validity checks run on CPU. A model is offered only when its code and weights both carry a licence that permits use by commercial and academic labs alike.
Skills carry the lab's way of reading results
A skill is a written procedure with a tested script: where a pipeline puts its outputs, which thresholds apply, and what the results can and cannot support. The agent follows the matching skill, so the same question gets the same treatment in every conversation.
How good are the MAGs from yesterday's mag run?
Using skillmag-quality-reportOfficial, version 9c41e07a2b5d
Of 41 refined bins, 6 are high-quality drafts under MIMAG: above 90% complete, under 5% contaminated, with 5S, 16S and 23S rRNA and tRNAs for at least 18 amino acids. Another 12 pass the completeness and contamination thresholds but lack a full rRNA operon, common in short-read assemblies, so they are counted as medium quality.
| Sample | High | Medium | Low | Contaminated |
|---|---|---|---|---|
| gut_S1 | 3 | 8 | 5 | 1 |
| gut_S2 | 2 | 6 | 4 | 0 |
| gut_S3 | 1 | 5 | 5 | 1 |
Counted by the skill's script in a recorded analysis, with mag_per_sample.tsv and a completeness-versus-contamination figure.
- Written for these pipelinesEach skill knows the output files of the pipelines it covers and states what it does not support, such as calling a bin high quality from CheckM2 alone.
- Chosen by the request, or by youThe agent reads a skill when a request matches it. Type / in the message box to choose one yourself; the chat marks each step that uses a skill.
- Recorded with the resultSkill scripts run as recorded analyses, and the methods record names each skill followed and its exact version.
- Reviewed, pinned and read-onlyOfficial skills are added by the lab administrator at one fixed version, which no conversation can edit. Each researcher can turn any of them off.
- Instructions, not permissionsA skill changes how the agent works, not what it may do. Run approval and data access rules apply unchanged.
Official skills
- blast-hit-triage Classify best hits by identity and coverage, flag uncharacterised matches, summarise per sample. blastpdiamond-uniref90
- mag-quality-report CheckM2 completeness and contamination as MIMAG tiers, with GTDB-Tk taxonomy and a figure. mag
- structure-confidence Per-residue pLDDT bands, PAE for domain placement, pTM and model ranking. proteinfold, ColabFold
- lab-figure-style Colour-blind-safe palette, journal column widths, 300 dpi PNG with SVG and PDF. Every figure
- skill-authoring Turn a repeatable analysis into a draft skill in your folder, for the administrator to review and add. New skills
Your dedicated computing resources
Each lab is provisioned with its own cloud machines rather than a slot in a shared cluster queue. A machine is created when a run is cleared to start (after the researcher's approval in manual mode, the default), runs the pipeline against the lab's data, and is released once it has been idle for 30 minutes.
GPU compute
Structure prediction and models
NVIDIA RTX PRO 6000
- GPU memory
- 96 GB
- CPU and memory
- 48 vCPU, 180 GB
- Provisioned
- On demand
Lab workspace
Always on
Private by default
- Files, conversations and run history
- Retained
- Finished results
- Cloud storage
- Sign-in
- Invitation only
- Cleared to run0 min
- Machine created2–4 min
- Pipeline runsOn your data
- Results archivedCloud storage
- Machine releasedAfter 30 idle min
Compute exists only while analyses need it. Idle machines are released automatically, and the workbench shows the hourly rate and the time remaining before shutdown. ColabFold is the one pipeline that sends sequences to an outside service; use ESMFold for unpublished sequences.
Get access today
02AI Workbench is available to research groups by invitation. Once a lab is onboarded, each invited member receives an individual account with their own conversations, run history and access to the lab's dedicated compute.
- Request an invitationYour lab administrator adds you to the workspace.
- Set your passwordOpen the invitation link within seven days and choose a password.
- Start your first analysisDescribe the question and the data; the agent proposes a run for your review.