Rank against the ask.
The catalogue is scored against what the researcher described, in their own words.
The pipeline you need, without knowing its name.
The challenge
Basepair runs genomic analysis pipelines for research labs. Each new assay, genome build and aligner combination earned its place in the catalogue, and every addition was worth having. But the catalogue could only be entered by name. A researcher had to know a pipeline existed, know what it was called, and know it was the right one, before the platform could help at all.
Search was the presenting problem. It was not the real one, and we only found that out by solving it first.
Interface concept
Catalogue
142 pipelines. Find one by name.
align RNA-seq reads from mouse liverSearch
No pipeline is called that. 0 of 142 match.
- rna-seq-star-mmRNA-seqmm39
- rna-seq-hisat2-mmRNA-seqmm39
- rna-seq-kallisto-mmRNA-seqmm39
- wgs-bwa-mmWGSmm39
- chip-seq-bowtie2-hg38ChIP-seqhg38
- atac-seq-macs2-mmATAC-seqmm39
and the three added last quarter
The catalogue could only be entered by the name of the thing you had not found yet, so the more the platform could do, the harder it was to reach.
01 / Discovery
Describe the experiment. Not the pipeline.
Product
Say what you are trying to do.
We indexed the catalogue into a knowledge base and put discovery inside a chat window, so the thing a researcher types is the experiment rather than the name.
Explain the exclusions.
Two of six pipelines drop away because the samples are mouse, and the interface says so.
Argue with the reasoning.
Showing why something was ruled out lets a researcher disagree with the logic, not the result.
Product
Interface concept
Discover
One conversation, from the first question to the finished run.
Align RNA-seq reads from mouse liver
Basepair
- rna-seq-star-mmSplice-aware, best for liver.94
- rna-seq-hisat2-mmFaster, slightly lower recall.81
- rna-seq-kallisto-mmQuantifies without aligning.66
Run rna-seq-star-mmCompare parameters
Run it on the four liver samples
Basepair
About 40 minutes. A notification arrives when the results are ready.
Results ready4 samples, 4,812 genes quantified
Discovery that ends in a pipeline name still leaves the work somewhere else. Two tools became one window.
02 / Execution
Finding it was only half the job.
Discovery that ends in a pipeline name still leaves a researcher to go and run it somewhere else. The next stage let them submit the run from the same conversation, with parameters confirmed in place, and be notified when the results landed.
03 / The question underneath
Why were there so many pipelines in the first place?
Our analysis
The same steps, shipped again under another name
rna-seq-star-mm
- QC
- Trim
- Align: STAR
- Count
- Differential
rna-seq-hisat2-mm
- QC
- Trim
- Align: HISAT2
- Count
- Differential
rna-seq-kallisto-mm
- QC
- Trim
- Quantify
- Differential
wgs-bwa-mm
- QC
- Trim
- Align: BWA
- Call variants
Four pipelines. Eighteen steps between them, nine of them distinct. Everything outlined is a step that already exists in another pipeline. The filled chips are all that make one pipeline different from the next.
Pipelines turned out not to be distinct things. They are combinations of a much smaller set of steps, so the catalogue was big because every useful combination had to be anticipated, built, named and shipped in advance.
Which meant the combination a researcher actually needed either already existed, or could not be run at all.
04 / Runtime pipelines
So we stopped indexing pipelines and started indexing steps.
A deterministic path, pick a pipeline and then move through its fixed stages in order, became a probabilistic search across every step in the catalogue.
That is what lets someone bring their own data into the middle of an existing pipeline, skip ahead to a result several steps down the line, or compose a run from steps that never shared a pipeline.
System diagram
Before / indexed as pipelines
rna-seq-star-mm
- QC
- Trim
- Align: STAR
- Count
- Differential
rna-seq-hisat2-mm
- QC
- Trim
- Align: HISAT2
- Count
- Differential
rna-seq-kallisto-mm
- QC
- Trim
- Quantify
- Differential
wgs-bwa-mm
- QC
- Trim
- Align: BWA
- Call variants
A step exists only inside its pipeline. You name the pipeline to reach one, and you take all of it or none of it.
After / indexed as steps
- QC
- Trim
- Align: STAR
- Count
- Differential
- Align: HISAT2
- Quantify
- Align: BWA
- Call variants
Your runCount, then Differential
Nine distinct steps, each embedded and scored on its own. A run is any subset in any order, so it no longer has to match something built in advance.
The same catalogue, reindexed. Nothing was added to it: what changed is what the index treats as a unit, and that is what lets a researcher start in the middle.
05 / The interface
Read twice. Once for the answer, once for its shape.
Interface concept
One agent result
{ matched: 4, excluded: 2,
reason: "samples are mouse (mm39)",
runtime: "about 40 min" }- Read one
What is being asked
- Read two
How it should be shown
4 of 6 pipelinesExcluded: 2
- rna-seq-star-mm
- rna-seq-hisat2-mm
- rna-seq-kallisto-mm
Counts by sampleAsked for
06 / The system
Six specialists. One context.
One model trying to do all of this reasons worse than six specialists doing one thing each. All six read and write the same shared context, so a researcher never repeats themselves moving from find me a pipeline, to tell me its parameters, to run it.
System diagram
Orchestrator
Routes each request and keeps the context intact
Pipeline Upload
Validates and embeds new pipelines
Pipeline Search
Runs the query across the index
Ask Pipeline Detail
Retrieves specifications and parameters
Run Analysis
Submits jobs and pulls back results
SynthesizerIn progress
Consolidates answers that span the others
Shared context
Read and written by all six, so a researcher never repeats themselves
From catalogue to conversation
What changed.
Nester Labs demonstrated exceptional thought leadership and development skills.
Coordinating six agents across one shared context is real overhead, and it is not worth spending on a search box.
It is worth spending for what sits at the other end. A new capability now arrives by embedding a new step into the index, and it is reachable the same day, by description, inside pipelines nobody has built yet.