Back To Projects

GeoMx Analysis Setup

GeoMx Analysis Setup

Designing for Failure in High-Stakes Science

Designing for Failure in High-Stakes Science

End-to-end workflow design for scientists preparing genomic sequencing data for spatial analysis. A single wrong decision can corrupt months of research.

Role

Product Designer

Timeline

6 months

PLATFORM

Web

UI/UX Design

Design system

Prototyping

DATA VISUALIZATION

Impact & Results

2x

development efficiency

10

Sprints from concept to launch

User Testimonial

"User-friendly, fast, and conveniently accessible."

When the science is at stake, designing for the ideal isn't enough

Scientists using GeoMx spatial analysis tools were losing accurate insights. The underlying science wasn't wrong but the software forced them to make consequential decisions without guidance. Traditional tools required coding expertise, applied thresholds indiscriminately, and gave no feedback when data quality was compromised. The result: false positives, misrepresented gene expression, and lost areas of interest.

The design challenge went beyond simplification and into building a trustworthy system, especially when things go wrong.

When the science is at stake, designing for the ideal isn't enough

Scientists using GeoMx spatial analysis tools were losing accurate insights. The underlying science wasn't wrong but the software forced them to make consequential decisions without guidance. Traditional tools required coding expertise, applied thresholds indiscriminately, and gave no feedback when data quality was compromised. The result: false positives, misrepresented gene expression, and lost areas of interest.

The design challenge went beyond simplification and into building a trustworthy system, especially when things go wrong.

When the science is at stake, designing for the ideal isn't enough

Scientists using GeoMx spatial analysis tools were losing accurate insights. The underlying science wasn't wrong but the software forced them to make consequential decisions without guidance. Traditional tools required coding expertise, applied thresholds indiscriminately, and gave no feedback when data quality was compromised. The result: false positives, misrepresented gene expression, and lost areas of interest.

The design challenge went beyond simplification and into building a trustworthy system, especially when things go wrong.

Four principles shaped every decision

Visualized QC

Indicate failures immediately, explain why they're happening, and give scientists clear options — don't make them hunt for problems.

Real Time Data Filtering

Let scientists adjust parameters and see the impact in real time. Uncertainty about threshold choices should be resolved through exploration, not guesswork.

Intelligent Recommendations

When the system can be confident, say so clearly. When it can't, show the reasoning and give the scientist what they need to decide for themselves.

Standardized Procedures

Automate decisions most prone to human error. Reduce the surface area where variability can corrupt results and warn the scientist before it's too late to recover.

Four principles shaped every decision

Visualized QC

Indicate failures immediately, explain why they're happening, and give scientists clear options — don't make them hunt for problems.

Real Time Data Filtering

Let scientists adjust parameters and see the impact in real time. Uncertainty about threshold choices should be resolved through exploration, not guesswork.

Intelligent Recommendations

When the system can be confident, say so clearly. When it can't, show the reasoning and give the scientist what they need to decide for themselves.

Standardized Procedures

Automate decisions most prone to human error. Reduce the surface area where variability can corrupt results and warn the scientist before it's too late to recover.

Four principles shaped every decision

Visualized QC

Indicate failures immediately, explain why they're happening, and give scientists clear options — don't make them hunt for problems.

Real Time Data Filtering

Let scientists adjust parameters and see the impact in real time. Uncertainty about threshold choices should be resolved through exploration, not guesswork.

Intelligent Recommendations

When the system can be confident, say so clearly. When it can't, show the reasoning and give the scientist what they need to decide for themselves.

Standardized Procedures

Automate decisions most prone to human error. Reduce the surface area where variability can corrupt results and warn the scientist before it's too late to recover.

Data import: The first step is most fragile

Data import is where most workflows break down. Files are missing, formats are wrong, samples don't match. Rather than letting scientists discover these issues downstream after hours of processing, the import flow was designed to surface every failure immediately, with specific explanations and clear paths forward.

Default import state (left) vs. blocking error state (right). When a scientist attempts to proceed without uploading both required file types, a full-width banner stops them with a specific explanation: "Cannot progress until both DCC files and Lab Worksheets have been uploaded." The system doesn't let them guess — it tells them exactly what's missing and why.

Default import state (left) vs. blocking error state (right). When a scientist attempts to proceed without uploading both required file types, a full-width banner stops them with a specific explanation: "Cannot progress until both DCC files and Lab Worksheets have been uploaded." The system doesn't let them guess — it tells them exactly what's missing and why.

Upload in progress. Two states of the same step: a single file being received (left) and multiple files processing simultaneously (right), each with individual status indicators. The scientist always knows where they are in the upload without having to wait for a final result to understand progress.

Upload in progress. Two states of the same step: a single file being received (left) and multiple files processing simultaneously (right), each with individual status indicators. The scientist always knows where they are in the upload without having to wait for a final result to understand progress.

Upload in progress. Two states of the same step: a single file being received (left) and multiple files processing simultaneously (right), each with individual status indicators. The scientist always knows where they are in the upload without having to wait for a final result to understand progress.

Mockup of GeoMx "Import Data" screen demonstrating blocking and non-blocking warning banners

Two simultaneous alert conditions requiring different user actions — one informational, one requiring a binary decision. The hierarchy of urgency is communicated through visual weight and placement, not just color.

Two simultaneous alert conditions requiring different user actions — one informational, one requiring a binary decision. The hierarchy of urgency is communicated through visual weight and placement, not just color.

Two simultaneous alert conditions requiring different user actions — one informational, one requiring a binary decision. The hierarchy of urgency is communicated through visual weight and placement, not just color.

Multiple things could go wrong during data import.

This screen handles three simultaneous conditions — files still processing, a blocking error (minimum sample requirements not met), and a non-blocking warning (lab worksheets contain extraneous information that doesn't match DCC files). Each condition required a different response: one dismissible, one requiring the scientist to fix the file, one offering a binary choice (accept or remove). The design had to communicate severity, urgency, and action clearly, without collapsing everything into a single undifferentiated alert.

Quality Control: Where results make or break

Quality Control is the highest-stakes section of the workflow. An AOI that silently passes when it should fail will corrupt everything downstream. An AOI that's removed unnecessarily means lost data that took weeks to collect. The design had to make failures visible, explainable, and reversible while keeping scientists in control of every decision.

Failed AOIs surface immediately. The table floats failing samples to the top with red flags, while the stacked bar chart breaks down failure reasons by category so scientists understand how something failed. Summary statistics and a plain-language explanation of the High ATC Count flag give context without requiring statistical expertise.

Failed AOIs surface immediately. The table floats failing samples to the top with red flags, while the stacked bar chart breaks down failure reasons by category so scientists understand how something failed. Summary statistics and a plain-language explanation of the High ATC Count flag give context without requiring statistical expertise.

Contextual explanation on demand. Rather than sending scientists to documentation, a tooltip on any flagged row explains what the flag means in plain language — in context, without disrupting the workflow.

Contextual explanation on demand. Rather than sending scientists to documentation, a tooltip on any flagged row explains what the flag means in plain language — in context, without disrupting the workflow.

Contextual explanation on demand. Rather than sending scientists to documentation, a tooltip on any flagged row explains what the flag means in plain language — in context, without disrupting the workflow.

The "Change Parameters" modal allows scientists to adjust ROSALIND's suggested thresholds. In the default state (left), fields are editable with current recommended values. In the validation state (right), a field that's been changed to an out-of-range value is immediately flagged. The system catches the change before the scientist saves it and triggers a reprocessing cycle on bad parameters.

The "Change Parameters" modal allows scientists to adjust ROSALIND's suggested thresholds. In the default state (left), fields are editable with current recommended values. In the validation state (right), a field that's been changed to an out-of-range value is immediately flagged. The system catches the change before the scientist saves it and triggers a reprocessing cycle on bad parameters.

The "Change Parameters" modal allows scientists to adjust ROSALIND's suggested thresholds. In the default state (left), fields are editable with current recommended values. In the validation state (right), a field that's been changed to an out-of-range value is immediately flagged. The system catches the change before the scientist saves it and triggers a reprocessing cycle on bad parameters.

Designed for the moment when there's nothing to show. When no AOIs remain after applying removal criteria, the table communicates the state clearly: "No AOIs at this Gene Selection date." The scientist knows the system is working correctly, not that something has gone wrong.

Designed for the moment when there's nothing to show. When no AOIs remain after applying removal criteria, the table communicates the state clearly: "No AOIs at this Gene Selection date." The scientist knows the system is working correctly, not that something has gone wrong.

Before applying a threshold change that will remove AOIs from the dataset, the system pauses and asks the scientist to confirm. The modal is specific: it names what changed, explains the consequence, and offers a clear binary choice. Irreversible actions require explicit confirmation, which is a pattern applied consistently throughout the workflow.

Before applying a threshold change that will remove AOIs from the dataset, the system pauses and asks the scientist to confirm. The modal is specific: it names what changed, explains the consequence, and offers a clear binary choice. Irreversible actions require explicit confirmation, which is a pattern applied consistently throughout the workflow.

Normalization: A guided workflow environment

Normalization method selection is the most consequential decision in the workflow, where even experienced bioinformatic scientists disagree. Getting it wrong doesn't produce an obvious error; it produces results that look valid but aren't. The design challenge was building an interface that could make a confident recommendation when the data supported one and give the scientist enough context to override the system intelligently.

First pass: the system makes a recommendation with supporting reasoning through citations. The core idea of demonstrating confidence without asserting the answer was right but the layout buried the recommendation logic and the supporting visualization wasn't present yet. Feedback from bioinformatic scientists revealed they wanted to see the data, not just the rationale.

First pass: the system makes a recommendation with supporting reasoning through citations. The core idea of demonstrating confidence without asserting the answer was right but the layout buried the recommendation logic and the supporting visualization wasn't present yet. Feedback from bioinformatic scientists revealed they wanted to see the data, not just the rationale.

Second and third pass: the visual comparison was introduced — box plots showing the distributional difference between normalization approaches side by side. The recommendation is still present but now sits alongside evidence, letting the scientist evaluate rather than just accept. This version was close but the visualization method and summary placement needed refinement before handoff.

Second and third pass: the visual comparison was introduced — box plots showing the distributional difference between normalization approaches side by side. The recommendation is still present but now sits alongside evidence, letting the scientist evaluate rather than just accept. This version was close but the visualization method and summary placement needed refinement before handoff.

Final design leads with a clear recommendation — "Quantile (ROSALIND Recommended)" — but shows its work: a side-by-side visualization of alternative methods and reference material for the scientist to choose differently if their data warrants it.

Final design leads with a clear recommendation — "Quantile (ROSALIND Recommended)" — but shows its work: a side-by-side visualization of alternative methods and reference material for the scientist to choose differently if their data warrants it.

The iteration from v1 to v4 reflects a core design conviction: in high-stakes scientific contexts, trust is built by showing reasoning and respecting the scientist's expertise, not by simply asserting confidence.

Analysis: Catching errors before they make their way downstream

Setting up a valid differential expression comparison requires simultaneous awareness of sample attributes, tissue types, group sizes, and statistical validity. The analysis builder was designed to make this process navigable for scientists at every expertise level, providing guidance and catching configuration errors before they became scientific ones.

Category selectors establish the structure and a drag area makes the interaction model immediately legible. As the scientist begins building their comparison, the system provides a contextual guidance banner, anticipating where scientists get stuck and intervening before confusion sets in.

Category selectors establish the structure and a drag area makes the interaction model immediately legible. As the scientist begins building their comparison, the system provides a contextual guidance banner, anticipating where scientists get stuck and intervening before confusion sets in.

When the system detects that the current data fits with a known valid subset, it lets the scientist know. The modal offers a clear path to accept or explore further.

When the system detects that the current data fits with a known valid subset, it lets the scientist know. The modal offers a clear path to accept or explore further.

When the system detects that the current data fits with a known valid subset, it lets the scientist know. The modal offers a clear path to accept or explore further.

ROSALIND detects a comparison setup that will produce unreliable outputs — samples configured in a way that violates the statistical assumptions of the model — and intervenes before processing begins. The warning names the problem, explains the consequence, and offers to either fix the setup or proceed with full awareness of the risk. The scientist stays in control, but the decision is never uninformed.

ROSALIND detects a comparison setup that will produce unreliable outputs — samples configured in a way that violates the statistical assumptions of the model — and intervenes before processing begins. The warning names the problem, explains the consequence, and offers to either fix the setup or proceed with full awareness of the risk. The scientist stays in control, but the decision is never uninformed.

ROSALIND detects a comparison setup that will produce unreliable outputs — samples configured in a way that violates the statistical assumptions of the model — and intervenes before processing begins. The warning names the problem, explains the consequence, and offers to either fix the setup or proceed with full awareness of the risk. The scientist stays in control, but the decision is never uninformed.

The most important guardrail in the comparison setup system is catching a configuration error before it produces misleading results.

For the scientist who wants to go deeper, a subset refinement modal allows further constraining of the comparison groups by attribute and value. Two states: selecting an attribute (left) and with a value chosen (right).

For the scientist who wants to go deeper, a subset refinement modal allows further constraining of the comparison groups by attribute and value. Two states: selecting an attribute (left) and with a value chosen (right).

For the scientist who wants to go deeper, a subset refinement modal allows further constraining of the comparison groups by attribute and value. Two states: selecting an attribute (left) and with a value chosen (right).

The final state of a correctly configured comparison is deliberately minimal: a single checkbox confirming the comparison name, and two clear actions. By this point the scientist has been guided, validated, and warned at every step. A quiet confirmation signals that the system is confident in what they've built.

The final state of a correctly configured comparison is deliberately minimal: a single checkbox confirming the comparison name, and two clear actions. By this point the scientist has been guided, validated, and warned at every step. A quiet confirmation signals that the system is confident in what they've built.

Components built to scale

Every new component designed for this project — alert banners, status indicators, chart variants, confirmation modals — was built to be reusable across future features. This was built to give engineering to build the next iteration faster.

1 consistent pattern for 3 severity levels. The Alert Flag component was designed to communicate urgency without interrupting the workflow unnecessarily. A color bar serves as the primary semantic signal, consistent typography and real-space padding that keep messages legible at a glance.

Error (red) — blocking. The scientist must resolve this before proceeding.

Alert (amber) — non-blocking but consequential. The scientist can continue, but should be aware.

Guidance (blue) — informational.

Built once, used everywhere.

Key Takeaways

Designing for a high-stakes scientific workflow meant designing for every way it could go wrong in addition to the happy path. The most meaningful design decisions in this project weren't the hero screens; they were the error states, the confirmation dialogs, the validation messages, and the moments where the system caught a mistake before it became a problem that affects scientific research.

This project established a component library and interaction pattern language that the engineering team built against for subsequent features, shaping how ROSALIND approaches guided, trustworthy UX across the platform.

NEXT PROJECT

NEXT PROJECT

Looking for design help?
Let's talk

Book A chat

Currently booking engagements through June

/aimartih

aitana@aitanamh.com

aitanamh.bio

Looking for design help?
Let's talk

Book A chat

Currently booking engagements through June

/aimartih

aitana@aitanamh.com

aitanamh.bio

Looking for design help?
Let's talk

Book A chat

Currently booking engagements through June

/aimartih

aitana@aitanamh.com

aitanamh.bio