EXPERIMENT PLANNING · 2018
Turning experiment results into the next design decision
An automated microbial-engineering pipeline connects computational design, experimental work and learning across iterations.
THE SCIENTIFIC QUESTION
How can each experimental cycle inform the next set of candidate designs?
How the study approaches it
The researchers implement an automated Design-Build-Test-Learn pipeline for microbial production of fine chemicals.
REPORTED FINDING
The study demonstrates an integrated workflow that connects design and experimental results to subsequent microbial-engineering decisions.
What the evidence supports
The results concern the reported pipeline and biological systems. They do not establish automated laboratory capability or performance for BioMedAna.
CONNECTION TO BIOMEDANA
BioMedAna's scientific workflows connect evidence, models and experiment decisions. Our example ranks candidates from synthetic observations; it does not operate laboratory equipment or reproduce this study's genetic-design pipeline.
Explore BioMedAna Twin ↗ANALYSIS PROTOCOL
Decision, data and execution method
Rank untested candidate settings and trace each proposal to an explicit fitted response model.
Input data and units
| Input | Value / range | Role in the calculation |
|---|---|---|
| Observed settings | 0, 1, 2, 4, 5, 6 | Six synthetic responses |
| Generator | 10 − (x − 3.2)² + 0.08 sin(2x) | Arbitrary response units |
| Surrogate | a + bx + cx² | Three fitted coefficients |
| Grid | 0–6; step 0.25 | 25 candidates before exclusions |
| Selection | Highest predicted response | No exploration or uncertainty bonus |
Analysis sequence
- Construct the polynomial design matrix.
- Solve normal equations with pivoted elimination.
- Predict on the candidate grid.
- Exclude previously observed settings.
- Rank remaining designs and report proposals and predicted responses.
Calibration and evaluation
Fit RMSE uses the same six observations used for estimation. There is no held-out or prospective result. Candidate selection maximizes a point prediction without accounting for uncertainty, experiment cost or feasibility.
Reading the results
- The leading candidate is a proposal until tested.
- A restricted grid is not continuous global optimization.
- Numeric settings do not encode genetic designs or automate laboratory equipment.
Next experiments and deployment evidence
- Record design identity, replicates, response and experimental conditions.
- Measure the next proposed candidate, compare with prediction, then refit.
- Add uncertainty and feasibility constraints for costly experiments.
ASK ASSIST · GUIDED EXAMPLE
Explore the question in BioMedAna's scientific workflow.
Choose a request to execute the educational model and inspect the results. This demonstration uses predefined requests and local calculations; it is not connected to the production Assist service.
Rank the next untested design candidates
Rank the next untested design candidates
BIOMEDANA WORKED EXAMPLE
Rank the next experiments from observed responses
This example fits a quadratic response to six synthetic observations and ranks untested operating points. It illustrates the learn-and-design steps of a cycle; no experiment is executed.
Synthetic response and fitted candidate predictions
Surrogate and candidate ranking
Synthetic response = 10 − (x − 3.2)² + 0.08 sin(2x) Fit y = a + bx + cx² by least squares a = -0.2336, b = 6.4172, c = -1.0042 Candidate settings: 0..6 in 0.25 increments Exclude tested settings; rank remaining by predicted response
| Rank | Candidate setting | Predicted response |
|---|---|---|
| 1 | 3.25 | 10.016 |
| 2 | 3.00 | 9.980 |
| 3 | 3.50 | 9.925 |
| 4 | 2.75 | 9.820 |
| 5 | 3.75 | 9.710 |
The fit uses all six observations at settings 0, 1, 2, 4, 5 and 6. The displayed RMSE is in-sample. No held-out evidence, predictive interval or prospective result is available, so candidate rankings are proposals for testing.
Complete calculated values
| Setting | Observed response | Predicted response | Status |
|---|---|---|---|
| 0.00 | -0.2400 | -0.2336 | Observed |
| 0.25 | — | 1.3080 | Proposed |
| 0.50 | — | 2.7240 | Proposed |
| 0.75 | — | 4.0145 | Proposed |
| 1.00 | 5.2327 | 5.1794 | Observed |
| 1.25 | — | 6.2189 | Proposed |
| 1.50 | — | 7.1328 | Proposed |
| 1.75 | — | 7.9212 | Proposed |
| 2.00 | 8.4995 | 8.5841 | Observed |
| 2.25 | — | 9.1214 | Proposed |
| 2.50 | — | 9.5333 | Proposed |
| 2.75 | — | 9.8196 | Proposed |
| 3.00 | — | 9.9804 | Proposed |
| 3.25 | — | 10.0156 | Proposed |
| 3.50 | — | 9.9254 | Proposed |
| 3.75 | — | 9.7096 | Proposed |
| 4.00 | 9.4391 | 9.3683 | Observed |
| 4.25 | — | 8.9014 | Proposed |
| 4.50 | — | 8.3091 | Proposed |
| 4.75 | — | 7.5912 | Proposed |
| 5.00 | 6.7165 | 6.7478 | Observed |
| 5.25 | — | 5.7789 | Proposed |
| 5.50 | — | 4.6844 | Proposed |
| 5.75 | — | 3.4644 | Proposed |
| 6.00 | 2.1171 | 2.1189 | Observed |
Interpretation and limits
A good surrogate fit does not prove an untested design will work. The next cycle must build and test the proposal and compare observed results with prediction. This reduced numeric example does not implement genetic design, laboratory automation or the original paper’s DBTL pipeline.
Inspect the full implementation
Download executable JavaScript ↗Inspect experimentRanking for the synthetic observations, polynomial least-squares fit and candidate selection.
Before applying it to your process
Define the intended decision and acceptance criteria. Evaluate the method on representative held-out runs, inspect assumptions and uncertainty, and confirm performance prospectively before relying on its recommendations.
Editorial review: October 13, 2026. This report combines a publication summary with a BioMedAna educational implementation. The simulated results belong to the worked example; they do not reproduce the published experiments or establish a product benchmark.
EXPLORE MORE RESEARCH