Fixing the Measurement Model Before the Selection Heuristics: Optimiser Repairs for Living-Cell Self-Driving Laboratories
- 1Duke University
Abstract
Closed-loop Bayesian optimisation assumes things that living cells do not provide. A measured cell is used up and cannot be measured again, cultures drift over a campaign, day-to-day batch offsets shift every reading, and cell-to-cell variation is heavy-tailed and mixed with probe failures. When an autonomous loop goes wrong under these conditions, the usual response is a selection-level repair: change the incumbent, replicate the current best, or explore more. The other option is to change what the model assumes about each measurement. We compared the two in a simulated living-cell laboratory with common random numbers, crossing two surrogate models with four selection repairs over three response surfaces, four pathology regimes and 60 seeds (8,820 campaigns of 30 rounds, five destroyed cells per round). A measurement-aware model, using a robust per-condition summary, shrunk per-condition noise and additive drift and day terms, lowered final simple regret against plain GP-EI by 0.042 [0.024, 0.061] under heavy tails and by 0.039 [0.013, 0.064] with all pathologies on, and raised 90% band coverage under drift from 0.83 to 0.95. It cost 0.015 [0.004, 0.027] on clean data. None of the selectionlevel repairs we tested helped either model once the data were contaminated, and using the best posterior mean as the incumbent raised regret in every regime for both models (+0.015 to +0.045). Because that variant is not integrated noisy expected improvement (NEI), we extended the study with an independent reimplementation of the simulator (10,260 further campaigns plus 4,320 for trajectory and sensitivity analyses). It reproduced the direction of the main effects with smaller magnitudes, showed that the measurement-aware gain persisted when drift was halved or doubled and when the probe-failure rate ranged from 5% to 20%, and found that integrated NEI did not hurt and gave the lowest regret when paired with the measurement-aware model. It also showed that offset-removed coverage overstates absolute calibration under drift. Within simulation, the measurement model is the first place to intervene; we have not tested real cell data.
Related
- Optimiser Repairs for Living-Cell Self-Driving Laboratories: Performance and Redundancy under Experimental Constraints
- RAxML-NG 2: Automatic model selection, novel tree search heuristics, and fast branch support metrics
- On the size of the model selection problem in a conjoint measurement experiment
- Confidence Filtering in Video-Based Gait Assessment: Measurement Error versus Selection Shift
- Complication rates of proximal hypospadias: meta-analyses of four surgical repairs
BibTeX
@misc{brimo2026fixing,
title = {{Fixing the Measurement Model Before the Selection Heuristics: Optimiser Repairs for Living-Cell Self-Driving Laboratories}},
author = {N. Brimo and D. C. Serdaroglu},
howpublished = {bioRxiv},
year = {2026},
month = oct,
doi = {10.64898/2026.10.27.762131},
url = {https://www.biorxiv.org/content/10.64898/2026.10.27.762131v1},
note = {CC BY},
}
RIS
TY - UNPB
TI - Fixing the Measurement Model Before the Selection Heuristics: Optimiser Repairs for Living-Cell Self-Driving Laboratories
AU - N. Brimo
AU - D. C. Serdaroglu
PY - 2026
DA - 2026/10/27
PB - bioRxiv
DO - 10.64898/2026.10.27.762131
UR - https://www.biorxiv.org/content/10.64898/2026.10.27.762131v1
AB - Closed-loop Bayesian optimisation assumes things that living cells do not provide. A measured cell is used up and cannot be measured again, cultures drift over a campaign, day-to-day batch offsets shift every reading, and cell-to-cell variation is heavy-tailed and mixed with probe failures. When an autonomous loop goes wrong under these conditions, the usual response is a selection-level repair: change the incumbent, replicate the current best, or explore more. The other option is to change what the model assumes about each measurement. We compared the two in a simulated living-cell laboratory with common random numbers, crossing two surrogate models with four selection repairs over three response surfaces, four pathology regimes and 60 seeds (8,820 campaigns of 30 rounds, five destroyed cells per round). A measurement-aware model, using a robust per-condition summary, shrunk per-condition noise and additive drift and day terms, lowered final simple regret against plain GP-EI by 0.042 [0.024, 0.061] under heavy tails and by 0.039 [0.013, 0.064] with all pathologies on, and raised 90% band coverage under drift from 0.83 to 0.95. It cost 0.015 [0.004, 0.027] on clean data. None of the selectionlevel repairs we tested helped either model once the data were contaminated, and using the best posterior mean as the incumbent raised regret in every regime for both models (+0.015 to +0.045). Because that variant is not integrated noisy expected improvement (NEI), we extended the study with an independent reimplementation of the simulator (10,260 further campaigns plus 4,320 for trajectory and sensitivity analyses). It reproduced the direction of the main effects with smaller magnitudes, showed that the measurement-aware gain persisted when drift was halved or doubled and when the probe-failure rate ranged from 5% to 20%, and found that integrated NEI did not hurt and gave the lowest regret when paired with the measurement-aware model. It also showed that offset-removed coverage overstates absolute calibration under drift. Within simulation, the measurement model is the first place to intervene; we have not tested real cell data.
ER -