Bayesian design of experiments for biomolecular assay design
File(s)
Author(s)
Sedgwick, Ruby
Type
Thesis
Abstract
Bayesian optimisation is a powerful tool for designing lab experiments in low data settings as it allows for incorporation of prior knowledge while effectively quantifying uncertainty. By using Bayesian optimisation as a sequential decision-making tool, we can reduce the experimental load of developing new bioengineering assays, making previously prohibitively expensive experiments feasible. This thesis focuses on two problems related to the development of biomolecular networks for medical diagnostics.
The first problem is the experimental validation of the mechanistic models that are often used to design biomolecular networks. We propose an efficient design of experiments strategy that uses Gaussian processes to construct a probabilistic model of the discrepancy between experimental results and the mechanistic model, then uses a design criterion to select the next sample points. We develop a stopping criterion, that quantifies this discrepancy over the whole surface while taking uncertainty into account and test the strategy on simulated data of biochemical processes.
The second problem tackled is the optimisation of many competitor DNA molecules needed for the biomolecular networks, for which we present a transfer learning design of experiments workflow. We first investigate the properties of transfer learning Gaussian process models on synthetic data and explore hyperparameter initialisation strategies. Then, by combining a transfer learning surrogate model with Bayesian optimisation, we show the total number of experiments can be reduced by sharing information between optimisation tasks. We use cross-validation on data from the development of DNA competitors to compare the predictive accuracy of different transfer learning Gaussian process models and compare their performance for both single objective and penalised optimisation.
While inspired by the design of biomolecular networks for medical diagnostics, the strategies developed here are application agnostic, as well as paving the way for fully automated lab experimentation when integrated into a design, build, test pipeline.
The first problem is the experimental validation of the mechanistic models that are often used to design biomolecular networks. We propose an efficient design of experiments strategy that uses Gaussian processes to construct a probabilistic model of the discrepancy between experimental results and the mechanistic model, then uses a design criterion to select the next sample points. We develop a stopping criterion, that quantifies this discrepancy over the whole surface while taking uncertainty into account and test the strategy on simulated data of biochemical processes.
The second problem tackled is the optimisation of many competitor DNA molecules needed for the biomolecular networks, for which we present a transfer learning design of experiments workflow. We first investigate the properties of transfer learning Gaussian process models on synthetic data and explore hyperparameter initialisation strategies. Then, by combining a transfer learning surrogate model with Bayesian optimisation, we show the total number of experiments can be reduced by sharing information between optimisation tasks. We use cross-validation on data from the development of DNA competitors to compare the predictive accuracy of different transfer learning Gaussian process models and compare their performance for both single objective and penalised optimisation.
While inspired by the design of biomolecular networks for medical diagnostics, the strategies developed here are application agnostic, as well as paving the way for fully automated lab experimentation when integrated into a design, build, test pipeline.
Version
Open Access
Date Issued
2023-10-27
Date Awarded
2024-06-01
Copyright Statement
Creative Commons Attribution NonCommercial Licence
License URL
Advisor
Misener, Ruth
Stevens, Molly
van der Wilk, Mark
Grant Number
EP/S023283/1
Publisher Department
Computing, Bioengineering, Materials
Publisher Institution
Imperial College London
Qualification Level
Doctoral
Qualification Name
Doctor of Philosophy (PhD)