Information and uncertainty in sequence-level prediction for genetic component engineering
File(s)
Author(s)
Beardall, William Alexander Victor
Type
Thesis or dissertation
Abstract
The field of engineering biology has pressing need for predictive tools to enable the design of genetic
components from the sequence level. However, the behaviour of engineered genetic constructs in cellular
environments is tightly coupled to the rest of cellular behaviour in a context dependent manner. As a
result of this context dependence, the development of engineering tools providing reliable predictions of
construct behaviour across a range of cellular conditions has proven challenging. Such challenges are
worsened in practice by the high cost of laboratory experimentation, which prohibits large-scale engi-
neering assessment of constructs in vivo.
This thesis studies machine learning approaches with a view to mitigate these challenges, leveraging
available resources to make predictions in information-limited cases. Particular focus is given to the task
of predicting the rate of translation of mRNA in prokaryotes. It is demonstrated that integration of
multimodal information is of high importance to prediction, as is data efficiency and a Bayesian view of
information and uncertainty.
The underlying biophysics of the problems of interest are studied, as are existing models built upon
them. The failure modes of these models are investigated, and machine learning approaches developed
to mitigate them by integrating sequence and structural information. In the very sparse data regime,
it is identified that biophysical features coupled with direct sequence information can produce superior
performance in low-cost predictors for specific local environments. These models are studied under
Bayesian inference, building on strong theoretical grounds for efficient prediction using kernel methods.
At higher information scales, it is shown that the inclusion of RNA structural information in expressive
models of gene expression enables improved predictive performance and increases information efficiency,
both under point estimation and an approximate Bayesian inference regime.
components from the sequence level. However, the behaviour of engineered genetic constructs in cellular
environments is tightly coupled to the rest of cellular behaviour in a context dependent manner. As a
result of this context dependence, the development of engineering tools providing reliable predictions of
construct behaviour across a range of cellular conditions has proven challenging. Such challenges are
worsened in practice by the high cost of laboratory experimentation, which prohibits large-scale engi-
neering assessment of constructs in vivo.
This thesis studies machine learning approaches with a view to mitigate these challenges, leveraging
available resources to make predictions in information-limited cases. Particular focus is given to the task
of predicting the rate of translation of mRNA in prokaryotes. It is demonstrated that integration of
multimodal information is of high importance to prediction, as is data efficiency and a Bayesian view of
information and uncertainty.
The underlying biophysics of the problems of interest are studied, as are existing models built upon
them. The failure modes of these models are investigated, and machine learning approaches developed
to mitigate them by integrating sequence and structural information. In the very sparse data regime,
it is identified that biophysical features coupled with direct sequence information can produce superior
performance in low-cost predictors for specific local environments. These models are studied under
Bayesian inference, building on strong theoretical grounds for efficient prediction using kernel methods.
At higher information scales, it is shown that the inclusion of RNA structural information in expressive
models of gene expression enables improved predictive performance and increases information efficiency,
both under point estimation and an approximate Bayesian inference regime.
Date Issued
2025-09-30
Date Awarded
2026-08-01
Copyright Statement
Attribution-NonCommercial 4.0 International Licence (CC BY-NC)
License URL
Advisor
Stan, Guy-Bart
Ellis, Tom
Sponsor
UK Research and Innovation
Grant Number
EP/S022856/1
Publisher Department
Department of Bioengineering
Publisher Institution
Imperial College London
Qualification Level
Doctoral
Qualification Name
Doctor of Philosophy (PhD)