Contextual representations of the chemical space for task agnostic machine learning methods
File(s)
Author(s)
Abdel Aty, Hisham
Type
Thesis
Abstract
The chemical space is a vast and complex multi-scale space that is difficult to represent in a
way that is both computationally efficient and contextually meaningful. In practice, chemists
represent molecular structures in a variety of ways, from skeletal and line formulae for organic
compounds, to spectroscopic representations such as mass spectra (MS) and nuclear magnetic
resonance (NMR) spectra. In this thesis, fast, contextually tailored computational representations of molecules and aggregate structures were developed. These representations not only
give predictive power to computational predictors, but are also used to aid experimentalists
with analysing more classical spectroscopic representations more rapidly.
The first representation developed within this thesis builds on the Small-angle X-ray scattering
(SAXS) representation of lipids, an experimental spectroscopic representation of aggregate
lipid structures. Synthetic modelling of lipid scattering patterns was used to generate training
datasets for a convolutional neural network (CNN) to quickly predict the lipid phase of a given
lipid system based on its SAXS pattern. This new 2D representation is more suitable for CNNs
and provides good and fast predictive power. Combining this with a user friendly WebApp
means that experimentalists are now able to analyse their results and adjust their experiments
at the beamline facility if necessary.
The second representation developed within this thesis is a flexible molecular representation that
initially abstracts structure data into a 768-dimensional vector, which can then be fine-tuned
to a specific task. Whilst the task-agnostic, pre-trained performance of this representation is
reasonable, it is the fine-tuning of the representation to a specific task that gives it its contextual
power, as it can be applied to almost any chemical task given a small sample of data. The model
architecture generates Molecular Fingerprints through Bidirectional Encoder Representations
from Transformers (MFBERT) applied to SMILES strings...
way that is both computationally efficient and contextually meaningful. In practice, chemists
represent molecular structures in a variety of ways, from skeletal and line formulae for organic
compounds, to spectroscopic representations such as mass spectra (MS) and nuclear magnetic
resonance (NMR) spectra. In this thesis, fast, contextually tailored computational representations of molecules and aggregate structures were developed. These representations not only
give predictive power to computational predictors, but are also used to aid experimentalists
with analysing more classical spectroscopic representations more rapidly.
The first representation developed within this thesis builds on the Small-angle X-ray scattering
(SAXS) representation of lipids, an experimental spectroscopic representation of aggregate
lipid structures. Synthetic modelling of lipid scattering patterns was used to generate training
datasets for a convolutional neural network (CNN) to quickly predict the lipid phase of a given
lipid system based on its SAXS pattern. This new 2D representation is more suitable for CNNs
and provides good and fast predictive power. Combining this with a user friendly WebApp
means that experimentalists are now able to analyse their results and adjust their experiments
at the beamline facility if necessary.
The second representation developed within this thesis is a flexible molecular representation that
initially abstracts structure data into a 768-dimensional vector, which can then be fine-tuned
to a specific task. Whilst the task-agnostic, pre-trained performance of this representation is
reasonable, it is the fine-tuning of the representation to a specific task that gives it its contextual
power, as it can be applied to almost any chemical task given a small sample of data. The model
architecture generates Molecular Fingerprints through Bidirectional Encoder Representations
from Transformers (MFBERT) applied to SMILES strings...
Version
Open Access
Date Issued
2023-06-03
Date Awarded
01/10/2023
License URL
Advisor
Gould, Ian
Sponsor
UK Research and Innovation
Grant Number
EP/R513052/1
Publisher Department
Chemistry
Publisher Institution
Imperial College London
Qualification Level
Doctoral
Qualification Name
Doctor of Philosophy (PhD)