The University of Southampton
University of Southampton Institutional Repository

Enhanced cancer diagnostics using machine learning assisted vibrational spectroscopy

Enhanced cancer diagnostics using machine learning assisted vibrational spectroscopy
Enhanced cancer diagnostics using machine learning assisted vibrational spectroscopy
Traditional cancer diagnostic modalities, including imaging and histopathology are increasingly being supplemented by molecular diagnostic approaches, such as those based on liquid biopsy from a minimally invasive blood draw. This has transformed the field of oncology by enabling precision medicine strategies where continuous monitoring is possible through repeat sampling, earlier detection of disease, real-time assessment of treatment response and resistance. Within this framework, circulating tumour DNA (ctDNA), DNA epigenetic modifications, and protein tumour biomarkers represent quantifiable liquid biopsy biomarkers with established clinical relevance. However, the current methods for measuring these biomarkers typically depend on separate analytical platforms, complex multi-step workflows and costly or large instrumentation, limiting their suitability for rapid or resource-constrained settings. Thus, this thesis aimed to demonstrate the use of vibrational spectroscopy for cancer biomarker analysis. Specifically, attenuated total reflection-Fourier transform infrared (ATR-FTIR) spectroscopy and Raman spectroscopy were combined with machine learning to analyse DNA fragment length, DNA hydroxymethylation and methylation, DNA mutations and protein tumour biomarkers within a rapid, label-free, and analytically simple test.

First, it was demonstrated that vibrational spectroscopy could quantify DNA fragment length, which is diagnostically informative because ctDNA is characteristically shorter than DNA released by healthy cells. Partial least squares regression (PLSR) models trained on monodisperse DNA (50-300 bp) achieved high predictive accuracy (R2 = 0.9-0.94, RMSE = 22-23 bp), with a multimodal fusion approach that combines ATR-FTIR and Raman together further improving performance to R2 = 0.96 and RMSE = 17 bp. A 1D-convolutional neural network (1D-CNN) successfully deconvoluted fragment length profiles from polydisperse mixtures, and transfer learning enabled adaptation to biological samples with low prediction error (RMSE =0.18-7.2%) and an error in weighted mean fragment length of Δμ = 12 bp.

The second study addressed global sample quantification of the epigenetic DNA modifications 5-methylcytosine (5-mC) and 5-hydroxymethylcytosine (5-hmC), levels of which are disrupted in cancer and carry diagnostic and prognostic value. Using APC promoter DNA standards spanning 0-100% modification, ATR-FTIR spectroscopy identified modification-sensitive spectral features, and PLSR achieved strong quantification performance for both 5-hmC (R2 = 0.98, RMSE = 4.3%) and 5-mC (R2 = 0.90, RMSE = 9.6%). Domain adaptation extended this methodology to ctDNA reference material, yielding a methylation calibration curve with R2 = 0.98 under cross-validation .

The third study investigated whether ATR-FTIR spectroscopy could detect DNA sequence differences relevant to ctDNA mutation profiling. Gene-level comparison of TP53 and BRAF plasmid DNA confirmed that differences in sequence length and base composition produce distinguishable spectral profiles. However, analysis of three TP53 mutations alongside a wild-type control revealed only subtle spectral variation that was not statistically separable by unsupervised principal component analysis (PCA). Supervised partial least squares discriminant analysis (PLS-DA) classification yielded poor generalisation establishing that single nucleotide mutations within a complex sequence background remain below the current sensitivity threshold of the technique.

The final study targeted protein tumour biomarkers, including CA125, CA15-3, CA19-9, AFP, and CEA, and demonstrated their spectral separability by PCA within the protein-associated infrared region between 1300 and 1700 cm-1. PLSR models achieved strong performance for quantifying CA125 in phosphate-buffered saline and were then extended to human serum, with stronger performance at elevated concentrations (R2 = 0.96) than around the diagnostic threshold (R2 =0.77). A semi-quantitative classification approach also identified clinically elevated CA125 with macro-average sensitivity of 0.86 and specificity of 0.92.

This thesis demonstrates a proof-of-principle potential of combining vibrational spectroscopy with machine learning for the analysis of both DNA- and protein-based cancer biomarkers. The DNA studies showed that fragment length, global methylation and hydroxymethylation, and sequence variation could be investigated from vibrational spectra, with an experimentally
determined limit of detection of 2.75 ng/µL. However, translation to liquid biopsy applications will require sensitivity approaching the ng/mL concentrations typically obtained for circulating cell-free DNA. In the protein studies, CA125 was investigated in a quantitative serum-based test case for tumour-marker analysis, with an experimentally determined limit of detection of 31 U/mL in human serum. Although this value is below the commonly used clinical threshold of 35 U/mL, its closeness to the threshold currently limits its diagnostic utility. Overall, further validation using more clinically representative samples, together with improvements in platform sensitivity, will be required. Nevertheless, these findings establish a foundation for the development of rapid, accessible and label-free spectroscopic approaches for cancer biomarker analysis.
University of Southampton
Fatayer, Rashad
f6d12299-9516-4ec8-8b15-ce32871157cf
Fatayer, Rashad
f6d12299-9516-4ec8-8b15-ce32871157cf
Ganapathy, Senthil Murugan
a867686e-0535-46cc-ad85-c2342086b25b
Sammut, Stephen-John
be71c4c8-4e37-4fc3-bcb0-fedbab5af3df
Wilkinson, James
73483cf3-d9f2-4688-9b09-1c84257884ca

Fatayer, Rashad (2026) Enhanced cancer diagnostics using machine learning assisted vibrational spectroscopy. University of Southampton, Doctoral Thesis, 224pp.

Record type: Thesis (Doctoral)

Abstract

Traditional cancer diagnostic modalities, including imaging and histopathology are increasingly being supplemented by molecular diagnostic approaches, such as those based on liquid biopsy from a minimally invasive blood draw. This has transformed the field of oncology by enabling precision medicine strategies where continuous monitoring is possible through repeat sampling, earlier detection of disease, real-time assessment of treatment response and resistance. Within this framework, circulating tumour DNA (ctDNA), DNA epigenetic modifications, and protein tumour biomarkers represent quantifiable liquid biopsy biomarkers with established clinical relevance. However, the current methods for measuring these biomarkers typically depend on separate analytical platforms, complex multi-step workflows and costly or large instrumentation, limiting their suitability for rapid or resource-constrained settings. Thus, this thesis aimed to demonstrate the use of vibrational spectroscopy for cancer biomarker analysis. Specifically, attenuated total reflection-Fourier transform infrared (ATR-FTIR) spectroscopy and Raman spectroscopy were combined with machine learning to analyse DNA fragment length, DNA hydroxymethylation and methylation, DNA mutations and protein tumour biomarkers within a rapid, label-free, and analytically simple test.

First, it was demonstrated that vibrational spectroscopy could quantify DNA fragment length, which is diagnostically informative because ctDNA is characteristically shorter than DNA released by healthy cells. Partial least squares regression (PLSR) models trained on monodisperse DNA (50-300 bp) achieved high predictive accuracy (R2 = 0.9-0.94, RMSE = 22-23 bp), with a multimodal fusion approach that combines ATR-FTIR and Raman together further improving performance to R2 = 0.96 and RMSE = 17 bp. A 1D-convolutional neural network (1D-CNN) successfully deconvoluted fragment length profiles from polydisperse mixtures, and transfer learning enabled adaptation to biological samples with low prediction error (RMSE =0.18-7.2%) and an error in weighted mean fragment length of Δμ = 12 bp.

The second study addressed global sample quantification of the epigenetic DNA modifications 5-methylcytosine (5-mC) and 5-hydroxymethylcytosine (5-hmC), levels of which are disrupted in cancer and carry diagnostic and prognostic value. Using APC promoter DNA standards spanning 0-100% modification, ATR-FTIR spectroscopy identified modification-sensitive spectral features, and PLSR achieved strong quantification performance for both 5-hmC (R2 = 0.98, RMSE = 4.3%) and 5-mC (R2 = 0.90, RMSE = 9.6%). Domain adaptation extended this methodology to ctDNA reference material, yielding a methylation calibration curve with R2 = 0.98 under cross-validation .

The third study investigated whether ATR-FTIR spectroscopy could detect DNA sequence differences relevant to ctDNA mutation profiling. Gene-level comparison of TP53 and BRAF plasmid DNA confirmed that differences in sequence length and base composition produce distinguishable spectral profiles. However, analysis of three TP53 mutations alongside a wild-type control revealed only subtle spectral variation that was not statistically separable by unsupervised principal component analysis (PCA). Supervised partial least squares discriminant analysis (PLS-DA) classification yielded poor generalisation establishing that single nucleotide mutations within a complex sequence background remain below the current sensitivity threshold of the technique.

The final study targeted protein tumour biomarkers, including CA125, CA15-3, CA19-9, AFP, and CEA, and demonstrated their spectral separability by PCA within the protein-associated infrared region between 1300 and 1700 cm-1. PLSR models achieved strong performance for quantifying CA125 in phosphate-buffered saline and were then extended to human serum, with stronger performance at elevated concentrations (R2 = 0.96) than around the diagnostic threshold (R2 =0.77). A semi-quantitative classification approach also identified clinically elevated CA125 with macro-average sensitivity of 0.86 and specificity of 0.92.

This thesis demonstrates a proof-of-principle potential of combining vibrational spectroscopy with machine learning for the analysis of both DNA- and protein-based cancer biomarkers. The DNA studies showed that fragment length, global methylation and hydroxymethylation, and sequence variation could be investigated from vibrational spectra, with an experimentally
determined limit of detection of 2.75 ng/µL. However, translation to liquid biopsy applications will require sensitivity approaching the ng/mL concentrations typically obtained for circulating cell-free DNA. In the protein studies, CA125 was investigated in a quantitative serum-based test case for tumour-marker analysis, with an experimentally determined limit of detection of 31 U/mL in human serum. Although this value is below the commonly used clinical threshold of 35 U/mL, its closeness to the threshold currently limits its diagnostic utility. Overall, further validation using more clinically representative samples, together with improvements in platform sensitivity, will be required. Nevertheless, these findings establish a foundation for the development of rapid, accessible and label-free spectroscopic approaches for cancer biomarker analysis.

Text
Thesis_Rashad_Fatayer_PDFA-3_unsigned. - Version of Record
Restricted to Repository staff only until 1 August 2028.
Available under License University of Southampton Thesis Licence.
Text
Final-thesis-submission-Examination-Mr-Rashad-Fatayer
Restricted to Repository staff only

More information

Published date: 2026
Additional Information: Parts of this thesis are associated with patent applications that have either already been filed or are expected to be filed shortly. For this reason, the thesis will remain under embargo for 2 years.

Identifiers

Local EPrints ID: 513194
URI: https://eprints.soton.ac.uk/id/eprint/513194
PURE UUID: 5105947f-fc38-43bc-aafe-e155e8746135
ORCID for Rashad Fatayer: ORCID iD orcid.org/0000-0002-6105-0760
ORCID for Senthil Murugan Ganapathy: ORCID iD orcid.org/0000-0002-2733-3273
ORCID for James Wilkinson: ORCID iD orcid.org/0000-0003-4712-1697

Catalogue record

Date deposited: 07 Aug 2026 13:35
Last modified: 14 Oct 2026 02:46

Export record

Altmetrics

Contributors

Author: Rashad Fatayer ORCID iD
Thesis advisor: Senthil Murugan Ganapathy ORCID iD
Thesis advisor: Stephen-John Sammut
Thesis advisor: James Wilkinson ORCID iD

Download statistics

Downloads from ePrints over the past year. Other digital versions may also be available to download e.g. from the publisher's website.

View more statistics

Atom RSS 1.0 RSS 2.0

Contact ePrints Soton: eprints@soton.ac.uk

ePrints Soton supports OAI 2.0 with a base URL of https://eprints.soton.ac.uk/cgi/oai2

This repository has been built using EPrints software, developed at the University of Southampton, but available to everyone to use.

We use cookies to ensure that we give you the best experience on our website. If you continue without changing your settings, we will assume that you are happy to receive cookies on the University of Southampton website.

×