Detecting tumour signals from multi-modal biological features of cell-free DNA fragments

In this thesis I present my work to advance the tumour-naive methods for cancer detection using shallow whole genome sequencing (sWGS) data of cell-free DNA (cfDNA) isolated from plasma. cfDNAs in body fluids inform non-invasive cancer detection. Multimodal Artificial Intelligence (AI) can improve sensitivity by exploiting various biomarkers when cancer signal is sparse. Tumour-informed assays depending on mutations from solid tissue have limited practicality in cancer early detection. Emerging fragmentomic and epigenetic features underpin tumour-naive approaches to cancer screening for individuals with low tumour burden. cfDNA feature extraction depends on instrumental processes such as library preparation and feature extraction. First, I used sWGS data of ten different healthy donors using six library kits and ten data processing routes. Variations across different conditions were evaluated. I collected 430 healthy plasma samples from seven published studies for validation. Trim Align Pipeline (TAP) and cfDNAPro R package were developed to accommodate the specific properties of cfDNA and establish data analysis standards where fragmentomics has relevance in contrast to genomic DNA. Second, I designed the “UNIversal cfDNA feaTure Ensemble” (UNITE) framework, a scalable and sensitive method based on Convolutional Neural Network (CNN) to detect cancer signal from sWGS. By systematically evaluating both CNN and XGBoost models across multi-dimensional feature spaces and tumour fractions (TF), I found that multi-modal strategy predominately achieved optimal sensitivity. In samples with less than 3% TF, CNN based on UNITE framework is more sensitive than a model based on XGBoost (32.9% vs 25.3% at 99% specificity). In summary, the results clarify the fundamental issue of variations introduced by experimental and analytical methods. The UNITE framework can sensitively detect inherent cancer signal represented by cfDNA sequencing data. These results jointly provide a roadmap for better feature integration in cfDNA liquid biopsies.

Name dc:type.qualificationname

Doctor of Philosophy (PhD)

Level dc:type.qualificationlevel

Doctoral

Grantor dc:publisher.institution

University of Cambridge

Year dc:date.issued

2024

Author dc:creator

• Wang, Haichao

Advisor dc:contributor.advisor

• Markowetz, Florian

• artificial intelligence
• cell-free DNA
• genomics
• liquid biopsy
• machine learning
• multiomics

Licence

• https://www.repository.cam.ac.uk/bitstreams/2a1a2897-f64d-4003-bdd5-552a7e186d0a/download
• http://purl.org/NET/rdflicense/allrightsreserved

Language dc:language

eng

Author Identifier

0000-0002-7648-916X

OAI identifier oai:identifier

oai:www.repository.cam.ac.uk:1810/377666