A deeper understanding of Multiple Sclerosis (MS) symptom progression is required for diagnostic accuracy and patient care. Remote monitoring through smartphones can provide continuous insights in the well-being of MS patients. This research aims to explore differences between MS and healthy typing behavior through collected typing behavior data (keystroke dynamics) on smartphones. The main contribution of this work was finding patterns within typing data that distinguish healthy typing behavior from MS with unlabeled time series data, except the diagnosis. A dataset of three combined studies was provided, which included 182 patients suffering from MS and 335 healthy subjects. An autoencoder was fitted on healthy samples to learn the temporal patterns in healthy typing behavior. Then, an anomaly detection method leveraged both the reconstruction loss and embeddings of these samples to determine whether MS samples deviated from healthy behavior. MS subjects were evaluated against healthy subjects through the ROC-AUC on both sample level and subject level. An initial hold out dataset of 20\% of the subjects was left out for validation. The overall performance was poor (AUC 0.6-0.8), because we aimed to find one general method for all MS subjects. Key limitations were the large variance in healthy typing behavior, relatively short window lengths, and difficulty in quantification of the training step of the autoencoder. Despite the difficulty of evaluating the performance of our suggested method, the described methods provide interesting monitoring possibilities for neurologists to enhance MS diagnosis. These monitoring possibilities could extend towards monitoring the effect of treatment, which is the first step towards personalised healthcare in MS.

In summary, the contributions within this thesis advance Alzheimer's research by introducing new computational tools and methods to better understand the genetics of the disease and cellular mechanisms. Additionally, showing that single-cell gene expression can be effectively analyzed in a binary format (expressed or not) simplifies genomic data analysis, making it more accessible, efficient, and applicable to a range of diseases and conditions.@en

Bacteria are everywhere and play essential roles in Earth's diverse ecosystems and human health. For example, humans harbor a complex and essential gut microbial community comprising thousands of bacterial species (in addition to numerous viruses, fungi, and microbial eukaryotes). This community helps break down and synthesize nutrients, trains the immune system, and keeps pathogens at bay. However, imbalances in this community are associated with several diseases, including obesity, inflammatory bowel syndrome, and recurrent urinary tract infections. Moreover, bacteria can cause deadly infections, and many are developing resistance to our most potent antibiotics. Studying bacteria is thus essential for identifying differences between pathogens and harmless commensals, countering antimicrobial resistance, and understanding their impacts on human health.

To study bacteria, we typically characterize and compare their genomes. The genome comprises all of an organism's hereditary information, and its genes encode the molecular machines necessary for cell function, providing an overview of an organism's capabilities. The hereditary information in genomes additionally enables inferring the organism's evolutionary history, which helps in understanding why specific traits evolved or aid in inferring transmission links in case of an outbreak.

A challenge with comparing large sets of bacterial genomes is the extensive variation in genome content among many species. For example, two Escherichia coli strains can share as little as 50% of their genes. Current computational tools offer biased or incomplete views of genetic variation among strains. This hinders the identification of genotype-phenotype associations, prevents tracking mobile genetic elements, and limits our understanding of the microbial communities they are part of.

The central question of this thesis is how to design computational tools that enable accurate characterization of genetic variation among diverse bacterial genomes. This thesis introduces new algorithms to identify and represent genetic variation using graph data structures. It additionally presents a tool that characterizes strain-specific genetic variation in microbial communities, even in the presence of same-species strain mixtures. Finally, this thesis uses the previously mentioned tools to investigate the role of the gut microbiome in women with recurrent urinary tract infections, offering novel insights into the gut and bladder dynamics of  E. coli.

Collectively, we expect this work to contribute to an improved mechanistic understanding of bacteria's role in human health, help track and counter the spread of antimicrobial resistance, and inform on the development of microbiome-mediated therapeutics.@en

The intensive care unit (ICU) is a hospital department where critically ill patients requiring organ support or intensive monitoring are admitted. Nowadays, the care provided in an intensive care unit has advanced so that more patients are being discharged alive. Advances in ICU care have increased patient survival rates, yet ICU admission is still associated with high morbidity and mortality both during and after the stay.

Identifying patients at high risk of complications during ICU admission is crucial. Even though there has been an increasing focus on predictive models, making accurate predictions with the data currently available remains challenging. A non-conventional parameter that appears to have promising value is heart rate variability (HRV), which reflects the fluctuations in time intervals between consecutive heartbeats and can be derived from the electrocardiogram. We investigated whether heart rate variability was able to predict ICU mortality and ICU length of stay. We employed a machine learning approach, assessing three models for each outcome. We used nested cross validation to estimate the performance and optimize and select the final models. Data from 468 adult patients admitted to the ICU for 48 hours or longer were analyzed and nine HRV measures were calculated. Two HRV measures, the power in the high frequency band and the standard deviation of 5-minute average RR intervals, showed a significant difference between the ICU survivors (n=398) and ICU non-survivors (n=71). While individual HRV measures had limited predictive power for ICU mortality, combining HRV with clinical features improved performance. The best performing model was an eXtreme Boosting Gradient classifier that used clinical features in combination with three HRV measures (power in the high frequency band, power in the low frequency band and the ratio between the power in the low frequency and high frequency band) achieving an AUC of 0.76. The models predicting ICU length of stay performed poorly, with the best model achieving a mean absolute error of 5.07 days. These findings suggest that HRV, when combined with clinical features, has potential in predicting ICU mortality, though further research is necessary before clinical implementation.

Monitoring ICU survivors after discharge is equally important, as half of these survivors suffer from Post Intensive Care Syndrome, which negatively impacts their quality of life and increases their healthcare needs. A team of researchers from the ICU conducted a pilot study establishing the feasibility of monitoring fifteen ICU survivors using eHealth, but a larger clinical trial was needed to assess feasibility on a larger scale. Therefore, we aimed to prepare for the ICU Recover Box 2.0 study. We evaluated the results and challenges of the first pilot. The study protocol was revised to include more participants and the use of new smart technology, specifically the Corsano CardioWatch, replacing the non-CE marked devices from the first pilot, which required 24/7 researcher availability. Preparations included finalizing the application to the Medical Ethical Review Committee and thinking out the study logistics, resulting in a comprehensive plan for the conduct of the study. Key takeaways from this process include the recognition that research is an iterative process of
continuous learning, that the purpose of the study must be carefully considered with ethical concerns in mind, and that the increasing importance of data requires careful planning for its security, processing and storage. The METC application has been submitted. Once approved, the study can proceed on a well-prepared basis.

Ever since the origin of human life, we have been infected by a wide range of viruses. These pathogens have invaded our cells, leaving behind traces of their presence in our genome, known as endogenous viral elements (EVEs). Among the affected cells are neurons. The infectious hypothesis for Alzheimer's disease (AD) proposes that viral infections may serve as an environmental factor contributing to AD. In our study, we explored this hypothesis for the first time from an endogenous perspective by identifying EVEs in the genomes of both AD patients and cognitively healthy centenarians (CHCs). Using a custom-built data processing pipeline, our findings reveal that the genomes of AD patients on average harbor more EVEs than those of CHCs (p=3.24e-4, incidence rate ratio (IRR) = 1.27). Furthermore, we identified specific chromosomal regions with a higher incidence of viral integration in the AD cohort across different virus families. Our data suggest that viral infections over time have increased the susceptibility to AD, underscoring the importance of preventive measures against infections.

Depression, a leading cause of disability worldwide, is challenging to diagnose due to its reliance on subjective clinical evaluations. Metabolomics, which analyzes small molecules to reflect physiological and pathological states, holds promise for enhancing the diagnosis and identifying biomarkers for depression, potentially leading to better understanding and treatment options. Despite the complexity of metabolomics data, deep learning methods have not been extensively explored due to issues with interpretability, which are crucial for gaining insights into biological mechanisms. This study evaluates a biologically interpretable deep neural network, MetaboNet, trained on metabolomics data, for predicting depression and identifying key metabolites and biochemical pathways relevant to the condition. Our results demonstrate that MetaboNet outperforms logistic regression, though the overall classification performance remains modest. Notably, the classification results revealed sex-related differences, with better performance observed in females. Our findings do not support the capability of MetaboNet to identify biologically relevant individual metabolites. However, MetaboNet shows promise in identifying biochemical sub-pathways and super-pathways relevant to depression, which are validated by existing literature.

Coronary artery disease (CAD) is a condition characterized by the narrowing or blockage of the arteries that supply blood to the heart. It is a major global health burden and is known to be correlated with genetics, but the details of the genetic contribution remain unclear. In this study, we explored the associations between structural variants (SVs) and single nucleotide polymorphisms (SNPs) identified through previous CAD genome-wide association studies (GWAS). We identified the most relevant associations by ranking them using a proposed composite score, combining effect sizes and p-values. Filtering was applied to keep the most significant associations within a 500kb genomic window, resulting in 968 SNP-SV associations. Cross-referencing with tissue-specific eQTL data from the GTEx Portal indicated 321 SNP-SV associations that impact gene expression in coronary artery tissue. GSEA identified associated pathways involving dopachrome isomerase, phenylpyruvate tautomerase, and
catalytic activity. The results suggest that the SVs make an important contribution to the regulation of CAD-related gene expression and the overall risk of CAD.

Cancer represents a huge challenge in the medical world, necessitating early detection methods to improve treatment outcomes. The field of fragmentomics emerged as a promising option towards developing efficient non-invasive cancer diagnosis tools. By analysing the differences between the cfDNA fragments from blood samples of healthy patients and patients with cancer, this study aims to determine the most important fragmentomics features for cancer detection. The methods present in this work involve extracting features from the cfDNA fragments available in the experimental dataset, applying a pipeline of feature selection techniques that removes the redundant features, training and evaluating a logistic regression and random forest classifiers to differentiate between healthy and diseased samples, and finally extracting the feature weights from the trained models to understand which features contributed the most to the classification task. Filter-based variance thresholding and Correlation-based Feature Selection (CFS) were employed to refine the dataset. Independent t-test and the Mann-Whitney U test are used to calculate the relationship between the cancer and healthy samples. The Pearson correlation coefficient calculates the correlation between each pair of features. The classification performance of the two proposed models is assessed using the train/test split and the nested cross-validation techniques. The evaluation reveals that logistic regression constantly outperforms the random forest and that removing the redundant features increases the performance of both classifiers. Certain genomic bins, mostly on chromosomes 1, 7 and 8, contain significant features for the classification task. These findings suggest that understanding the importance of the fragmentomics features can lead to improved diagnostic tools such as cancer detection based on blood tests.

Detecting cancer at an initial stage could change the course of the disease's development. A non-invasive examination consists of the liquid biopsy of blood, revealing biomarkers that could provide information about the existence of a tumour or not in the organism. The research touches upon the relevance of DNA fragments, precisely the length of fragments, in the detection of cancer. An in-depth interpretation of the fragment length distribution for predicting the state of a patient as being healthy or sick with cancer was approached. The distribution was explored from four perspectives: the complete fragment length distribution, the size range from 90 to 150 bp, important lengths selected by the feature extraction methods and the Fourier Transform of the initial data. These were input in three machine learning models. Using the fragment lengths between 93 and 98 produced accuracy and AUC scores of over 0.85 for all supervised classification models. Processing the data with the Fourier Transform and using the amplitude of spectrums as features in the Random Forest model resulted in an AUC of 0.99.

In recent years, a new way of cancer diagnostics has emerged, the analysis of DNA fragments circulating in the blood of cancer patients known as fragmentomics. This DNA, known as cell-free DNA (cfDNA), is an easily available biomarker for cell types. Deducing the tissue origin of cfDNA can reveal anomalies in cell death caused by diseases. That holds great potential in cancer detection and monitoring. The process of establishing the cell composition of a blood sample is called cell deconvolution. This research paper focuses on the comparison of two methods of cell deconvolution. The first one UXM, solves this problem by employing a reference-based technique using a methylation atlas. The second one reference-free cfSort, utilizes a Deep Learning Neural Network to perform the sample analysis. In the paper, however, a simpler architecture was trained due to the difficulties in reproduction. Experiments have been conducted to assess the sensitivity of both methods. Experiments consisted of 5 major cell types together mixed with white blood cell DNA fragments to assess the sensitivity of each method. Furthermore, different metrics such as Pearson's correlation coefficient have been used to determine the accuracy of both methods. In the end, UXM outperformed cfSort in most metrics, including Pearson's correlation coefficient, indicating its superior accuracy in deconvolution tasks. However, cfSort showed potential for higher prediction accuracy with further development and better documentation. The findings highlight the strengths and limitations of both methods. This study suggests that while UXM is currently more reliable, future improvements in cfSort could make it a viable alternative. Continued research is recommended to enhance the accuracy and transparency of these methods, ensuring their effectiveness in real-world healthcare applications.

Alzheimer’s disease (AD) is a neurodegenerative disease affecting roughly 40 million people. 70% of the heritability of AD is expected to be explained by Structural Variants (SVs), however these have been scarcely studied in the context of AD. This study aims to find SVs associated with AD and to investigate the properties of these correlated SVs. To this end, a dataset created using Third Generation Sequencing of Single Nucleotide Polymorphisms (SNPs) and their correlation with SVs was utilised and combined with the full summary statistics and fine-mapped results of a large AD Genome Wide Association Study. This resulted in 85 unique SVs with significant correlations to known AD associated SNPs, of which 5 were also discovered in previous research, and 80 were novel. SVs were then associated with their nearest genes, however this resulted in a relatively low overlap with known AD genes and few to no results when applied to Gene Set Enrichment Analysis. Additionally, the data was tested for enrichment of Tandem Repeats (TRs), Transposable Elements, regulatory elements and mechanisms of gene expression, which found enrichment of TRs and heterochromatin areas and a depletion of deletions, weak enhancers, and transcription elongation areas.

Recent research has indicated attributes of cell-free DNA (cfDNA) called fragmentomics
as a promising method for late stage cancer detection in a non-invasive manner. The pri-
mary objective of this research is to uncover hidden patterns and interactions that could
enhance the accuracy and sensitivity of blood-based cancer diagnostics. This study explores
the complementarity between three fragmentomics features; fragment length distribution,
and nucleotide fragment end sequence diversity and nucleosome positioning for four dif-
ferent sample groups; breast cancer, colorectal cancer, lung cancer and healthy controls.
Various machine learning techniques such as linear regression were employed to quantify
any complementary relationships between the features

Cancer poses a significant clinical, social, and economic burden, necessitating the development of effective treatments. Understanding how drugs interact with cancer cells and their downstream effects is critical for creating new therapies and overcoming drug resistance. This paper compares the predictive performance of the Geneformer model with traditional machine learning methods in predicting the response of cancer cells to perturbation combinations using the sciplex2 dataset.

The research involves preprocessing the sciplex2 dataset, training the models, and evaluating their performance in binary classification of cells as either treated or untreated, and the prediction of gene perturbation impacts. While traditional ML models demonstrated higher accuracy in binary classification tasks, Geneformer excelled in predicting the impact of gene perturbations due to its advanced architecture and extensive pre-training on single-cell transcriptomes.

Key findings reveal that highly expression-correlated gene pairs cause the largest shifts in cell classification, underscoring the importance of gene correlations in biological predictions. Geneformer showed a deeper understanding of gene network dynamics, achieving higher maximum Cosine Shifts compared to PCA embeddings and placing less emphasis on highly differentially expressed (HDE) Single Genes. Instead, it focused on HDE Gene Pairs, indicating its potential ability to capture complex downstream effects of gene perturbations.

This study highlights the potential of integrating advanced deep learning models like Geneformer into drug discovery, offering a pathway for more effective and targeted therapeutic interventions.

Powerful new machine learning models in biomedicine are being developed constantly, further hastened by the advent of transformer-based architectures. These advanced systems can be used for various applications, from diagnostics to assessing drug effectiveness. Many of these are fundamentally cell classification problems. Models like Geneformer [1] use gene expression data to learn how to distinguish between these cell classes. This information is usually obtained through single-cell RNA sequencing. However, the alternative source, bulk RNA sequencing, offers some advantages that make exploring the feasibility of using it to train Geneformer enticing, such as its greater availability and lower cost. In this paper, pseudo-bulk datasets are created from single-cell data by aggregation of gene expressions. A method to generate synthetic single-cell-like data from a bulk dataset is used to create new datasets. Some remain purely synthetic, while others are mixed with real single-cell data. Geneformer is fine-tuned on all generated datasets separately, and its performance in a cell classification problem is measured. It is shown that the more a dataset resembles real single-cell data, the better the model’s performance. Using bulk data to fine-tune Geneformer is proven to be infeasible. The synthetic data fails to effectively fine-tune the model and is proven to not have a meaningful impact when added to a singlecell dataset. It is concluded that the generated synthetic data is of too low quality and that alternative generation methods should be explored.

in bio-informatics visualisations are often used to relay the results of genome-wide association studies (GWAS), which can be used to get a better inside into the genetics of diseases. Over the years many websites have been developed, which can create visualisations for a variety of GWAS, with one of the most famous being locuszoom. Recent advancements in technology have allowed for collection of data which wasn't previously possible and which no visualisation tool currently supports. For the first time it is now possible to collect data on the interactions between two different types of genetic mutations: Single Nucleotide Polymorphisms, which consist of a single change in one nucleotide (one letter in a string of DNA) and Structural Variants, which consist of a change in a series of nucleotides. This paper builds upon an already existing visualisation tool snpxplorer and the goal is to design a visualisation, which captures thesse interactions and to add a new, easily integratable section to snpxplorer containing this new visualisation.

Geneformer is a transformer which is pretrained on Geneformer-30M, a dataset consisting of 29.9 million healthy cells. This paper focuses on how Geneformer shifts its attention, when fine-tuned on a dataset of cancer cells, whose gene expression is expected to be distinct, and which genes are key when making cell-state predictions within such an environment. In this paper we compare the shift in attention, and which genes receive the most attention, in a weight-based analysis.

The observed shift in attention was significant, however the accuracy of the prediction increased minimally. The attention weights were mapped back to individual genes, which showed that while Geneformer shifts its attention towards key genes, largely it is still subject of a batch-effect, namely the amount of expressed genes. Further research into designing a data representation of consistent size might be beneficial.

Accurately predicting how cancer cells respond to drug treatment is important to advance drug development. This paper presents a comparative analysis of Geneformer, a deep-learning transformer pre-trained on transcriptomic data, and Support Vector Machine. Using the Sciplex2 dataset, which includes transcriptomic data from lung cancer cells treated with three drugs, both models were trained to predict the response of cancer cells to drug treatments.

This paper investigates how Geneformer and SVM perform in predicting the treatment label of cells across different drugs and doses, which drug doses are suitable for conducting single-gene perturbation experiments, how accurately can these experiments replicate drug effects, and what are the differences in results between Geneformer and SVM regarding their ability to identify significant genes affecting drug response.

Results indicate that while SVM generally achieves higher accuracy in predicting treatment labels of cells, Geneformer demonstrates better capability in identifying genes whose perturbations mimic drug effects. Geneformer's embeddings show significant shifts towards treated cell states after single-gene perturbations, indicating a deeper understanding of gene interactions in drug response. On the other hand, SVM's predictions rely more on differential gene expression. This comparative analysis underscores the strengths and limitations of each approach in modelling complex biological systems and predicting the drug response of cancer cells.

Advancing protein design is crucial for breakthroughs in medicine and biotechnology, yet traditional approaches often fall short by focusing solely on representing protein sequences using the 20 canonical amino acids. This thesis explores discrete diffusion models for generating novel protein sequences with an all-atom representation, specifically SELFIES a widely used molecular string representation. This all-atom approach considers the atomic composition of each amino acid in the protein. Enabling the inclusion of non-canonical amino acids and post-translational modifications. Using a modified ByteNet architecture and the D3PM framework, we compare the effects of this all-atom representation to the standard amino acid representation on the generated proteins' quality, diversity and novelty. Additionally, we see how a uniform or absorbing noise process affects the results. While models trained on the all-atom representation struggle to generate fully valid proteins consistently, those successfully designed showed improved novelty and diversity. Moreover, the all-atom representation can achieve comparable structural reliability results from OmegaFold to the amino acid models. Lastly, our results show that the use of an absorbing noise schedule is the most effective for both the all-atom and amino acid representation.

Accurately predicting enzyme-substrate interactions is critical for applications in drug discovery, biocatalysis and protein engineering. Building upon the ProSmith algorithm, a machine learning framework with a multimodal transformer for protein-small molecule interaction prediction, this study introduces protein 3D structural data as an additional modality. To integrate this data, we explore additive and multiplicative modality fusion strategies without requiring retraining the original transformer from scratch. Our experiments demonstrate that while the incorporation of structural data does not offer improved performance in random splits, it has the potential to surpass ProSmith in challenging data splits involving unseen small molecules. Notably, the model shows better generalization for underrepresented substrates.

Computer vision systems, such as image classifiers, object detectors and video analysis tools, serve diverse applications, ranging from autonomous vehicles and drone navigation to medical image analysis and anomaly inspection in the manufacturing industry. The development of these systems relies heavily on well established practices, which include the adoption of conventional training and evaluation metrics and benchmark datasets. However, we argue that standard approaches are sub-optimal with respect to the ultimate objectives of the computer vision systems. In this thesis, we question whether the training and evaluation of computer vision systems for object detection and long-term action recognition are typically aligned with human-defined end goals.

Object detectors are deployed for object tracking in autonomous vehicles and drones, but also as user assistive tools in medical image analysis and anomaly inspection in industry. Regardless of the end use, object detectors are trained with standard optimization and evaluation strategies. By investigating whether the optimization and evaluation methods of object detectors correlate with human quality judgments, we discover a discrepancy between established metrics and human preferences. To address this, we propose an alternative training loss that better aligns object detectors with human preference.

Subsequently, we ask whether object detections can be used to improve longterm human action recognition in videos. We find that explicitly focusing on the region containing the detected human is beneficial to long-term action recognition models. Unexpectedly, we also find that including a temporal attention module does not help recognizing the videos. Motivated by this result, we investigate how much temporal information is needed to solve long-term action recognition in three popular video datasets. Our results show that most of these videos can be recognized without any long-term temporal information. This suggests that models trained on these videos might exploit short-term shortcuts, instead of learning long-term temporal dependencies. Importantly, these models would not perform successfully on new videos where long-term reasoning is necessary.

As a follow-up, we investigate the impact of the temporal receptive field in longterm action recognition models. The size of the temporal receptive field determines the capability to encode long-term information in videos, like the actions order and duration. We experimentally verify that large temporal receptive fields are sensitive to order and can overfit on the exact action orders seen at training time. Contrarily, short temporal receptive fields are more robust to order permutations and perform better on a current long-term video dataset. This result further demonstrates the irrelevance of long-term information in current long-term action recognition datasets. Our research findings highlight the importance of using training and evaluation metrics that match the intended use of the computer vision systems and choosing training and evaluation datasets that carefully represent the problem at hand.@en