The shift to precision medicine in cancer focuses on providing therapies targeting vulnerabilities of each individual patient tumor. This approach involves identifying cancer subtypes and discovering targets, such as genetic interactions, to treat patients who lack effective therapy. While computational tools, especially machine learning methods, are essential to analyze complex high-dimensional molecular data and suggest new candidate treatment strategies, their effectiveness is often questioned due to data-related challenges. Specifically, limitations in data collection result in sparse or biased biological data, hindering accurate decision-making and the identification of correct patterns. This thesis proposes state of the art solutions to learn improved prediction models for precision medicine and beyond by leveraging relevant data that was previously ignored, and addressing issues of data sparsity and bias.

Prediction of gene synthetic lethalities to identify novel therapeutic targets has overlooked sequence similarity, which is both a notable indicator of functional relation and available for every gene pair, unlike sparser data sources often used for this prediction task. Existing models also struggle to generalize beyond known synthetic lethalities due to an over reliance on data affected by prominent biases. Similarly, the stratification of cancer cohorts without effective treatments is challenging due to the small sample sizes of cancer (sub)cohorts such as oncogene-driven cohorts. In addition, stratification might not directly uncover an actionable treatment opportunity. The integration of dense protein sequence similarity and comprehensive drug response data each, together with methodological advances, led to significant improvements and revealed promising therapeutic opportunities.

Although these integrations improved the performance of computational methods, selection bias, a nonrandom sampling of training data, remained a significant issue affecting fair evaluation and generalizability. Thus, this thesis also introduces strategies to evaluate and mitigate the impact on model generalizability and fairness when the selected training data is not representative of the underlying population. We first artificially induce multivariate selection bias by favoring the selection of specific clusters of samples to study the fair evaluation of model generalizability. Then, to mitigate selection bias, we advance semi-supervised learning methods that use unlabeled data to gain insight into the distribution of the population beyond the labeled training data and promote sample diversity to counter confirmation bias typical of existing approaches. Our approaches include bias mitigation designed for specific machine learning models, such as forest ensembles and neural networks, and model-agnostic methods that operate under fewer assumptions. We show that diversity-guided semi-supervised learning strategies outperform existing domain adaptation techniques in the presence of various selection biases.

The computational methods proposed in this thesis enhance therapeutic target discovery in cancer and address selection bias in machine learning to advance precision medicine in cancer and improve the generalizability and fairness of bioinformatics models.

Motivation. DNA molecules mutate thousands of times every day. Some mutations are harmful to human cells, and may lead to the loss of function in important genes involved in DNA damage repair (DDR) mechanisms. Diseases such as tumors can exploit mutations in important, driver DDR genes to rapidly proliferate. Specific patterns of mutations (or signatures) are insightful indicators for the presence of DDR malfunctioning, which can be exploited to provide targeted treatment (e.g., by leveraging synthetic lethalities). Different methods have been developed to successfully extract relevant mutational signatures from the genomes of tumor patients. Most approaches are unsupervised and thus do not optimize toward distinguishing DDR deficiencies (DDRd). Supervised approaches achieve this, but rely on labeled in vitro data from tumor cell line genomes during training, due to the lack of DDRd ground truth for tumor patient genomes. Semi-supervised learning could bridge the gap and jointly exploit labeled cell line and unlabeled patient mutation profiles to generalize to patient tumors and provide more clinically relevant DDRd mutational signatures.
Results. We propose Pseudo-labeling Semi-Supervised NMF (PSS-NMF), a novel integrated signature extraction and label prediction method, which extends supervised non-negative matrix factorization (NMF) with the ability to incorporate unlabeled samples into the training via pseudo-labeling. Models learned using PSS-NMF were benchmarked on two different tasks, cancer type and DDRd prediction. PSS-NMF consistently improved prediction for patient tumors over the supervised NMF baseline for both tasks, learning signatures that better transferred to the patient tumor domain: the models achieved Macro F1 scores of 0.3842 and 0.1331 respectively for cancer type prediction, and 0.4928 vs 0.4704 for DDRd prediction. We further validated that PSS-NMF identified DDRd signatures were biologically relevant, by comparing them to known DDRd-related mutational signatures curated in COSMIC and investigating their exposures in patient tumor genomes.

Targeted and successful cellular therapies for disease treatment require an extensive mapping of the complex structure and dynamics of molecular mechanisms which determine the behaviour and function of cell. CELL-seq is a genome-wide screening procedure measuring specific and targeted protein quantities as phenotypic readouts and is employed by the Netherlands Cancer Institute to analyze which genes regulate the protein state of interest. This research aims to explore the current compendium of CELL-seq screens that investigate a range of phenotypes, to create a mapping of gene-gene associations that share similar phenotypic profiles and elucidate biology that is hard to uncover with more conventional screening techniques.

We perform exploratory research to investigate the ability of the screen compendium to show network structures that reflect known biological processes. We find that with stringent requirements on interactions the screen compendium shows enrichment for a wide range of biological processes and known protein-protein interactions. We further conclude that the experimental design biases network behaviour and needs to be accounted for when constructing networks. We recommended a mutual k-nearest neighbor network construction approach, which yielded networks with the most biological relevance.
We compare the CELL-seq screens using findings from the approaches to the DepMap dataset, a well-known collection of synthetic lethality CRISPR screens, and find that the behaviour of these datasets is in many ways mirrored. We conclude that this is both due to the biology they represent and the differences in the number of screens in each dataset. Finally, we compare the coverage of biological processes between the HAP1 compendium and DepMap, and show large overlap in their coverage. Nonetheless, the differences they do show leads us to bring forward two hypotheses for gene-gene interactions that score strongly uniquely in the CELL-seq networks which are biologically plausible but are not found in DepMap or curated literature, warranting future investigations.

All code pertaining to the methods and figures in this work are hosted on GitLab by the High Performance Computing Facility of the Netherlands Cancer Institute. As such the code can be viewed by supervisors, but further details could be shared upon request.

Sample selection bias occurs when the selected samples in a subset of the original data set follow a different distribution than the samples from the original data set. This type of bias in the training set could result in a classifier being unable to predict samples from a testing data set optimally. Domain adaptation techniques try to adapt classifiers to a possible bias in the training or testing set. Subspace mapping techniques specifically do this by trying to find common subspaces between the source and target domain, where the source domain is the domain with all samples used for training, and the target domain is the domain with samples that must be predicted. This project aims to evaluate the effectiveness of two subspace mapping techniques in mitigating sample selection bias. This research assumes that no data samples from a target domain are available, but only unlabelled samples coming from an underlying global domain. The two subspace mapping techniques that will be tested in this paper are subspace alignment (SA) and transfer component analysis (TCA). This paper will show that the subspace alignment method is more effective on data sets with fewer features and where the source and target domains are further away from each other. The transfer component analysis method is more effective when more training samples are available on data sets with fewer features and where the distance between the source and target domain is not too big. The effectiveness of both methods also depends on the type and form of the data sets they are used on.

Sample selection bias is a well-known problem in machine learning, where the source and target data distributions differ, leading to biased predictions and difficulties in generalization. This bias presents significant challenges for modern machine learning algorithms. To tackle this problem, researchers have focused on developing domain adaptation techniques that aim to create robust methods against sample selection bias. One approach is the use of minimax estimation techniques, which belong to the category of inference-based techniques. Despite the extensive research in developing these domain adaptation methods, there remains a critical need to evaluate their performance. This thesis explores the performance differences of various minimax estimation techniques in the presence of sample selection bias, providing insights into their effectiveness in mitigating the challenges posed by biased data. By understanding and evaluating the performance of these techniques, this research contributes to the advancement of domain adaptation methods and their application in real-world machine learning scenarios.

Importance weighting is a class of domain adaptation techniques for machine learning, which aims to correct the discrepancy in distribution between the train and test datasets, often caused by sample selection bias. In doing so, it frequently uses unlabeled data from the test set. However, this approach has certain drawbacks: it requires retraining for each new test set and fails when the number of test samples is very small. Therefore, we seek to study the performance of importance weighting techniques when the unlabeled data comes from an underlying domain, instead of one specific test set. We propose an evaluation framework inspired from scenarios traditionally known for posing difficulties to importance weighting and apply it to two popular algorithms, KMM and KLIEP. Our results reveal that both algorithms produce statistically significant classification improvements in most experiments. However, their performance is highly dependent on the characteristics of the dataset and the sampling bias. In particular, class overlap seems to influence adaptation ability in the case of unequal conditional probabilities of the source and target domains, while the "intensity" of the sampling bias is an important confounding factor when the train set size is small.

Domain adaptation allows machine learning models to perform well in a domain that is different from the available train data. This non-trivial task is approached in many ways and often relies on assumptions about the source (train) and target (test) domains. Unsupervised domain adaptation uses unlabeled target data to mitigate a shift or bias difference between the domains. Deep domain adaptation (DDA) is a powerful class of these methods, which utilizes deep learning to extract high-level features that are common across the domains and robust against the shift. These algorithms adapt to a specific target domain. This has two possible downsides. Firstly, the model might not generalize and thus require retraining for each new domain. Secondly, obtaining data for the target domain(s) might be difficult. There can be situations where both source and target domains originate from a “global” domain, from which the samples are selected in a biased way. We explore a new application of existing DDA methods and answer the question: How effective is deep domain adaptation when adapting with the global domain, instead of the target domain, in case of sample selection bias?. Results with synthetic data show that where target adaptation works, global adaptation also improves accuracy compared to supervised learning, although to a lesser extent.

This paper tackles the problem of sample selection bias in machine learning, where the assumption of train and test sets being drawn from the same distribution is often violated. Existing solutions in domain adaptation, such as semi-supervised learning techniques, aim to correct this bias, but their ability to generalize to unseen test sets remains unexplored. To address this issue, specific semi-supervised methods (self-training and co-training) are trained on biased training sets and tested with an unbiased test set drawn from the same distribution. The results of this paper demonstrate that the semi-supervised methods consistently outperformed or matched the baseline models, with self-training exhibiting greater improvement. Through this study, a promising approach is presented to mitigate sample selection bias in machine learning.

Synthetic lethality (SL) is a relationship between two genes, exploited for targeted anti-cancer therapy, whereby functional loss of both genes induces cell death, but the functional loss of either gene alone is non-lethal. Computational prediction of SL gene pairs is sought after because it is expensive to do lab screening for SL. Existing SL labeled pairs from wet- lab experiments often focus on specific genes or pathways, resulting in notable selection bias. Current SL prediction methods ignore this bias when training on available SL labels, and fail to generalize if test sets follow a different selection bias. One way to mitigate bias is to incorporate unlabeled pairs during model learning. However, conventional semi-supervised methods such as self-training can reinforce bias by adding confidently pseudolabeled pairs, which tend to be most similar to previously included samples. We present DBST, a self-training strategy that addresses the issue by promoting diversity in the selection of pseudolabeled samples. This is achieved using metric learning to find a class-contrastive representation of the feature space, based on which DBST selects diverse (or dissimilar) pseudolabeled pairs. In results for five cancer types, semi-supervised models, including DBST, delivered improved SL prediction performance over the supervised model. Additionally, DBST successfully incorporated unlabeled samples that were more dissimilar among them compared to standard self-training. In experiments with differing biases between train and test sets, DBST showed a slight improvement in performance compared to the supervised model.

Double-strand break (DSB) repair is a critical cellular process which repairs breaks in both strands of the DNA double helix. Different repair mechanisms are tasked with repairing such breaks. Predicting deficiencies in repair mechanisms has been widely used for therapeutic purposes, such as targeting cancer cells that have specific DNA repair deficiencies. DSB repair, however, is not error-free, resulting in mutations. These mutations are also influenced by the DNA sequence surrounding the break site. To the best of our knowledge, sequence representations have not been considered when predicting DNA repair deficiencies. We hypothesise that higher-order information can be extracted from sequence representations. In this study, we research the problem of predicting Non-Homologous End Joining (NHEJ) repair deficiencies. Initially, we evaluate how accurately we can predict NHEJ repair deficiency using only the mutational outcome frequencies (mutational spectra). Afterwards, we examine how combining mutational spectra with representations of the sequence surrounding the break site can improve the prediction of NHEJ repair deficiency. We demonstrate that adding DNABERT sequence representations to mutational spectra features significantly improves prediction accuracy from 94.44% to 96.12%. We also show that even simple sequence representations, such as 1-mer frequencies, can lead to significant improvements. Our findings highlight the importance of including sequence representations with mutational spectra in repair deficiency prediction.

The inclusion of intronic reads in the downstream analysis of RNA-sequencing (RNA-seq) data has long been controversial. Recent studies show that intronic reads do contain relevant biological signal. Additionally, studies have discovered differential expression unique to intronic reads in certain diseases. Nevertheless, most disease prediction studies only use exonic read counts as input to their models. In this study, we investigate the informativeness of intronic read counts for RNA-seq-based machine learning prediction tasks. Furthermore, we explore possibilities to combine exonic and intronic read counts to increase predictive performance. To this end, we use an RNA-seq dataset originating from four different brain regions and try to predict multiple different clinical labels, including Alzheimer's disease and dementia. We start by identifying differently expressed genes by performing differential gene expression (DGE) analysis. Next, we evaluate the predictive performance of both exonic and intronic read counts using logistic regression. Subsequently, we explore some basic machine learning techniques to combine the information contained in both sets. Furthermore, we construct our own model architectures with the aim of gaining information by using both sets. We show, for this dataset, that exonic and intronic reads have overlapping but also unique differentially expressed genes. Using these genes we show that the predictive performance using the exonic and intronic reads is very similar for all predicted labels. We further show that even though different genes are identified, the biologically relevant signal for the prediction task appears to be the same in exonic and intronic read counts. We are not able to leverage the combination of the counts to further increase predictive performance. Existing disease prediction models have neglected the inclusion of intronic reads. In light of our findings, machine learning models that incorporate intronic reads could potentially discover novel biological insights.

Genomics has revolutionized our understanding of evolution, hereditary diseases, and more. The advent of long-read DNA sequencers i.e. Oxford Nanopore Technologies' innovations, has opened many new research potentials in genomics. These sequencers produce significantly longer DNA reads, facilitating novel applications. However, this technological leap brings challenges, particularly in accurate basecalling which is the process of converting raw sequenced measurements into digital base pair sequences. While advances in basecalling accuracy have been steadily improving over the years, the computational intensity remains a bottleneck in genomic analysis workflows, demanding costly high-end GPUs for probabilistic neural network models.

The main problem this thesis addresses is the implementation of an accelerated hardware solution for the compute-intensive process of basecalling long-read sequences. The thesis presents an FPGA-based implementation of the computationally demanding Long Short-Term Memory (LSTM) layers within the basecalling network known as Bonito. However, due to the lack of floating-point arithmetic units available on the FPGA, the FPGA implementation could not achieve competitive performance compared to GPUs.

While the FPGA implementation falls short of GPU performance, it serves as a possible stepping stone toward developing an ASIC solution for implementing the Bonito LSTM layers or potentially implementing the entire Bonito model. An ASIC implementation has the potential for superior performance up to 9 times faster than a GPU implementation while additionally being cost-effective. This suggests that ASICs hold promise as a future direction for accelerating long-read sequence basecalling, allowing for faster and more affordable genomics research.

Motivation: Many tumors show deficiencies in DNA damage repair. These deficiencies can play a role in the disease, but also expose vulnerabilities with therapeutic potential. Targeted treatments exploit specific repair deficiencies, for instance based on synthetic lethality. To decide which patients could benefit from such therapies requires the ability to determine the repair deficiency status of a tumor. It has been suggested that mutational signatures could be better predictors of DNA repair deficiency than loss of function in select genes. However, current models for prediction of repair deficiency rely on mutational signatures extracted using unsupervised learning techniques. As a result, the signatures are not optimized to discriminate between repair deficiency status or pathway. We argue that the supervised learning of mutational signatures guided by repair deficiency status could enable the identification of signatures that are predictive of repair deficiency, and capture underlying mechanisms of DNA repair.
Results: We propose S-NMF, a supervised non-negative matrix factorization method, which jointly optimizes two objectives: (1) learning of signatures shared across tumor samples using NMF, and (2) learning of signatures predictive of repair deficiency using logistic regression. We apply S-NMF to mutation profiles of human induced pluripotent cell lines carrying knockouts of genes involved in three DNA repair pathways: homologous recombination, base excision repair, and mismatch repair. We show that S-NMF achieves high prediction accuracy (0.971) and learns signatures that better distinguish the repair deficiency of a sample. Signatures extracted by S-NMF are similar to cancer-related signatures associated with the same repair deficiency. Additionally, S-NMF can capture signatures of deficiencies affecting distinct subpathways within a main repair pathway (e.g. OGG1 and UNG mechanisms in base excision repair).

Recent advancements in quantification of repair outcomes of CRISPR-Cas9 mediated double-stranded DNA breaks (DSBs) have allowed for the use of machine learning for predicting the frequencies of these repair outcomes. Local DNA sequence context influences the frequencies of mutations that arise when DNA gets repaired after it is targeted by CRISPR (CRISPR outcomes). Contemporary models exploit this and can predict what the frequencies are of CRISPR outcomes at predetermined genomic loci. Predictions of such models are reasonably precise, but there may be opportunities for improvement in how the DNA sequence context is leveraged for making predictions. Some models only utilize a set of hand-crafted features, limiting the available information for the model. Other models do utilize broader sequence context but disregard sequence order or only predict a limited set of outcome classes. In this work we present an attention-based deep learning model that uses DNA sequence context to make fine-grained CRISPR outcome predictions. We present a custom input embedding for representing DSB repair outcomes and we expand on existing methods for analyzing attention-based models.