Large Vision-Language Models (LVLMs) have demonstrated impressive capabilities across a wide range of visual tasks, yet their performance often degrades in specialised domains such as Remote Sensing, where spatial reasoning in Visual Question Answering (VQA) remains particularly difficult. This limitation is primarily driven by a scarcity of high-quality training data, which comes from the rarity of complex real-world events and the labour-intensive process of annotating spatial relationships. To address this, this thesis proposes a synthetic data generation pipeline based on generative inpainting, using parked aircraft as a case study. By inserting synthetic objects into real remote sensing backgrounds, the pipeline creates complex spatial scenarios while preserving realistic context. Because the scene layout is fully controlled, it can also automatically generate grounded image-question-answer triplets together with Chain-of-Thought (CoT) reasoning traces for each answer. To measure this spatial reasoning, we construct an extensive evaluation benchmark on real overhead imagery whose questions span five question domains, from single-object properties to multi-object relational reasoning, which is necessary to properly test complex spatial reasoning in VQA. We fine-tune three open-source LVLMs from different model families on the synthetic data and evaluate them against generalist and domain-specific baselines on this benchmark. Fine-tuning on the synthetic data improves spatial reasoning across all tested domains and across all three models, raising the strongest model, Qwen3.5 9B, from 43.4\% to 71.9\% accuracy, a gain of 28.5 percentage points (pp) over its zero-shot baseline. The consistency of this improvement across architectures indicates that it comes from the data itself rather than from any single model. Adding CoT supervision raises performance further, to a best of 75.5\% (a 3.6 pp increase over standard fine-tuning, and 32.1 pp over the zero-shot baseline). The gains are not uniform: intrinsic orientation (distinguishing the front of an aircraft from its back) remains the lowest-scoring domain for every fine-tuned model, and performance declines as the number of aircraft in a scene grows. Overall, these findings show that synthetic data can serve as an effective alternative or extension to real remote sensing VQA data, helping to address the data scarcity present in this domain, and because the pipeline was built to be modular, its components hold potential beyond this case study, possibly extending to other objects and tasks. ... Read more

Gene regulatory networks (GRNs) inferred at single-cell resolution offer insight into regulatory mechanisms underlying complex diseases such as Alzheimer's disease, but the reliability of the GRNs produced by methods such as ScReNI is not fully explored. This work investigates the stability and reliability of GRNs inferred by a Python reimplementation of ScReNI applied to the Seattle Alzheimer's Disease Brain Cell Atlas (SEA-AD) and a mouse retinal dataset. Three experiments were conducted. First, gene-level prediction accuracy, measured via random forest out-of-bag (OOB) R², was found to vary across genes and to be driven primarily by expression level rather than network position. Second, two main candidate cell-level reliability metrics were evaluated against precision and recall: OOB R² showed no significant association with precision once network density and expression were statistically controlled for, whereas edge weight variance across repeated inference runs retained significant correlation with both, making it the more promising candidate metric identified. Third, comparing the GRNs of cells with high versus low edge weight variance within cell types revealed that such networks share similar topological structure but diverge substantially in edge weight magnitude. These findings indicate that out-of-bag accuracy alone is an insufficient proxy for GRN reliability, and that edge weight variance, while more informative, does not yet provide a fully cell-type-consistent reliability metric, motivating further work on alternative approaches to quantifying single-cell GRN stability. ... Read more

High-quality, interpretable metadata is essential for enabling data reuse and reproducibility in the life sciences. Frameworks such as the Investigation–Study–Assay (ISA) model provide a structured way to describe experimental workflows, yet in practice metadata remains incomplete, inconsistent, and costly to produce. Lab notebooks, which record experimental procedures, offer a promising source to extract metadata from, but their unstructured nature makes extraction challenging.

This thesis investigates extracting ISA metadata from lab notebooks using LLM-based multi-agent systems. It also addresses the lack of suitable methods for evaluating the performance of such tasks. To this end, this thesis contributes two artefacts: (1) a prototype multi-agent system that extracts ISA metadata from lab notebooks, and (2) a rubric-based evaluation framework which formalises and assesses extraction performance.

A feasibility study is conducted on a set of real-world lab notebooks to evaluate the proposed multi-agent system. The results show that partial extraction is achievable with the current prototype, with 41% of ISA entities represented well, and another 27% sufficiently. Therefore, the prototype can provide a practical starting point for researchers in their ISA metadata creation task. While fully automated extraction remains challenging with the current prototype due to missing and insufficiently represented ISA entities, this work provides recommendations for future system design based on the performed feasibility study. Specifically, to operationalise multi-agent ISA extraction in practical workflows, this thesis recommends building a human-in-the-loop system where researchers and agents collaborate to construct metadata. ... Read more

Motivation: Clustering is an unsupervised learning task with broad applications. Traditional clustering methods often rely on point estimates of model parameters, which can limit their ability to capture uncertainty. Bayesian clustering addresses this by incorporating uncertainty into parameter estimation. However, existing Bayesian inference methods like Markov Chain Monte Carlo and Variational Inference are computationally intensive and can produce biased approximations. To overcome these limitations, we propose Cluster-PFN, a transformer-based model inspired by Prior-Data Fitted Networks. Cluster-PFN simultaneously approximates the posterior distributions over the cluster assignments of individual data points and the total number of clusters for the given dataset in a single forward pass. Our model provides fast and accurate Bayesian clustering, supporting data with up to five features and conditioning on user-specified cluster counts.

Results: Our results demonstrate that Cluster-PFN can predict the number of clusters up to 20% more accurately than standard heuristics. It also outperforms the Bayesian Gaussian Mixture Model using Variational Inference (Bayesian GMM VI), achieving up to 60% higher scores on certain external metrics while being up to 20 times faster during inference. Additionally, Cluster-PFN surpasses both the traditional Gaussian Mixture Model and K-means++ across the same external evaluation metrics. ... Read more

Alzheimer's Disease (AD) is a complex heterogeneous disease and is the leading cause of dementia around the world. Treatment options remain limited and the underlying mechanisms are not yet fully understood. To get more insight on this celular level, single-cell gene expression data can be used. It has proven to be effective with machine learning for tasks like cell type classification. While prior studies have explored AD classification using scRNA-seq, this has only been a binary classification. Severity of AD is classified using multiple measures, ranging from cognitive ability scores, to neuro pathological measures. This research explores the possibility of expanding the binary prediction of AD by including these measures for AD severity. In addition, given that these measures are associated, we also investigate if Multi Task Learning (MTL) models can improve the predictions by learning multiple AD related data points. If successful, this approach can give additional analysis into key tasks, genes and/or cells (sub)types that drive the models, which would lead to more possibilities for personalized treatment options, alongside more insight into the development of AD in the brain. We used a three-layer neural network architecture alongside a translation from cellular level to individual level to make individual-level predictions. Results show that Cognitive Ability can be classified best, but overal performance is only slightly above Naive Bayes. Furthermore, MTL does not appear to have any measurable positive effect on scores compared to single task models. A link to the github repository is available at \url{https://github.com/WillemDieleman/ADseverityCSE3000}. ... Read more

The Cytomegalovirus (CMV) serostatus, of both donor and patient, plays an important role in allogeneic hematopoietic stem cell transplantation, yet it is only known for 19% of donors in the global database. In this research, the other available data in the global database of the World Marrow Donor Association will be used to predict CMV serostatus for donors who’s status is unknown. In a statistical analysis, features such as sex, registry, ethnicity, age and height were found to be informative. A particular focus was put on investigating the relation between Human Leukocyte Antigen (HLA) and CMV. Previous literature on this relation consists of studies on small cohorts, Machida et al., 1998 (N=125) and Hassan et al., 2016 (N=1 849). The large global database (N=8 707 407 allows us to evaluate this relation for many more different HLA groups and validate their findings using a meta-analysis across registries. In this analysis, the majority of HLA groups showed small effects on the CMV serostatus. However, a few groups showed consistent large increases in likelihood to be CMV seropositive. This indicates that there is a biological relation between specific HLA and CMV serostatus. To predict CMV serostatus, a simple logistic regression classifier was iterated on by adding features and using more complex models. The best performing classifier uses XGBoost on all features in the database. This classifier has an AUC performance of 0.70. Although this AUC is too low to rely on it for patient-donor matching, it does show non-trivial predictive power. The classifier could be further improved by using a better embedding for the HLA that includes the similarities between different HLA alleles. ... Read more

Integrating single-cell multi-omic data is crucial for comprehensive biological discovery, yet it remains challenging due to the weak correlation between modalities, data heterogeneity, and stringent privacy regulations. Conventional integration methods that depend on shared features or matched cells, which are rarely available in practice. While some diagonal integration approaches might mitigate some of these limitations, they are sensitive to noise, prone to overfitting, and challenging to validate, especially in the absence of centralized data access. This thesis introduces Federated Matching xcross modalities via Fuzzy smoothed embeddings (MaxFuse), a novel adaptation of MaxFuse within a Federated Learning (FL) framework, which enables privacy-preserving diagonal integration through fuzzy smoothing, federated Canonical Correlation Analysis (CCA), and iterative matching without exchanging raw data. We validate Federated MaxFuse on benchmark single-cell datasets, demonstrating that it achieves matching accuracy and embedding quality comparable to centralized baselines across supervised and unsupervised metrics. These findings establish Federated MaxFuse as a practical and scalable solution for privacy-preserving integration of multi-omic data, enabling robust cross-institutional analyses under real-world constraints. ... Read more

... Read more

Support Vector Machines (SVMs) are widely used in various domains, with their performance heavily dependent on hyperparameter selection. However, hyperparameter tuning is computationally demanding due to the SVM training complexity, which is at best $O(n^2)$, where $n$ represents the number of training samples. To mitigate this challenge, we propose integrating a validation-based early stopping criterion into the Sequential Minimal Optimization (SMO) algorithm to enhance tuning efficiency.

We evaluate this approach within Random Search and Successive Halving frameworks, aiming to reduce tuning runtime while preserving model performance. We introduce a composite score function to facilitate a balanced assessment of accuracy and efficiency. Our empirical analysis reveals that incorporating early stopping into SMO significantly reduces hyperparameter tuning time under RS but provides limited benefits in successive halving, given its inherent efficiency. Additionally, while dataset characteristics influence the effectiveness of early stopping, we found evidence that dimensionality does not. We also observe that frequent early stopping objective assessments introduce computational overhead, which can offset runtime improvements. Reducing assessment frequency alleviates this issue, but it diminishes the effectiveness of early stopping.

Our findings highlight the potential of early stopping in SMO for optimizing SVM hyperparameter tuning, particularly within random search-based approaches. They also identify trade-offs in assessment frequency and dataset-specific factors. ... Read more

Stroke remains a leading cause of morbidity and mortality worldwide, despite advances in treatment modalities. Endovascular thrombectomy (EVT), a revolutionary intervention for ischemic stroke, is limited by its reliance on 2D fluoroscopic imaging, which lacks depth and comprehensive vascular detail. We propose a novel AI-driven pipeline for 3D CTA to 2D DSA cross-modality registration, termed DeepIterReg. The proposed pipeline integrates neural network-based initialization with iterative optimization to align pre-intervention and peri-intervention data. Our approach addresses the challenges of cross-modality alignment, particularly in scenarios involving limited shared vascular structures, by leveraging synthetic data, vein-centric anchoring, and differentiable rendering techniques. We assess the efficacy of DeepIterReg through quantitative analysis of capture ranges and registration accuracy. Results indicate that our method is able to accurately register 70% of a testset of 20 patients, and is able to improve capture ranges when performing an initial pose estimation using a convolutional neural network. ... Read more

Large language models enable rich communication and flexible planning in embodied agents, yet their updates to internal state remain correlation driven, making them prone to hallucinations that undermine reliability in interactive settings. To address this limitation, we investigate whether a true causal model can replace template based belief, desire, and intention (BDI) updates and provide more stable reasoning. We introduce CausalMind, an embodied agent architecture that integrates a causal model trained on text based BDI transitions drawn from a synthetic dataset. We evaluate both the causal model and the full agent in iTHOR across solo and cooperative tasks.

The results indicate that learning causal variables from text embeddings is difficult. The model focused primarily on action related features, failing to separate belief, desire, and intention and instead overfitting to biases in the dataset. Expanding and refining the synthetic dataset improved performance but did not resolve the core issues, highlighting the need for richer data and more suitable encoders. At the agent level, perception errors from the vision module frequently led to hallucinated task completion, limiting the usefulness of teacher guidance and making complex tasks difficult in the iTHOR environment.

Although the current implementation does not outperform template based baselines, the study clarifies the obstacles that arise when extending causal representation learning to natural language BDI states. These findings outline the key requirements for future work toward embodied agents capable of robust causal reasoning rather than correlation driven updates. ... Read more

Synthetic polymers are crucial in diverse industries, but current AI-driven design methodologies primarily target linear homopolymers, with limited emphasis on developing customized approaches for copolymers. To address this gap, we introduce a generative model for goal-directed synthetic copolymer design using reinforcement learning. Our model operates in a graph generation environment, facilitating efficient monomer unit design while incorporating domain-specific constraints to ensure high validity rates. In a case study optimizing for Hydrogen Evolution Rate (HER) and synthetic accessibility, our approach showcases the efficacy of reinforcement learning in advancing copolymer design. Furthermore, experimental results underscore the challenges in designing effective scoring functions due to the sparse nature of polymer datasets, emphasizing the need for robust property predictors in polymer design methodologies before integrating more complex generative models into the design process. ... Read more

Physics-Informed Neural Networks (PINNs) offer a promising approach to solving partial differential
equations (PDEs). In PINNs, physical laws are incorporated into the loss function, guiding the network to learn a model that adheres to these laws as defined by the PDEs. Training PINNs involves using three types of points: collocation points within the domain of the PDEs, boundary points on the edges of the domain, and initial points at the starting time (t = 0). A common challenge in training PINNs is the imbalance between the number of boundaries and initial points compared to collocation points, which can negatively impact the training process. Typically, the number of each type of point is manually set, and the number of collocation points is usually ten times that of boundary and initial points, leading to an uneven loss distribution over time. Our work introduces a method called Relative Residual Resampling (R3) to address this issue. This method dynamically adjusts the number of each type of training point to ensure a more balanced distribution. Consequently, the loss is more evenly spread across the time domain, enhancing the performance of the PINN. We tested our method on the 1D Heat Equation and 1D Diffusion Equation. The results demonstrate that our approach reduces the overall loss by at least 6%, and balances the loss distribution over time by reducing the loss range for at least 65% compared with the state-of-the-art residual-based resampling strategy. This improvement makes PINNs more accurate and efficient for solving time-dependent PDEs, providing a practical solution for applications where understanding temporal dynamics is crucial. ... Read more

Recent advancements in machine learning (ML) have shown promise in accelerating polymer discovery by aiding in tasks such as virtual screening via property prediction, and the design of new polymer materials with desired chemical properties. However, progress in polymer ML is hampered by the scarcity of high-quality, labelled datasets, which are necessary for training supervised ML models. In this work, we study the use of the very recent ’Joint Embedding Predictive Architecture’ (JEPA) type for self-supervised learning (SSL) on polymer molecular graphs, to understand whether pretraining with the proposed SSL strategy improves downstream performance when labelled data is scarce. By doing so, this study aims to shed light on this new family of architectures in the molecular graph domain and provide insights and directions for future research on JEPAs. Our experimental results indicate that JEPA self-supervised pretraining enhances downstream performance, particularly when labelled data is very scarce, achieving improvements across all tested datasets. ... Read more

The emergence of Language Language Models (LLMs)-based agents represents a significant advancement in artificial intelligence (AI), offering new possibilities for complex problem-solving and interaction within a virtual environment. Our work is based on the Voyager paper [1], which is a state-of-the-art LLM-based agent for Minecraft. However, this system suffers from some significant limitations, such as its reliance on closed-source LLMs and lack of social awareness. Indeed, current open-source LLMs often fail to match closed-source ones in the agent setting, leaving research reliant on third-party closed-source technology [1] [2]. This gap highlights the need for alternative strategies to enhance LLM performance without the high costs associated with fine-tuning. To address these challenges, we propose the Collaborative Voyager, a new architecture designed to enable agent collaboration and social awareness using open-source LLMs. Inspired by the social intelligence hypothesis, which suggests that intelligence emerges from social interactions, we propose collaboration as an alternative learning paradigm for LLMs. This alternative learning paradigm could potentially supplement the expensive fine-tuning currently needed to bridge the performance gap between open-source and closed-source models in the agent setting [2]. Our approach involves developing a framework that allows agents to communicate, understand, and learn from each other, enabling them to correct errors and adapt to new tasks dynamically. By using a memory module, our agent is able to remember interactions and learn from them in order to accomplish a task that it was previously unable to do on its own. Through various experiments, we demonstrate that collaboration significantly enhances the performance of LLM agents in both task completion and adaptability, addressing issues like hallucinations. This study provides insights into developing more sophisticated, adaptable AI systems capable of dynamic interactions and problem-solving. These findings have potential applications extending beyond Minecraft and virtual environments to fields such as robotics, where collaboration and social awareness are crucial. ... Read more

Cancer is one of the leading causes of death. To reduce the amount of deaths caused by cancer, a number of different screening methods are used to detect cancer in an earlier stage, to improve sur vival rates when treating patients with cancer. Cur rent screening methods are often invasive, costly and not very accurate. Therefore, new methods are being sought that aim to be cheaper, less in vasive and provide more accurate results. One of these methods is fragmentomics. Multiple methods have been proposed to use fragmentomics analy sis in the context of screening for cancer, includ ing using the short/long ratio as well as investigat ing the nucleotides at the ends of the fragments. Across previous works using fragmentomics anal ysis to predict cancer, different pre-proccessing steps are used, with limited explanation why the pre-processing methods were chosen. Research into the effects of pre-processing steps used when using fragmentomics analysis is lacking. Two main pre-processing steps in the field are correct ing GC-bias and filtering on MAPQ. Here we in vestigated the impact of three GC-correction meth ods by applying the correction method and then analyzing the resulting fragmentation profiles us ing short/long fragment ratios. Furthermore, three different MAPQ filtering thresholds were studied. This showed that Deeptools correction of the GC bias lowered performance, with the accuracy drop ping from 77.8% to 69.4%. Applying LOESS cor rection using all fragments at the same time re sulted in an accuracy of 83.3%, while applying LOESS correction using the short and long frag ments separately resulted in an accuracy of 91.7%. The impact of filtering the data based on mapping quality was determined by comparing the results of analysing all fragments, analyzing only fragments with mapping quality 5, 20 or 30. This showed that not filtering by mapping quality has a big impact on the profiles of cancer samples, with a KS-test statistic of 0.08 for MAPQ 5 and MAPQ 20 and larger differences in correlations between healthy and cancer samples. The performance of classi fication was much higher when not filtering, with an accuracy of 97.3%, which dropped whenever the filtering threshold was raised, bottoming out at 62.7% for a threshold of MAPQ 30. Due to limita tions with the study, the combined pre-processing of not filtering on MAPQ and using the LOESS separate correction were not studied. ... Read more

The field of causal inference provides a variety of estimators that can be used to find the effect of a treatment on an outcome based on observational data. However, many of these estimators require the unconfoundedness assumption, stating that all relevant confounders are observed within the data. This assumption is quite strict and many real-life problems would violate it, due to confounders being too expensive, immoral, or abstract to observe. To weaken this assumption, causal sensitivity analysis attempts to quantify the level of hidden confounding and use it to create a possible bound on the causal effect. This study compares a well-established Marginal Sensitivity Model (MSM) with a newly proposed f-sensitivity model. Given f-sensitivity is relatively new, the current approaches for computation can be inefficient or non-deterministic. A new method of computing the f-sensitivity bound is proposed, which can lead to closed-form solutions in some estimator-specific cases. The differences between the MSM and f-sensitivity models are outlined and illustrated with examples, from which a set of guiding questions is created for researchers to decide between the two sensitivity models. All of the code required to reproduce this study is located in this GitHub repository. ... Read more

Proteins are fundamental biological macromolecules essential for cellular structure, enzymatic catalysis, and immune defense, making the generation of novel proteins crucial for advancements in medicine, biotechnology, and material sciences. This study explores protein design using deep generative models, specifically Denoising Diffusion Probabilistic Models (DDPMs). While traditional methods often focus on either protein structure or sequence design independently, recent trends emphasize a co-design approach addressing both aspects simultaneously. We propose a novel methodology utilizing Equivariant Graph Neural Networks (EGNNs) within the diffusion framework to co-design protein structures and sequences. We modify the EGNN architecture to improve its effectiveness in learning intricate data patterns. Experimental results show that our approach effectively generates high-quality protein sequences, although challenges remain in producing plausible protein backbones and ensuring strong sequence-structure correlation. ... Read more

Studying the interactions of genes within a cell is an area of significant interest in the field of medicine as it can provide answers to what exactly drives the behavior of a cell under specific circumstances, such as diseases. Once understood, gene interactions can enable the synthesis of efficient, possibly personalized treatments for these diseases and other disorders. However, studying gene interactions requires a large number of samples which might be costly and laborious to obtain in the case of rare disorders for which there is not much recorded data. Geneformer, a context-aware, attention-based deep-learning model, was created specifically for solving this problem. The model makes use of transfer learning to apply any relevant knowledge gained from a larger, similar domain onto a downstream domain with limited data which can be used to further train the model. In this paper, we assessed four fine-tuning strategies, including the one used throughout the in silico experiments presented in the original Geneformer paper. We did this to assess whether the accuracy of Geneformer on the downstream task of predicting the sensitivity of cancer cells to different treatments can be improved versus the default implementation as found within the model's paper. The model was firstly fine-tuned using a training dataset compiled from the sciplex2 dataset, followed by the prediction of the dosage levels to which samples from a test set were exposed. Upon performing the experiment, we concluded that, depending on the way in which knowledge from the source domain is stored inside the pre-trained model and the similarity between the source and the target domains, different fine-tuning strategies were suitable for a given task. Hence, there is no single optimal fine-tuning method which can be used to predict the level to which cancer cells were exposed to treatments such as nutlin-3A. ... Read more

This paper describes ZygosDB, a novel and efficient read-only database designed specifically for querying positional genomic data required for Genome-Wide Association Studies (GWAS). ZygosDB addresses limitations of existing solutions like Tabix by offering optimized data structures, compression techniques, and parallel query execution.

Our evaluation shows that ZygosDB achieves a significant speedup of 2-5 times in query throughput compared to Tabix. This improvement comes from our focus on efficient data storage and retrieval tailored to the specific needs of GWAS.

The paper also explores the impact of multi-threading on query performance and the role of compression algorithms in optimizing query throughput. We identify a decrease in performance beyond a certain number of threads and discuss the influence of compression algorithms like Gzip and LZ4.

While ZygosDB offers substantial performance gains, future work should explore avenues for further optimization, such as measuring query latency, refining memory usage, and investigating specialized column support. Overall, ZygosDB establishes itself as a powerful tool for efficient querying of large genomic datasets, facilitating more effective GWAS. ... Read more