Oportunidades de Dissertação

Instituição: Universidade de Utrecht.

Nome do Responsável: Wilson Silva.

Contacto: Assistant Professor | AI Technology for Life | Dep. of Information and Computing Sciences | Dep. of Biology | Utrecht University | Buys Ballot building room 4.79 | w.j.dossantossilva@uu.nl | Mon-Fri
Guest Scientist | Dep. of Radiology | The Netherlands Cancer Institute | w.santos@nki.nl | Often at the office on Fridays

Público Alvo: Alunos de mestrado em Engenharia Biomédica/Bioengenharia.

Requisitos: Conhecimentos básicos de Machine Learning.

Temas de Dissertação:

Explainable AI for Spatial Transcriptomics from Histology Images; 

Machine Unlearning for Case-Based Explanations: Privacy vs Robustness;

Synthetic Data for Concept-Based Explainable AI;

Ordinal Representation Learning for Single-Cell RNA Sequencing;

Pretraining General-Purpose Aggregators for Crosslevel Biological Data; 

Incorporating Biological Priors into Wave-PDE Networks;

Deep Learning for Compositional Biological Data.

Descrições das Dissertação: 

Spatial transcriptomics is expensive and limited, while WSIs are widely available. This
project explores learning a mapping from WSIs to spatial gene expression with a focus on
interpretability. We will study which tissue regions drive gene expression predictions and
whether explanations align with known biology;

Case-based explanations retrieve similar training examples, raising privacy concerns and
propagating biases. This project studies machine unlearning from two angles: removing
sensitive samples to reduce leakage, and removing spurious cases to improve robustness.
We will analyze how unlearning affects retrieval, explanations, and generalization;

Concept-based models rely on costly annotations. This project explores using synthetic
images to define and control concepts. We will study whether synthetic data can improve
scalability of concept-based explanations while preserving reliability and avoiding biases;

Many biological processes follow ordinal structure (e.g., dosage, progression). This project
explores incorporating ordinal constraints into deep models for single-cell RNA-seq. We will
analyze how this affects representation structure, robustness, and biological interpretability;

Biological data is often limited, making it hard to train complex models. Foundation models
address this by learning general representations but typically operate at a single biological
level. This project explores pretraining aggregation mechanisms that combine information
across levels (e.g. protein → bacterium), facilitating crosslevel explainability. We will study
alternative training strategies, as standard objectives like masking do not directly apply;

Wave-PDE Nets are a recent, data-efficient alternative to attention, with explicit parameters
that control information propagation across positions, providing a natural way to incorporate
prior knowledge. We will explore how to incorporate biological priors into wave-PDE nets,
and whether it improves performance and explainability;

Many biological datasets, such as microbiome and RNA sequencing data, are compositional
and naturally live in the simplex rather than Euclidean space. While common transformations
(e.g. log, clr) enable the use of standard models, they do not guarantee that learned
representations preserve the underlying geometry. This project explores how to design or
adapt deep learning models that better capture simplex geometry and learn robust
representations from compositional data.

Período: Flexíveis, idealmente para realização de tese de mestrado durante o ano letivo de 2026/2027. Estamos também abertos a receber estudantes em Erasmus, tanto no primeiro como no segundo semestre.