AI for Discovery

The traditional drug discovery process is notoriously slow and expensive, often failing because standard tests only look at how a drug affects a single biological target. Cell Painting offers a powerful and unbiased alternative: by using a multiplexed set of fluorescent dyes to highlight various organelles and structures within a cell, this technique captures a holistic snapshot of how cells react to different chemical compounds.

The result is a massive dataset of high-resolution images containing thousands of hidden morphological features. For data scientists, this presents an exciting challenge. This project aims to develop a state-of-the-art machine learning pipeline that analyses high-dimensional Cell Painting data to automate drug screening, predict drug toxicity, and uncover new therapeutic mechanisms before clinical trials begin.

Objectives

1. Build a feature extraction module using modern computer vision architectures to identify subtle morphological changes in raw cell images — moving beyond manual feature engineering, and including explainability techniques to ensure the extracted profiles capture true biological variance rather than image noise.
2. Implement clustering and dimensionality reduction algorithms to analyse the high-dimensional phenotypic profiles, grouping unknown compounds with known landmark drugs to reliably predict their mechanisms of action and identify potential off-target effects across diverse biological pathways.

References

1. Gopalakrishnan, V., Ma, J., and Xie, Z. "Grad-CAMO: Learning interpretable single-cell morphological profiles from 3D Cell Painting images." CVPR (2024).
2. Seal, S., et al. "Cell Painting for image-based profiling: a decade of discovery." Nature Methods (2025).
3. Camilleri, F., et al. "Cell Painting and Chemical Structure Read-Across Can Complement Each Other for Rat Acute Oral Toxicity Prediction." Chemical Research in Toxicology (2024).

Work environment

The student will join a diverse team of researchers from the Informatics Institute (IvI) and AmsterdamUMC, including experts in AI development, computer vision, experimental design, and pharmacology. You will contribute to a computational pipeline aimed at accelerating drug discovery and reducing clinical failure rates.

The Cellular Potts Model (CPM) is an agent-based framework widely used to simulate complex multicellular systems such as tumour invasion, cell sorting, and tissue development at single-cell resolution. While running forward simulations from known physical rules is well understood, solving the inverse problem remains a major challenge: how can we automatically infer underlying biophysical parameters (such as cell adhesion energies, surface tension, and motility forces) directly from image snapshots?

Solving this inverse problem is essential for calibrating mechanistic models to experimental microscopy data and building biological digital twins. This project aims to design and benchmark modern machine learning and computer vision pipelines to extract simulation parameters directly from static and time-series image snapshots.

Objectives

1. Build an automated data generation pipeline across diverse parameter configurations and tissue morphologies using open-source CPM simulation engines.
2. Implement and benchmark deep learning architectures (CNNs, Vision Transformers, Graph Neural Networks) to map spatial cell arrangements and morphological features back to ground-truth simulation parameters.
3. Formulate simulation-based inference (SBI) and neural posterior estimation methods to evaluate parameter identifiability, quantify prediction uncertainty, and test robustness on sparse or noisy snapshot data.

References

1. Arruda, J., et al. "Simulation-Based Inference of Cell Migration Dynamics in Complex Spatial Environments." npj Systems Biology and Applications (2026).
2. Minartz, K., et al. "Deep neural cellular Potts models." arXiv:2502.02129 (2025).
3. Braat, Q. J. S., et al. "Shape matters: inferring the motility of confluent cells from static images." Soft Matter 21.28 (2025).

Work environment

The student will join an interdisciplinary research team working at the intersection of AI for science, computer vision, and computational biology at UvA and AmsterdamUMC. You will work closely with researchers in machine learning and mathematical modelling, gaining hands-on experience in deep learning, probabilistic inference, and mechanistic simulation modelling.

3D in-vitro models enable quantification of proliferation dynamics of cancer cells. Real-time imaging reveals changes in growth and morphology arising from the addition of drugs to 3D spheroids. Currently, there is a lack of AI tools for analysing such morphological changes arising from drug treatments.

The goal of this project is to develop segmentation algorithms to identify and determine the metrics associated with drug efficacy from time-series image capture of treated 3D spheroids. The metrics comprise change in area of coverage, dissociation of adherent cells, and compactness of spheroids subjected to drug treatment.

Objectives

1. Develop segmentation algorithms for time-series images of treated 3D spheroids.
2. Derive drug-efficacy metrics: area of coverage, dissociation of adherent cells, and spheroid compactness.
3. Validate the extracted metrics against experimental drug-response readouts.

Expectations

Basic experience in AI development and programming.

Work environment

The student(s) would join a diverse team of researchers and clinicians from the Informatics Institute (IvI) and AmsterdamUMC, comprising experts in AI development, experimental design, and validation. You will work closely with medical researchers and develop AI tools to quantify drug efficacies.

Clinical assessment in dental education relies heavily on case-based multiple-choice questions and realistic patient vignettes to evaluate diagnostic reasoning and treatment planning. However, authoring high-quality exam items that mirror modern clinical standards is labour-intensive and difficult to scale.

Recent breakthroughs in Retrieval-Augmented Generation (RAG) offer powerful tools for automated item creation, but applying them to specialized dental curricula requires solving essential software engineering challenges around clinical accuracy, provenance, guideline alignment, and verifiable reasoning. This project aims to design, implement, and evaluate a robust, end-to-end AI system that automatically generates, validates, and refines clinically authentic dental examination questions grounded in trusted medical knowledge bases.

Objectives

1. Architect and implement a modular RAG pipeline that ingests accredited dental textbooks, clinical guidelines, and curricula to ground question generation and prevent hallucinations.
2. Build an automated multi-agent reasoning framework that produces plausible clinical distractors, contextual patient vignettes, and step-by-step explanatory rationales aligned with pedagogical standards such as Bloom's taxonomy.
3. Develop an automated validation and psychometric benchmarking suite to evaluate item difficulty, clinical fidelity, and distractor quality against expert-written faculty benchmarks.

References

1. Xiong, G., et al. "Benchmarking retrieval-augmented generation for medicine." arXiv:2402.13178 (2024).
2. Riehm, L., et al. "The use of large language models in generating multiple choice questions for health professions education: a systematic review and network meta-analysis." PLOS One (2026).
3. Kong, M., et al. "A Scoping Review of Large Language Models in Dental Education: Applications, Challenges, and Prospects." International Dental Journal (2025).

Work environment

The student will join an interdisciplinary team bridging software engineering, computational modelling and clinical dental education at UvA and ACTA. This setup provides access to advanced language model tooling, domain-specific corpora, and experienced dental faculty for human-in-the-loop evaluation.