Cancer

The tumor microenvironment (TME) is a dynamic ecosystem where tumor cells, immune cells, and stromal cells interact through physical contact and diffusible signals like cytokines and chemokines. Simulating this system typically requires computationally expensive hybrid models that couple a Cellular Potts Model (CPM) for cellular dynamics with reaction-diffusion partial differential equations for chemical signals. These simulations are slow, which limits their use in parameter sweeps, uncertainty quantification, and therapy design. This project asks a simple question: can we learn a fast surrogate that predicts the future of such a system directly from its current state, skipping the expensive solver entirely?

Recent work has shown that deep learning surrogates, including Physics-Informed Neural Networks (PINNs), can approximate these kinds of spatiotemporal systems. Meanwhile, state-of-the-art approaches elsewhere in biology now model tissues as spatially coupled systems whose future states can be predicted and even virtually perturbed, as seen in frameworks such as GraphTME and TMEformer. This project pushes that idea further by training a Joint Embedding Predictive Architecture (JEPA), a self-supervised world model, to predict the evolution of coupled reaction-diffusion and Cellular Potts systems without reconstructing every pixel. Instead of learning in raw pixel space, the model learns a hidden representation where prediction happens in latent space, making it far more efficient and scalable than PINN-style pixel-level reconstruction.

Objectives

1. Build a hybrid ground-truth simulator coupling a Cellular Potts Model with reaction-diffusion PDEs, and use it to generate a large, diverse dataset of simulated TME trajectories.
2. Train a JEPA-style world model that learns to predict future latent states of the system from current states, with an optional conditioning signal for external perturbations such as drug delivery or cytokine gradients.
3. Benchmark the learned world model against PINN surrogates from prior work, evaluating both predictive accuracy and computational speedup over the original solver.
4. Explore whether the model enables long-horizon rollouts, uncertainty estimation, and in silico perturbation experiments that are infeasible with the full simulation.

References

1. Papapanagiotou, I., et al. "From simulations to surrogates: Neural networks enhancing burn wound healing predictions" (2025).
2. Sheraton, M. V., et al. "Emergence of spatio-temporal variations in chemotherapeutic drug efficacy: in-vitro and in-silico 3D tumour spheroid studies." BMC Cancer 20 (2020).
3. LeCun, Y. "A path towards autonomous machine intelligence." (2022) — JEPA position paper.

Expectations

Prior experience with physics simulation, graph neural networks, or self-supervised learning is a plus but not required.

Work environment

The student will join a diverse team of researchers from the Informatics Institute (IvI) and AmsterdamUMC, with expertise in AI development, physical simulation, and experimental validation.

Cancer progression and drug resistance are driven by complex cellular conversations within intact tissue ecosystems. While traditional assays look at single snapshots of dissociated cells, next-generation spatial multiomics and image-based pooled perturbation screens allow us to observe both the molecular makeup and the spatial behaviour of cancer and immune cells under genetic or drug interventions.

However, mapping these high-dimensional, multi-modal measurements to the underlying causal mechanisms remains a major computational hurdle. This project aims to build an interpretable computational framework that integrates multiplexed imaging, spatial multiomics, and perturbation data to uncover the causal signalling pathways and spatial rules governing tumour microenvironment remodelling.

Objectives

1. Develop a multi-modal representation learning model using spatial graph neural networks and optimal transport to align multiplexed tissue imaging with spatial transcriptomic and proteomic profiles.
2. Formulate a generative or neural optimal transport framework to predict single-cell phenotypic and spatial state transitions in response to genetic and pharmacological perturbations.
3. Reconstruct spatially resolved, mechanistic cell–cell communication networks to identify how localized microenvironmental niches drive therapy resistance and immune evasion.

References

1. Bunne, C., et al. "Learning single-cell perturbation responses using neural optimal transport." Nature Methods 20.11 (2023): 1759–1768.
2. Cang, Z., et al. "Screening cell–cell communication in spatial transcriptomics via collective optimal transport." Nature Methods 20.2 (2023): 218–228.
3. Barylli, M., et al. "Network methods for diagonal integration of unpaired single-cell multiomics data: a review." Bioinformatics (2026).

Work environment

The student will join an interdisciplinary team bridging computational science, machine learning, and oncology at UvA and AmsterdamUMC.