---

The Challenge of Predictive Oncology

Cancer treatment is increasingly moving toward personalized medicine, yet clinicians often still rely on "average" patient responses to guide individual care^{1, 2}. Whether dealing with the extreme unpredictability of pediatric neuroblastoma, which can range from spontaneous regression to fatal progression^{3, 4}, or the diagnostic uncertainty of prostate cancer monitoring via blood tests ^{5, 6, 7}, the need for patient-specific "Digital Twins" is clear.

Our computational team has developed a suite of computational tools that combine multiscale modeling, biomechanics, and artificial intelligence algorithms. By integrating imaging and biochemical biomarkers, we aim to provide a "what-if" platform for testing treatments in silico.

1. The "Eyes" of the Twin: Extracting Accurate Biomarkers

A Digital Twin requires high-quality data. We use Dynamic Contrast-Enhanced MRI (DCE-MRI) to map tumor vascularization, but traditional pharmacokinetic models often ignore the physical diffusion of contrast agents between neighboring voxels, leading to errors. In Sainz-DeMena et al. (2024)⁴, we introduced Physics-Informed Neural Networks (PINNs) to solve this inverse problem. Unlike standard AI, PINNs are constrained by the physical laws of mass conservation. By embedding these Partial Differential Equations (PDEs) into the neural network’s loss function, we retrieve vascularization parameters (K_trans) with significantly higher accuracy, even from noisy or incomplete clinical data.

2. The "Brain": Learning Cell-Level Decisions

How do thousands of cells self-organize? In Camacho-Gomez et al. (2023)³, we developed a hybrid framework where a Deep Learning algorithm manages the decision-making of individual agents (cells) in a 3D model.

Instead of fixed rules, the neural network evaluates the simulation in real-time and "evaluates" if a cell should proliferate, stay quiescent, or secrete fluid to form a lumen. This model successfully reproduced the complex morphogenetic patterns of tumor organoids, proving that AI can help us unravel the underlying principles of how cancer cells self-organize.

3. The "Orchestra": Multiscale Frameworks for Neuroblastoma

In Borau et al. (2023)¹, we presented a massive multiscale orchestrated framework. This "orchestrator" integrates four distinct modules: macro-scale transport, a subcellular machine learning model of drug-protein interactions, a GPU-accelerated agent-based model (ABM), and a macro biomechanical model. This infrastructure allows clinicians to run an 80-day treatment simulation in less than 12 hours on High-Performance Computing (HPC) clusters.

4. Clinical Validation: Biomechanical Responses

In Hervas-Raluy et al. (2024)⁵, we turned MRI scans into biomechanical Finite Element models. By feeding the models image-based biomarkers like cellularity (from ADC maps) and vascularization (K_trans), our simulations predicted the dramatic 90% volume reduction in a responsive patient versus the minimal 15% reduction in a resistant one. This provides a window into "what-if" scenarios, such as the potential impact of proangiogenic treatments to improve drug delivery.

5. Reconstructing Trajectories: PSA-Driven Digital Twins

Moving from the mechanical to the biochemical, our most recent work in Camacho-Gomez et al. (2025)² addresses prostate cancer monitoring. Current methods often fail to detect tumor growth if Prostate-Specific Antigen (PSA) levels don't rise significantly, a phenomenon known as "hidden growth"^2,5,6. We developed a Physics-Informed Machine Learning (PIML) digital twin that reconstructs a patient's tumor growth trajectory directly from PSA blood tests². This model is geometry-agnostic; it uses voxelized representations that allow the same neural network architecture to be applied across different patients without retraining. In clinical tests, the twin reconstructed 2.5 years of tumor evolution with errors as low as 0.8%, revealing growth even in stable-PSA scenarios⁵.

The Path Forward

These seven works represent a shift toward personalized computational oncology. By bridging the gap between clinical data (MRI, PSA), physics-based modeling, and deep learning, we are moving closer to a future where every patient’s treatmentis guided by a digital simulation of their unique disease.

References

1. Borau, C. et al. A multiscale orchestrated computational framework to reveal emergent phenomena in neuroblastoma. Comput. Methods Programs Biomed. 19, (2023). https://doi.org/10.1016/j.cmpb.2023.107742



2. Camacho-Gomez, D. et al. Physics-informed machine learning digital twin for reconstructing prostate cancer tumor growth via PSA tests. npj Digit. Med. 5, (2025). https://doi.org/10.1038/s41746-025-01890-x



3. Camacho-Gomez, D. et al. A hybrid physics-based and data-driven framework for cellular biological systems: Application to the morphogenesis of organoids. iScience 40, (2023). https://doi.org/10.1016/j.isci.2023.107164



4. Sainz-DeMena, D. et al. Exploring the potential of Physics-Informed Neural Networks to extract vascularization data from DCE-MRI in the presence of diffusion. Med. Eng. Phys. 9, (2024). https://doi.org/10.1016/j.medengphy.2023.104092



5. Hervas-Raluy, S. et al. Image-based biomarkers for engineering neuroblastoma patient-specific computational models. Eng. Comput. 25, (2024). https://doi.org/10.1007/s00366-024-01964-6



6. Pérez-Benito, Á. et al. Patient-specific prostate tumour growth simulation: a first step towards the digital twin. Front. Physiol. 15, (2024). https://doi.org/10.3389/fphys.2024.1421591



7. Pérez-Benito, Á. et al. In-silico patient-specific modelling of prostate cancer: Predicting PSA dynamics and treatment response. Comput. Methods Programs Biomed. 270, 108931 (2025). https://doi.org/10.1016/j.cmpb.2025.108931