Deep Learning Models of Cancer Mechanisms
Abstract
Cancers are heterogeneous in terms of mutation profile, cell type of origin, and micro-environment, and precision cancer medicine aims to find effective treatments for each individual cancer. This is challenging since cellular processes are highly interconnected and complex. We develop deep learning models of the molecular networks of cells, to identify systems-level mechanisms in cancer.
Deep learning models excel at finding complex relations in large quantities of data but are not always easy to interpret. To address this, we constrain the models, so that their predictions align with physical interactions between bio molecules. Our models cover metabolism, signal-transduction, and gene regulation and are trained on high throughput datasets such as transcriptomics, metabolomics, and proteomics from different cell types, interventions, and conditions.
Using our models, we aim to find new drug targets and biomarkers, predict resistance mechanisms, and improve the understanding of cell-cell interactions in the tumor microenvironment. Our long-term goal is to enable computer aided design of cancer medicine.
For this work we use the LEMBAS framework [1], and updated versions thereof [2,3], as well as our extension that covers cellular metabolism. LEMBAS is a deep learning framework for modeling cellular signaling that has proven useful in predicting transcription in response to ligand stimulation and drugs, and for generating biologically interpretable predictions.
Sub project 1: Cancer metabolism. Predicting the metabolic state of cancer cell lines from gene expression. A new architecture has been developed that recurrently combines deep learning prediction of fluxes with linear updates of metabolite levels.
Sub-project 2: Cancer signaling. A mechanistic deep learning model of gene regulation in cancer. A new architecture has been developed that allows for state-of-the-art modelling of cancer cells, modelling aspects such as mutation, knockouts, drug interactions etc., while allowing for interpretability where molecular mechanistic cause can be attributed to specific interactions.
Sub-project 3: Cell types in cancer. The prediction of cell type–specific cancer responses are improved by incorporating a more detailed representation of the cell-type-of-origin into the model. The project uses both cancer data, including epigenetics, and healthy reference atlases (e.g. Tabula Sapiens) to investigate tissue-specific tumorigenesis pathways.
Sub-project 4: Signaling and metabolism integration. This sub-project aims to develop a mechanistic deep learning model that integrates metabolism, signaling, and gene regulation into a single interpretable architecture. Prior knowledge of molecular interaction networks (~117,000 interactions among >20,000 molecular entities) is embedded directly into the model structure, allowing predictions to be mechanistically attributed to specific biomolecules.
[1] Nilsson, A., Peters, J.M., Meimetis, N. et al. Artificial neural networks enable genome-scale simulations of intracellular signaling. Nat Commun 13, 3069 (2022). https://doi.org/10.1038/s41467-022-30684-y
[2] Nordenstorm, O., Baghdassarian, H., Xu, X., Lauffenburger, D. A. & Nilsson, A. Biologically informed neural network models are robust to spurious interactions via self-pruning. Bioinformatics 42, btag556 (2026).
[3] https://github.com/AvlantNilssonLab
