Cancer is one of the leading causes of death worldwide and has recently surpassed cardiovascular disease as the number one cause of death in high-income countries. Glioblastoma is an especially belligerent brain tumor known for its notorious behavior of adapting its cellular machinery in ways that allow it to become unresponsive to treatment. It has an incidence of 0.59 to 5 per 100,000 people per year. The average life expectancy after diagnosis is around 17 months due to recurrence. Therapeutic failure has multiple causes, but the two most important underlying factors are the heterogeneity of the genetic makeup of glioblastoma and the instability of the cancer cell regulatory processes at the genetic and epigenetic level. The diverse phenotypic traits that result from the combined effect of these microscopic factors exhibit high plasticity at the macroscopic level, which further contributes to therapy resistance. Plasticity here refers to the ability of cells to transition from one phenotypic state to another due to external and/or internal microenvironmental factors. This is typically observed as a response to drug treatment, rendering tumors likely to switch to a more resistant state.
 

Faced with the poor understanding of the plasticity phenomenon in general, the heterogeneity of patient-specifc phenotypic plasticity profiles and the high dimensionality of the associated data, the use of AI methods is crucial to tackle the complexity characterizing the plasticity landscape and to advance towards individualized treatments against a resilient tumor. Given the diversity of cellular states driven by the malignant transformation and the increased plasticity, the current datasets generally represent very sparse temporal snapshots with little predictive power regarding the temporal evolution of the tumor. In our project, we tackle these challenges via two approaches. Firstly, we augment the datasets with synthetic data using datadriven style transfer techniques and generative models, such as GANs (generative adversarial networks), to enhance the representation of sparse cellular states. Secondly, we develop multiscale models based on hybrid AI modeling, which captures both discrete states and continuous trajectories by leveraging autoencoders and optimal transport approaches. Multi-modal experimental datasets including transcriptomic, epigenomic and proteomic data will be combined in integrative AI models, in which they will be projected into a common phenotypic space. In such a reduced space, dimensions correspond to phenotype features (signaling and metabolic pathway activities corresponding to cancer hallmarks and master transcription factor activities) of the normal and malignant cells. The projection into a lower-dimensional space, whose dimensions have a biological interpretation, is based on interpretable AI models, incorporating a priori biological knowledge in the form of biological graphs within an autoencoder structure. These models will be enhanced by applying state-of-the-art methods such as self-supervised learning (contrastive learning or masked autoencoders) to increase their phenotypic resolution.
 

In light of above facts, the three most important research questions of this project are:

• How can we characterize an interpretable phenotypic space to delineate the phenomenon of tumor plasticity in glioblastoma?
• What is the survival strategy that a cancer cell applies and how does it relate to the pathway selection and motion dynamics in the phenotypic landscape?
• How can we use AI-models for devising groundbreaking personalized glioblastoma therapy strategies?
   

• Glioblastoma
• Therapy reistance
• Personalized medicine
• Artificial Intelligence
   

January 2024 - December 2029

Prof. Dr.-Ing. Naim Bajcinca
Gottlieb-Daimler-Str. 42
67663, Kaiserslautern
+49 (0)631/205-3230
naim.bajcinca(at)mv.uni-kl.de