Research Areas AI for Cancer

Artificial Intelligence Models for Cancer Classification

 At the COMBINE Lab, we develop artificial intelligence and machine learning methods to extract clinically and biologically meaningful information from complex cancer data. Our research spans multiple data modalities, including genomic, epigenomic, transcriptomic, single-cell, histopathological, and clinical data. We investigate how modern representation-learning approaches, including foundation models, capture cancer-related molecular and morphological patterns, giving particular attention to the robustness, interpretability, and computational efficiency of these models. Potential clinical applications of our research include molecular tumor classification and subtyping, computational analysis of histopathological images, characterization of intratumoral heterogeneity and the tumor microenvironment, and the identification of clinically relevant biomarkers to support cancer diagnosis and patient stratification.

Our research actively translates into concrete computational solutions. For instance, our recent work on single-cell foundation models showed that selecting representations from intermediate network layers substantially improves the reconstruction of cellular differentiation trajectories and the modeling of perturbation responses. In practice, this enables a more reliable identification of disease-relevant cell states and context-specific responses to genetic or pharmacological interventions, which are crucial for studying tumor progression and treatment resistance.

In the field of digital pathology, we focus on making the analysis of histological images significantly more efficient by identifying and retaining only the image regions most relevant to the model’s prediction. On breast and colorectal histopathology benchmarks, our approach processed image tiles up to 3.53 times faster while retaining only 10–20% of the original image tokens and fully preserving classification performance. This type of approach reduces the computational cost of analyzing large collections of whole-slide images, making advanced AI models accessible to laboratories with limited computing resources, while the retained regions provide clear spatial attribution maps that show exactly which tissue structures contributed most to the model’s output.