Specific aims
1) Investigate Joint Embedding Models to Integrate WSI Representations with Multi-Omics Data
We aim to develop joint embedding models that combine representations from H&E stained WSIs and multi-omics data (e.g., transcriptomics, genomics, proteomics) to enhance the understanding of cancer biology. By integrating multi-omics data during the pre-training phase, we seek to improve the capacity of WSI-based models to capture complex biological features associated with cancer progression and treatment response. Techniques such as contrastive learning or multi-modal autoencoders will be used to create a unified embedding space that reflects both histological and molecular characteristics. We will evaluate these joint embeddings in downstream tasks, such as predicting patient outcomes and identifying biomarkers, to determine whether they outperform models trained with WSI or omics data alone.

2) Evaluate the Prognostic Value of H&E WSIs Compared to Established Genomic Signatures
We will compare the predictive accuracy of WSI-based models with established genomic signatures, including HRDscore, HRDetect, and other FDA-approved assays, in predicting cancer prognosis and treatment response. Our goal is to determine if histological features extracted from WSI patches can serve as surrogates for certain genomic signatures, potentially offering a non-invasive and cost-effective alternative for assessing molecular risk factors. We will conduct a comprehensive analysis of the concordance between WSI-derived features and genomic signatures across multiple cancer types and patient cohorts, exploring the potential for WSI models to enhance or complement existing genomic risk models.

3) Explore Spatial Features Between Tissue Types Using Novel Models Based on Correlation Functions
We plan to develop models that leverage correlation functions between tissue "prototypes" (distinct histological regions) to capture spatial relationships within WSIs. This approach will investigate the prognostic significance of spatial features, such as tissue organization, interaction, and distribution, and their association with clinical outcomes. Correlation-based models will help identify patterns reflecting biological processes like immune response and tumor microenvironment dynamics. We will apply interpretability techniques to understand which spatial features and tissue interactions are most predictive of outcomes, enabling the generation of new biological hypotheses.

4) Enable Automatic Hypothesis Generation Using AI-Based Models
Our AI models will automatically identify potential research questions or hypotheses by detecting novel patterns, biomarkers, or associations between tissue and omics data. Using unsupervised learning and clustering, we will uncover previously unknown patient subgroups linked to distinct clinical outcomes or therapeutic responses. We aim to validate these AI-generated hypotheses through retrospective and prospective studies, ensuring their clinical relevance. This approach facilitates a shift from predictive modeling to discovery-driven research, employing the developed models as tools for continuous learning and exploration in oncology.

These aims address critical challenges in precision medicine, integrating digital pathology and AI to enhance cancer understanding, prognosis accuracy, personalized treatment, and hypothesis generation. Together, they push the boundaries of pathology and computational biology, aiming to transform clinical decision-making.

Istvan Csabai, Dept. of Complex Systems, Eotvos University, Budapes, Hungary